Determination of Environmental Pollutants by Gas. Chromatography/Mass Spectrometry with Chemometrics. A dissertation presented to.

Transcription

1 Determination of Environmental Pollutants by Gas Chromatography/Mass Spectrometry with Chemometrics A dissertation presented to the faculty of the College of Arts and Sciences of Ohio University In partial fulfillment of the requirements for the degree Doctor of Philosophy Mengliang Zhang May Mengliang Zhang. All Rights Reserved.

2 2 This dissertation titled Determination of Environmental Pollutants by Gas Chromatography/Mass Spectrometry with Chemometrics by MENGLIANG ZHANG has been approved for the Department of Chemistry and Biochemistry and the College of Arts and Sciences by Peter de B. Harrington Professor of Chemistry and Biochemistry Robert Frank Dean, College of Arts and Sciences

3 3 ABSTRACT ZHANG, MENGLIANG, Ph.D., May 2015, Chemistry Determination of Environmental Pollutants by Gas Chromatography/Mass Spectrometry with Chemometrics Director of Dissertation: Peter de B. Harrington This dissertation focuses on the determination of environmental pollutants (i.e., polychlorinated biphenyls, trichloroethylene) using gas chromatography/mass spectrometry (GC/MS) with chemometrics. First, a novel fast screening method was developed for the determination of polychlorinated biphenyls (PCBs) that are constituents of the commercial mixture, Aroclor 1260, in soil matrices by GC/MS combined with solid-phase microextraction (SPME). The advantages of the use of potassium dichromate and sulfuric acid solution for the extraction of PCBs from soil are fully discussed. It is the first time to use SPME-GC/MS for the quantitative study of Aroclor. Second, a fast method for the identification of 7 Aroclors by GC/MS and chemometrics is described. The modulo compression was introduced and evaluated for the classification of complex mixtures of PCBs by GC/MS for the first time. A fuzzy rule-building expert system (FuRES), fuzzy optimal associative memories (FOAMs), and partial least-squares discriminant analysis (PLS-DA) are compared with

4 4 respect to classification rates. Aroclor soil samples were classified with greater than 95% prediction rates based on models built from pure Aroclor standards. Third, a method for the quantitation of Aroclor mixtures in soils by using SPME-GC/MS and PLS models was developed. Compared with other reported methods which usually need expensive high resolution GC/MS, tedious sample preparation, and long run times, this method is simple, fast and automated, and performed the analysis on a low resolution mass spectrometer. Fourth, PCBs in soils were measured by portable GC/MS instrument. The method has the advantage of the high sample throughput, with a soil sample being prepared and analyzed about every 37 min. By adapting the headspace SPME method with portable scale and heating block, the on-site sampling and sample preparation can be perform on the field. Although the capability of the quantitative analysis for PCBs and Aroclors using this method is understated, this method can be very beneficial and cost-effective for the fast decision of environmental sample investigation. At last, a novel sample preparation method, liquid-liquid microextraction assisted solid phase microextraction (LLME-SPME) with portable gas chromatography/mass spectrometry, for the

5 5 determination of trichloroethylene (TCE) was developed. This method significantly improved the extraction efficiency compared with SPME and was suitable for field analysis because of its simplicity. Two chemometric methods, classical least-squares (CLS) and inverse leastsquares (ILS), were applied to resolve overlapping TCE and deuterated TCE (TCE-d) mass spectral signals and evaluated for the determination of TCE. CLS was demonstrated for the first time as a method to resolve overlapping isotopic peak clusters between the analyte and its internal standard, thereby allowing the use of less expensive deuterated standards. The method enables simple isotopic analogs of analytes (one H or C atom is isotopic labeled) to apply as internal standards. It is the first application of chemometrics to overcome overlapping peaks between an analyte and its corresponding isotopic internal standard.

6 6 DEDICATION To my wife Wen and my son Yi, every day they make of this world a better place to be.

7 7 ACKNOWLEDGMENTS I would like to give my gratitude to my advisor Dr. Peter de B. Harrington for his professional, scientific guidance, and encouragement. He accepted me to join his group on my third-year of PhD study and gave me much freedom and trust on research. He taught me the critical thinking and writing scientific papers with patience and kindness. I am grateful for the friendship from Peter and his family. I would also like to thank Dr. Glen P. Jackson, Dr. Shiyong Wu and Dr. Natalie A. Kruse for agreeing to serve on my dissertation committee and for their thoughtful advice and precious suggestions. I especially thank Dr. Glen P. Jackson, a gentleman with patience and kindness, who was my advisor for first two years of PhD study. He instructed me and offered the research assistant fellow to allow me to concentrate on my research. Bascom French and Paul Schmittauer are acknowledged for their assist on my research. Jennifer R. Bowman, Stephanie Howe, Robert Eichenberg are thanked for their effort on the PCB project. Stephen A. Lammert, Edgar D. Lee, and Kevin Winder are thanked for their help on troubleshooting of portable instrument.

8 8 Past and current members in Jackson and Harrington s groups are thanked for their support, help and friendship. I also thank the Department of Chemistry and Biochemistry at Ohio University for the support of my research and I am also grateful for the support from my beloved parents.

9 9 TABLE OF CONTENTS Page Abstract... 3 Dedication... 6 Acknowledgments... 7 List of Tables List of Figures Chapter 1: Introduction Introduction of dioxins and PCBs Methods for the determination of PCBs Congener-specific method and limited congener Aroclor estimation methods Instrumentation of PCB analysis methods General introduction of PCB sample extraction methods Other aspects of PCB analysis methods Introduction of TCE Introduction of selected chemometric methods Chemometric methods for Aroclor identification Chemometric methods for Aroclor quantification Chemometric methods for TCE quantification Data preprocessing methods Chapter 2: Determination of Aroclor 1260 in Soil Samples by GC/MS with Solid Phase Microextraction Experimental Reagents Instruments Sample preparation Results and discussion... 61

10 2.2.1 Optimization of SPME conditions GC/MS analysis Analytical method performance Recovery evaluation of SPME-GC/MS method Validation of method by certified soil samples Conclusions Chapter 3: Automated Pipeline for Classifying Aroclors in Soil by Gas Chromatography/Mass Spectrometry using Modulo Compressed Twoway Data Objects Materials and methods Reagents Instruments Data collection Data format Data preprocessing Data processing Results and discussion SPME-GC/MS analysis Baseline correction Modulo compression FuRES, PLS-DA, and FOAM classification Classification of Aroclor soil samples Conclusions Chapter 4: Simultaneous Quantification of Aroclor Mixtures in Soil Samples by Gas Chromatography/Mass Spectrometry with Solid Phase Microextraction using Partial Least-Squares Regression Experimental Reagents Instruments

11 4.1.3 Sample preparation Results and discussion PLS method Validation of method Detection limit determination Conclusion Chapter 5: Expedited Field Survey and Sampling Method for PCBs, PCDDs and PCDFs using Portable Gas Chromatography/Mass Spectrometry Experimental Portable GC/MS Devices for SPME extraction Reagents Sample preparation and analysis Results and discussion Extraction condition optimization Peak identities and general performance Quantitation Applications Analysis of PCDDs and PCDFs Conclusions Chapter 6: Determination of Trichloroethylene in Water using Liquidliquid Microextraction Assisted Solid Phase Microextraction and Classical Least Squares Resolution of Overlapping Trichloroethylene and Its Isotopic Internal Standard Peak Clusters Materials and methods Reagents and materials Instruments Data collection

12 6.1.4 Data format Retention time alignment Data processing Results and discussion LLME-SPME method optimization Recoveries and enrichment factors CLS and ILS model and data analysis Applications Effectiveness of LLME-SPME on other volatile organic contaminants in water Conclusion Chapter 7: Summary and Future Work References Appendix A: Publications Appendix B: Presentations

13 13 LIST OF TABLES Page Table 1-1. Comparison of CLS and ILS algorithms for calibration * Table 2-1. Accuracy and precision of developed method Table 2-2. The percentage recoveries of Aroclor 1260 by SPME-GC-MS Table 2-3. Application of the method to certified soil samples (n = 4) Table 3-1. FuRES, PLS-DA and FOAM Classification Rates with 95% Confidence Intervals Obtained by Using Different Numbers of Modulo Features in Modulo Compression with Baseline Correction of Training Data Using 3 Components Table 3-2. FuRES and PLS-DA Classification Rates with 95% Confidence Intervals Obtained by Using Different Data Representations with Baseline Correction of Training Data Table 3-3. FuRES, PLS-DA and FOAM Classification Rates with 95% Confidence Intervals Obtained by Using Different Data Representations with Baseline Correction Using 30 Components of the Calibration Data Table 3-4. Classifier Prediction Rates (%) of Standard Aroclor Soil Samples with No Parametric Changes Table 4-1. Concentration matrix of calibration data set Table 4-2. Comparison of PLS-1 and PLS-2 using calibration set by cross validation

14 14 Table 4-3. Linear regression coefficient ( ), RMSPE and RPEs results using different data representations Table 4-4. Comparison of prediction rates between PLS-1 and PLS-2 for certified soil samples (number of samples = 8) Table 4-5. Prediction results of certified soil samples using different data representations a by PLS Table 4-6. Comparison of prediction rates between PLS-1 and PLS-2 for real soil samples (number of samples = 3) Table 4-7. Prediction results of real soil samples using different data representations a by PLS-2 and the comparison with prediction results using EPA 8082 method Table 5-1. Paired t test results (the two tailed p-values) for the comparison of peak areas for five different PCB peaks under different conditions during headspace SPME extraction Table 5-2. Comparison of PCB distributions in different Aroclors [163] Table 6-1. Concentration matrix of modelling set (n = 3) Table 6-2. Retention indices and densities of selected organic compounds Table 6-3. Absolute recoveries and enrichment factor of TCE by LLME- SPME (n = 3) Table 6-4. Effect of normalization, RT alignment and two data representations on RMSPE by CLS Table 6-5. Predicted concentrations by weighted least squares regression and linear least squares regression for calibration set (n = 3)

15 15 Table 6-6. Prediction results of synthetic water (n = 3) and TCE contaminated river water (n = 4) with PLS regression Table 6-7. Prediction results of spiked tap water (n = 3) and TCE contaminated river water (n = 4)

16 16 LIST OF FIGURES Page Figure 1-1. Chemical structures of A: polychlorinated Biphenyls (PCBs), B: polychlorinated dibenzo-p-dioxins (PCDDs) and C: polychlorinated dibenzofurans (PCDFs) Figure 1-2. The geometry of toroidal ion trap (figure is interpreted from reference 61) [61] Figure 1-3. The principle of SPME: A, two modes of SPME (inserted and headspace); B, scheme of SPME syringe in our study; C, mechanisms for SPME Figure 1-4. Reconstructed mass spectra for TCE (top), TCE-d (middle) and stacked TCE and TCE-d (bottom) Figure 1-5. Flowchart of PLS as a multiple linear regression method. T is the latent two-way GC/MS data matrix, W is the two-way GC/MS data loading matrix, U is the latent concentration matrix, Q represents the concentration loading matrix, Ex, Ey, and Ed are residual error matrices Figure 2-1. The extraction efficiency in different extraction conditions (n = 3) Figure 2-2. The comparison of KMnO4 solution, CrO3 solution and K2Cr2O7 solution on extraction efficiency of PCBs from Aroclor 1260 in soil (A) (n = 4) and fiber contamination (B) Figure 2-3. The effect of concentration and addition volume of K2Cr2O7 and CrO3 solution on extraction efficiency of PCBs from Aroclor 1260 in soil. (n = 3)... 66

17 17 Figure 2-4. The effect of different SPME fiber on extraction efficiency (n = 3) Figure 2-5. The effect of extraction temperature (A), extraction time (B), and electron ionization energy (C) on the extraction efficiency of PCBs from soil samples spiked with Aroclor (n = 3) Figure 2-6. GC/MS total ion current (TIC) chromatograms of Aroclor 1260 before (A) and after (B) baseline/background correction for 30 ng g -1 soil sample after headspace SPME extraction. On the right (C) are EICs for tetra-, penta-, hexa-, hepta-, and octa-cbs at zoom-in retention time window Figure 3-1. Total ion current (TIC) chromatograms of solvent blank samples (A), Aroclor samples before (B), and after (C) baseline correction. PDMS peaks are designated with a label Figure 3-2. FuRES and PLS-DA prediction rates with respect to the numbers of components for baseline correction using different data sets representations. The 95% confidence intervals are given in dashed lines Figure 3-3. The minimum PDR value among 7 classes in modulo compression data sets varies with the modulo numbers Figure 3-4. Principal component analysis score plot for two-way modulo compressed (35 features) data sets of seven Aroclor standard samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 64%. The second number in the parentheses denotes the absolute variance

18 18 Figure 3-5. Prediction plot of Aroclor soil samples using FuRES (A) and FOAM (B) classifiers constructed by modulo compressed (35 features) Aroclor standard data sets. Each spot represents an Aroclor soil sample. The correct classified samples are colored in green, and misclassified samples are colored in red Figure 3-6. Principal component analysis score plot for two-way modulo compressed data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 57%. The second number in the parentheses denotes the absolute variance Figure 3-7. Principal component analysis score plot for two-way data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 47%. The second number in the parentheses denotes the absolute variance Figure 3-8. Principal component analysis score plot for one-way TIC data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 48%. The second number in the parentheses denotes the absolute variance Figure 3-9. Principal component analysis score plot for one-way TMS data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by

19 19 the ellipses. The percent variance of the first two principal components given in parentheses is 61%. The second number in the parentheses denotes the absolute variance Figure 4-1. GC/MS total ion current (TIC) chromatograms of Aroclor 1254 (A) and Aroclor 1260 (B) for 300 ng g -1 soil sample after headspace SPME extraction. On the right of each TIC chromatogram is zoom-in retention time window of PCB peaks for each Aroclor Figure 4-2. Plot of the RMSPE with respect to the number of PLS components for the EIC-Both data Figure 5-1. Photographs of the Guardion -8 GC-TMS showing A) external and B) internal components Figure 5-2. Bar charts showing the influence of (A) SPME sorption time, (B) agitation, (C) addition of KMnO4 and H2O and (D) addition of acid and base on extraction efficiency of PCB 66, PCB 153, PCB 138, PCB 180 and PCB 170 from soil, as measured on the bench-top GC/MS instrument. Error bars show ±1 standard deviation (n = 3). Significance tests are shown in Table Figure 5-3. Total ion current (TIC) chromatogram of headspace SPME of 40 µl of 100 ppm Aroclor 1260 (red) and EPA 8082A mix (blue) in the absence of soil matrix on the Torion Guardion -8 GC/MS. Peak assignments for the Aroclor mix (red) were made from the EI mass spectra and retention times but cannot exclude the possibility of congener co-elution Figure 5-4. Example of mass spectra comparison for pentachlorobiphenyl between Torion Guardion -8 GC/MS data (top) with NIST database (bottom). The chlorine isotope

20 20 distributions are identifiable in both spectra, especially around m/z 254 and Figure 5-5. Portable GC/MS chromatograms (TIC) for headspace SPME analyses of 10 µg spikes of (A) Aroclor 1016, (B) Aroclor 1232, (C) Aroclor 1242, (D) Aroclor 1248, (E) Aroclor 1254 and (F) Aroclor 1260 in the absence of soil matrix. The retention time windows of chromatograms for each Aroclor (upper chromatogram of each Aroclor) are shown from 3.1 min to 4.9 min. The lower chromatograms of each Aroclor show the same data in larger scale. (2CB: Dichlorobiphenyl; 3CB: Trichlorobiphenyl; 4CB: Tetrachlorobiphenyl; 5CB: Pentachlorobiphenyl; 6CB: Hexachlorobiphenyl; 7CB: Heptachlorobiphenyl; 8CB: Octachlorobiphenyl) Figure 5-6. Comparisons of headspace SPME calibration curves of Aroclor 1260 in the absence of soil matrix for the peak tentatively assigned as PCB 180 collected on (A) Portable Torion Guardion -8 GC/MS (R = 0.96) and (B) Bench-top Thermo Polaris Q GC/MS (R = 0.98). Note that the bench-top calibration curve covers significantly lower quantities Figure 5-7. GC chromatogram of Aroclor 1260 (top) and representative MS spectrum of PCBs (bottom) for 10 ppm Aroclor simulated soil sample Figure 5-8. Dioxin mix standards tested on portable Torion Guardion 8 GC-TMS system using headspace SPME extraction. Spike consisted of 50 µl of a 5 ppm stock solution of EPA 8280B mix in the absence of soil matrix

21 21 Figure 5-9. Dioxin mix standards tested on bench-top Thermo Polaris Q GC-MS system using headspace SPME extraction. Spike consisted of 50 µl of a 5 ppm stock solution of EPA 8280B mix with different matrices. The dry dioxin standard (top figure) was measured in the absence of any chemical modifiers or matrix. The dry dioxin-spiked soil (bottom figure), was analyzed with and without treatment with acidified KMnO4. Wet dioxin-spiked soil (bottom figure) was also extracted following treatment with KMnO4 + H Figure 6-1. Flowchart of data analysis Figure 6-2. Effect of different extraction solvent for LLME-SPME (A) and effect of hexane volume for LLME-SPME with 95% confidence intervals (n = 3) (B). TCE peak area: m/z 132 (molecular ion of TCE) was extracted from total ion current chromatogram and integrated in TCE retention time window. 164 Figure 6-3. Data points in the two factors (extraction time and extraction temperature) central composite design Figure 6-4. Response surface of the second-order polynomial model with the zoom-in window of interested region (A) and response surface model at 15 min (B) Figure 6-5. Effect of dispersive solvents (A), volume of acetonitrile (B) and extraction time (C) on LLME-SPME extraction efficiency. (n = 3) Hex: hexane; ACN: acetonitrile; MeOH: methanol. Note that in Figure 4B, volume of acetonitrile refers to additional volume of acetonitrile as dispersive solvents added to the solution, and the acetonitrile in TCE standard solution was not counted

22 22 Figure 6-6. Effects of fiber coatings (A), stirring (B) and salting out (C) with 95% confidence intervals. (n = 3) Figure 6-7. Effect of RT alignment: TCE chromatograms before RT alignment (A) and TCE chromatograms after RT alignment Figure 6-8. Example calibration curve for TCE/TCE-d ratios predicted by CLS model with respect to prepared TCE concentrations (A) and a plot of predicted concentration with respect to prepared concentration of TCE in water samples; the line is a reference line (B) Figure 6-9. Comparison of regular SPME and LLME-SPME for the extraction of toluene (A) and TCB (B). (n = 3)

23 23 CHAPTER 1: INTRODUCTION 1.1 Introduction of dioxins and PCBs Dioxins usually refer to two structure-related groups of chemicals: polychlorinated dibenzo-p-dioxins (PCDDs) and polychlorinated dibenzofurans (PCDFs), and polychlorinated biphenyls (PCBs) are described as dioxin-like compounds because of their similar structure and toxic effects (see Figure 1-1) [1]. The configurational isomers of PCBs or dioxins are referred to as congeners that depend on the numbers and positions of chlorine atoms on the benzene rings [2]. As persistent organic pollutants (POPs), PCBs and dioxins are very stable and resistant to degradation under natural conditions [3, 4]. At the same time, their lipophilic and hydrophobic properties lead to high affinity to fat tissues and cause bioaccumulation of toxins in the food chain [5-7]. To compare the toxicity of various PCDD, PCDF, and PCB congeners, toxic equivalency factors (TEFs) are used and TEF values of 29 PCB and dioxin congeners have been assigned by World Health Organization (WHO). Each TEF is determined by the relative toxicity as compared with 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD), the most toxic dioxin congener. And the sum of the products of the concentration of each congener multiplied by its TEF is the total toxic

24 24 equivalency (TEQ), which is used to estimate the total toxicity of a mixture [8-10]. Figure 1-1. Chemical structures of A: polychlorinated Biphenyls (PCBs), B: polychlorinated dibenzo-p-dioxins (PCDDs) and C: polychlorinated dibenzofurans (PCDFs). Dioxins, PCDDs and PCDFs (PCDD/Fs), have 210 different congeners (75 congeners for PCDDs and 135 congeners for PCDFs), however none of them were synthesized for commercial purposes but were unwanted byproducts of industrial processes and combustion activities [2]. Many studies have proved that the toxicity of dioxins was depended on a cytosolic protein, aryl hydrocarbon receptor (AHR) [11-14]. At the same time, it was found that the toxicity of dioxins was structurally related. The dioxin congeners with lateral chlorines, which refers to the dioxin congeners with chlorine substituted in positions 2,3,7 and 8 (see Figure 1-1), are more likely to bind AHR as well as are more abundant in the environment and organisms. So they are especially toxic, and other congeners are more easily metabolized and/or have a low binding affinity to AHR [15, 16].

25 25 Therefore, only 17 of them (7 PCDDs and 10 PCDFs) with 2,3,7,8- structure are toxicologically the most relevant and arouse the most public concern [8]. PCBs have 209 possible congeners, about 133 of which were produced intentionally as marketable products [17]. Aroclors, the trade name of products from the Monsanto Company, USA, were the major commercial PCB products of United States [18]. They were produced from 1930 to 1975 for the use in insulating oil, lubricating oil and heat medium, and different Aroclors were named as Aroclor 12xx, in which the last two digits represent the percent Chlorine by mass in the mixture. For example, the weight percentage of Chlorine in Aroclor 1260 is about 60%. But Aroclor 1016 was the only exception which contains 41% Chlorine by mass [19, 20]. The manufacture of Aroclors was banned in 1977 in North America because of PCBs bioaccumulation as toxic compounds to environmental and human health [21, 22]. PCBs were found to have potential carcinogenicity [23-28], reproductive toxicity [29-33], adverse growth and development effects [34-38], irritation and sensitization effects to skin [39-41], and so on.

26 Methods for the determination of PCBs Congener-specific method and limited congener Aroclor estimation methods As persistent organic pollutants (POPs), PCBs are very stable and resistant to degradation under natural conditions [4]. PCBs were produced and disposed of in the environment as Aroclors, so the quantification of Aroclors is important. Two methods are commonly used for the quantitation of Aroclors in environmental samples. One is the comprehensive congener-specific method (e.g., US EPA Method 1668c) [42] which requires high-resolution gas chromatography/mass spectrometry (GC/MS) to separate and determine each individual PCB congener in the sample. This method is expensive, difficult to perform, and requires expensive equipment, individual isotopic standards, and expertise that are not always available in many laboratories. The recent publication determined the concentrations of a single Aroclor in soil by summing together the concentrations of each PCB congener calculated by response factors of each PCB to PCB 209 [43]. In their method, sample preparation required 60 min and for sample analysis 65 min were required. The calculation of concentrations for each PCB congener (105 peaks and 9 response factors) was tedious and this method cannot analyze samples that

27 27 contain more than one Aroclor. Congener-specific methods are preferable in some cases because of the detailed PCB information offered. By providing concentrations of all the PCB congeners, the toxicity of each sample can be accurately estimated, even though the proportions of most toxic PCB congeners are considerably smaller than the major, less-toxic components of the Aroclors. The congenerspecific method is recommended in carcinogenic risk-assessment [44, 45]. On the other hand, the result of congener-specific method is not influenced by source and type of PCBs, the type of environmental media (air, water, soil, sediment, and biota), physical-chemical properties of the media (temperature, ph, organic carbon content), the congeners present in technical mixtures, or the type and abundance of microfauna and flora. Congener-specific methods can reflect physiological, spatial, and temporal changes. The measurement of complicated samples, which combines several Aroclors can also be achieved by using congener-specific methods. The other method is a limited congener Aroclor estimation method (e.g., US EPA Method 8082a) [46]. The Aroclors should be identified first based on the similarity of the sample chromatogram with that of the most similar Aroclor standards. Then 3 to 5 major characteristic peaks were selected to build the calibration models and

28 28 the average of the concentration results from these models was used to estimate the concentration of the Aroclor. A decision tree using PCB peak patterns (e.g., PCB118:203 for Aroclor 1260 and 1254) was required for the identification of samples with two or more Aroclors [19]. The selection of characteristic peaks is experience-based, and although the full separation of all the PCB congeners was not required, this method required a relatively long GC program ( 45 min) to separate 3 to 5 selected characteristic peaks from the other PCB congeners [46]. Most of limited congener Aroclor estimation methods cannot determine PCBs in the samples which contain more than one Aroclor Instrumentation of PCB analysis methods General instrumentations Both the limited congener Aroclor estimation methods and the congener-specific methods of PCB analysis rely on gas chromatography [19]. Gas chromatography (GC) is a powerful analytical technique to separate individual PCB congeners or combinations of congeners based on their volatility (boiling point) and polarity. Open tubular capillary GC columns rather than older packed GC columns are now in common use because of their improved resolution, and better selectivity.

29 29 After GC separation, PCBs can be detected with electron capture detection (ECD), electrolytic conductivity detection (ELCD), or mass spectrometry (MS). Since the 1960s, PCBs have been detected by GC- ECD using packed columns. Capillary GC-ECD and capillary GC/MS have been in common use since the 1980s [47]. For GC/MS systems, PCBs are identified by order/time of elution from the GC, as well as through molecular and fragment mass to charge (m/z) ratios. Highresolution mass spectrometry (HRMS) techniques can be used to distinguish between certain coplanar PCBs, which are carcinogens at low concentration levels, but normally additional sample cleanup and special instrumentation are required. For classification purpose, the mass spectrometer is more powerful because of its greater informing power compared to GC/ECD [48]. However, many PCB isomers exist in different Aroclors, so long gas chromatographic programs were required to provide enough chromatographic resolution for the identification of unique PCB congeners assigned to a specific Aroclor [49]. For Aroclor identification, manual selection of PCB peak patterns for classification of Aroclors is tedious and error prone because it depends on the skill and experience of the analyst as well as the precision of the measurement.

30 Portable GC/MS Most of methods for the analysis of PCBs are laboratory-based. For the laboratory-based methods, samples have to be transported with the risk of analyte volatilization or degradation, and these performing methods often require time-consuming sample pretreatment or expensive equipment. On-site analysis using portable instruments is a good alternative, and is cost effective [50]. Portable instrument without losing analytical performance significantly compared with bench top instrument [51] is necessary and desirable for on-site studies. Over the past 20 years, the miniature mass spectrometers have been significantly developed. Although all the instrument components are required to be miniaturized for portable instrument, miniaturization of mass analyzer is often believed the key to reduce the instrument size overall [52]. The mass analyzer is the component for mass spectrometer which accepts ions, separates them based on mass to charge ratio and outputs them to the detector, and almost all the traditional mass analyzers including sector analyzer [53, 54], linear quadrupoles [55-57], time of flight [58, 59], ion trap (IT) [60-63], and ion cyclotron resonance [64], have been miniaturized for portable instruments. By miniaturizing the mass analyzer, the mean free path

31 31 of ions can be shorter in which systems of high pressure can be tolerated and low-power pumps can be used. The power can be further reduced because the smaller volumes make lower voltages necessary to obtain the required electric field strength which in turn facilitates smaller power supplies or batteries [52]. Ion trap mass analyzers have been chosen for miniature mass spectrometers by virtue of their good pressure tolerance, tandem MS (MS 2 ) capability, simplicity and readiness for the combination with GC [51, 52, 65]. The ion storage capacity is reduced dramatically during miniaturization which more easily results in the ion-ion repulsion (space charge) effects and causes broadening of peaks in the mass spectrum. Various geometries have been explored for miniature ion traps to increase the ion storage capacity such as cylindrical ion traps [66, 67], arrays of cylindrical ion traps [68, 69], arrays of rectilinear ion traps [70, 71], halo ion trap [72], and toroidal ion trap [65]. The toroidal ion trap increases the ion storage capacity by distributing the ions within a circular storage ring. The device is augmented from conventional ion trap and can be viewed as either a conventional ion trap cross section that has been rotated on an edge through space or as a linear quadrupole curved and connected end to end (Figure 1-2) [65].

32 32 Some advantages of the toroidal ion trap as a miniature mass analyzer are as follows: 1) It remains all the advantages that ion trap has including simplicity, pressure tolerance, MS n, etc. 2) It s easier for machining and assembling than an array of ion traps. 3) Theoretically, the ion storage area would be hundreds of times larger than conventional ion trap with the same cross-section radius. The portable gas chromatography-toroidal ion trap mass spectrometer (GC-TMS) has been developed by Stephen A. Lammert and his coworkers [61], and now the commercial product, Guardion -8 which is available from Torion Technologies (American Fork, Utah, USA). This system has been successfully applied to on-site analysis of VOCs from different matrices [73, 74], trihalomethanes in water [75], chemical warfare agents [74], and other hazardous compounds [76].

33 33 Figure 1-2. The geometry of toroidal ion trap (figure is interpreted from reference 61) [61] General introduction of PCB sample extraction methods Both limited congener Aroclor estimation methods and congenerspecific methods are gas-chromatography methods, which require the extraction of PCB from the environmental matrix General introduction of PCB sample extraction methods Many classical extraction techniques have been applied such as Soxhlet extraction [42, 77, 78], liquid/liquid extraction (LLE) [42, 79], solid-phase extraction (SPE) [80-82], and pressurized fluid extraction [83-86]. However, these methods have experienced some shortcomings that limit their application to high-throughput or on-site

34 34 analysis. For example, Soxhlet extraction usually takes 16 to 24 h and uses large volumes of solvent. Even though automatic Soxhlet extractors are available, about a 2-h extraction time is still needed [87]. LLE-based methods are laborious and also require large volumes of solvent. At the same time LLE is usually not as exhaustive as Soxhlet extraction and are often coupled with other extraction technics such as solid phase extraction (SPE) [88]. SPE has been used as an alternative method to LLE for the extraction of PCBs from aqueous samples such as ground water and serum because of the smaller solvent consumption and shorter extraction time than LLE [81, 82, 89]. SPE has a few downsides such as clogging due to small particles and pore size of the sorbent in cartridges when directly performing extractions with complex-matrix samples such as soil and serum [90]. Pressurized fluid extraction requires the use of an expensive and large extraction device, which is inconvenient for on-site analysis [84, 91]. All the methods above need to use large volumes of organic solvents, which is a major concern in the light of green chemistry goals [92]. Therefore, more extraction methods with high efficiency, short time, and low cost have been developed in recent years such as vortex assisted liquid-liquid microextraction (VALLME) [79], dispersive liquid liquid microextraction (DLLME) [93], hollow-fiber liquid-phase

35 35 microextraction (HF-LPME) [94, 95], ultrasound assisted emulsification-microextraction (USAEME) [96] and solid-phase microextraction (SPME) [97-100]. In VALLME, dispersion of the microvolume level extraction solvent (organic phase) into the aqueous solution has been assisted by vortex mixing and it was able to take only 2 min to achieve equilibrium [79]. DLLME is very similar to VALLME but without vortex mixing. DLLME also requires an additional disperser solvent which can be miscible in both water and the extraction solvent such as acetone, acetonitrile, and/or methanol [101]. It has been reported that the combination of SPE and DLLME could achieve higher enrichment factors (EF) and was more suitable for the determination of PCBs even in complex matrices such as plant samples and milk [93, 102, 103]. HF-LPME works based on the principle of supported liquid membrane [92]. First, a polypropylene membrane is immersed into the organic solvent several times to immobilize the solvent in the pores of the polymer. The extraction in the sample solution is then performed followed by filling the lumen of hollow fiber with an acceptor solution. The nonpolar organic solvent forms a thin layer within the wall of the HF to exclude the aqueous solution from the lumen. After the extraction, the acceptor solution can be injected to the GC for analysis [104]. USAEME is another type

36 36 of liquid-liquid microextraction with the help of ultrasonic excitation. The ultrasound accelerates the mass-transfer process between two phases and facilitates the emulsification effects which shortens the extraction time and improves the extraction efficiency [105]. SPME was first reported by Pawliszyn and co-workers in 1989 [106], and it has been greatly developed and widely applied over the past 20 years [107]. The thin polymer-coated fiber is the key part of the SPME devices. The fiber is placed in the sample or the sample headspace for the adsorption/absorption of analytes. After reaching equilibrium (ideally, but not always), the fiber is removed from the sample and the analytes are transferred from the fiber to a chromatographic column by either thermal desorption in the hot GC injector or mobile phase or through elution, as with high pressure liquid chromatography (HPLC) [107]. SPME has some advantages compared with the traditional sample preparation techniques, such as LLE and SPE. First, it is a fast, simple, sensitive and solvent-free method. Second, SPME can integrate sampling, extraction, concentration, and sample introduction to an instrument into one step. Third, it is compatible with major separation systems such as GC, HPLC, and capillary electrophoresis [108]. Forth, commercialized fibers with various coating and combinations [i.e., PDMS, polyacrylate

37 37 (PA), carboxen (CAR), carbowax (CW) and divinylbenzene (DVB)] are available and more fiber coatings such as polypyrrole (PPY) [109], poly(phthalazine ether sulfone ketone) (PPESK) [110, 111] and polyurethane (PU) [112] foams have been developed. SPME is easy for automation and several instrument companies such as Thermo Scientific, Agilent have GC instruments with SPME autosamplers. Fifth, the device for SPME is small which is convenient for portable instrument and field/on-site analysis SPME phase selection The mechanisms of SPME are both adsorption and absorption, and the relative proportion of each depends on the phase material and analytes. As the only manufacture, Supelco supplies SPME fibers with various single or mixed polymer materials [113]. For the liquid-coated fibers like PDMS, the analyte molecules can partition and penetrate the entire coating phase within a certain extraction time. For solid-coated fibers like polyacrylate, the analyte molecules are very difficult to diffuse into the coating phase because of the complex crystalline structures [114]. Therefore, absorption fibers have larger extraction volume which means a wider dynamic range but a longer time to reach equilibrium compared with adsorption fibers [108]. The interactions between analytes and the materials of fiber also follow the principle of

38 38 like dissolve like. For example, PDMS (nonpolar) coating fiber provides high extraction efficiency for nonpolar compounds whereas PA (polar) coating fiber has more applications for polar compounds such as phenols and alcohols [115]. Typically, SPME can be used in inserted mode or headspace mode (Figure 1-3). In inserted mode, the fiber is completely immersed in the liquid samples; in headspace mode, it is exposed to the vapor phase above the liquid or solid samples. For complex mixtures, headspace SPME is more preferable because it can protect the fiber from damage or carryover. Under headspace mode, the mass transportation of analytes from sample to headspace and then to fiber coating is easier for volatile compounds. So, headspace SPME typically works better for analytes of high-to-medium volatility and low-to-medium polarity [116]. The most important parameters for optimizing SPME method development include fiber coating, extraction mode, agitation method, sample volume, water/organic solvent composition, ph, extraction temperature, extraction time, ionic strength, desorption condition, and sometimes sample derivatization.

39 39 Figure 1-3. The principle of SPME: A, two modes of SPME (inserted and headspace); B, scheme of SPME syringe in our study; C, mechanisms for SPME. SPME has been applied to the determination of PCBs in different matrices such as water [117], soil [118], ash [119], and tissues [120]. All of these applications focused on the quantitative analysis of selected PCB congeners rather than modeling the Aroclors directly Other aspects of PCB analysis methods Typical detection limits For limited congener Aroclor estimation methods, method detection limits (MDLs) in the literature for Aroclors vary in the range of to 0.9 mg/l in water and 57 to 70 ng/g in soils, with higher (worse) MDLs for the more heavily chlorinated Aroclors. For congener-

40 40 specific methods, MDLs of the GC-MS methods can be low to from parts per billion to parts per trillion Cost The limited congener Aroclor estimation methods are relatively inexpensive in comparison to congener-specific methods. The cost on a per sample basis ranges from $50 to $500 for the limited congener Aroclor estimation methods, and from $500 to $2,000 for congenerspecific methods depending on instrumentation used, the sample matrix, number of samples to be analyzed, and extent of qualityassurance-quality-control required. 1.3 Introduction of TCE Trichloroethylene (IUPAC name: trichloroethene, TCE) is a volatile compound that has been used for over 100 years as a chemical intermediate, anesthetic, ingredient in food processing, dry cleaning agent, and metal degreaser [ ]. Some studies have shown that the symptoms such as central nervous system (CNS) depression, reduced fertility, abortions, cardiac arrhythmias, and some cancers could be caused by exposure to TCE [ ]. TCE has been found extensively in the environment, especially in ambient air and water due to its widespread use. Previous studies have shown that 42% of marine sediments obtained from the US EPA s Superfund sites

41 41 and 9~34% of US drinking water supplies are contaminated with TCE [121, 128]. At least 60% of hazardous waste sites on the US EPA's National Priorities List (NPL) contain TCE [129]. Therefore TCE is considered as the most frequently reported organic contaminant in groundwater [130]. Gas chromatography (GC) based methods are the most widely used for the determination of TCE including GC coupled with either electron capture detectors (ECD) [ ], mass spectrometers (MS) [ ], or flame ionization detectors (FID) [138]. Traditional sample preparation methods for TCE such as liquid-liquid extraction and solid phase extraction are labor-intensive and time-consuming [132]. Solid-phase microextraction (SPME) is suitable for the extraction of volatile organic compounds (VOCs) and has been introduced for the analysis of TCE in recent years [133, 134, ]. As a volatile compound, TCE is generally extracted from the sample headspace with faster extraction times and improved selectivity than direct extraction [107]. The condensed phase, the headspace gas phase, and the SPME polymer film are involved in a regular headspace SPME process and the diffusion of analytes happens across two interfaces, the condensed/gas interface and the gas/polymer interface [139]. The extraction efficiency is limited by

42 42 mass transfer between the two interfaces, especially between the condensed/gas interface [139]. Once the extraction conditions such as extraction temperature, extraction time, sample agitation, ph, ionic strength, volume, etc.; for an analyte are optimized, it is difficult to improve the extraction efficiency further. In my dissertation, liquidliquid microextraction (LLME) was used to assist the headspace SPME for the first time. The analytes in organic phase are more concentrated than without using LLME, which greatly benefits the mass transfer between condensed/gas interface and further improves the extraction efficiency of the headspace SPME. For TCE and other VOCs quantification, internal standards (IS) are used and recommend by US Environmental Protection Agency (US EPA) Method 8260C [140]. Isotopic analogs of the analytes with three or more 2 H-atoms (or less commonly, 13 C-atoms) at appropriate positions are considered the most effective and commonly used ISs for GC/MS methods [141]. However, usually the number of isotopic labels for a molecule will increase the cost of the standard reference material. Additionally, at least 15 of the VOCs listed in Method 8260 cannot have three or more 2 H-atom analogs such as TCE, malononitrile and bromoform because of the limited number of H-atoms in the structure [140]. Using one or two 2 H-atoms labeled analogs as ISs

43 43 can cause a cross-contribution problem which is the contribution to the intensities of ions designated for the counter-component of the analyte or ISs (Figure 1-4) [141]. This problem is especially serious for multiply chlorinated and/or brominated VOCs because stable isotopes of chlorine and bromine give broad isotopic peak clusters in the mass spectra. However, in this situation the overlapping peak clusters of the analyte/ and ISs can be resolved by multivariate chemometric methods such as classical least-squares (CLS) and inverse least-squares (ILS). In my study, CLS and ILS were applied to overcome the cross-contribution problem between analyte (TCE) and its isotopic ISs (deuterated TCE) for the first time.

44 44 TCE mass spectrum Response Response Response m/z TCE-d mass spectrum m/z TCE and TCE-d stacked mass spectra TCE mass spectrum TCE-d mass spectrum m/z Figure 1-4. Reconstructed mass spectra for TCE (top), TCE-d (middle) and stacked TCE and TCE-d (bottom). 1.4 Introduction of selected chemometric methods Chemometric methods for Aroclor identification Chemometric methods especially pattern recognition including soft independent modeling by class analogy (SIMCA) [142], partial least-squares discriminant analysis (PLS-DA), K-nearest neighbor (KNN) [143], probabilistic neural network (PNN) [49], and sequential classification method [144] for classification of Aroclor samples with GC/ECD and GC/MS data have been used because of their superb capability to obtain categorical information from complex data sets. None of these methods have been applied to soil samples. Of pattern

45 45 recognition methods, the fuzzy rule-building expert system (FuRES) has been developed and utilized to construct reliable and robust classifiers for many applications [ ]. The FuRES algorithm was described in the early 1990s [146]. PLS-DA is used as a second reference method because of its widespread use for classification. FuRES has three key advantages: (1) interpretation of classification results from inductive classification trees; (2) no adjustable parameters that require optimization; (3) the nonlinear fuzzy logistic function accommodates overlapping classes and fits better with the target categories than other methods that attempt to model discrete data with a continuous linear function [146, 149]. The fuzzy optimal associative memory (FOAM) is another powerful fuzzy classifier [150]. The difference between FuRES and FOAM is that, FuRES is a classification method which is better at tweezing out the differences of the features among objects that belong to different classes; but FOAM is a modeling method which exploits the similarities of the features within a class, is softer, and can be used when only one class is known or present [149] The FuRES classifier For the FuRES classifier an inductive classification tree is constructed from fuzzy multivariate rules. Each rule is a branch of the

46 46 tree that partitions data objects based on a fuzzy logistic value (i.e., consequent of the rule). Each data object is projected onto a normalized weight vector so that the objects are mapped from a multidimensional space to a single dimension. The scalar projections are then processed by a fuzzy logistic function. The fuzzy logistic values are then used to measure the entropy of classification. The orientation of the weight vector is optimized to achieve the lowest fuzzy classification entropy while constraining the fuzziness through a computational temperature that maximizes the magnitude of the first derivative of the fuzzy entropy with respect to temperature. The objects are partitioned by their fuzzy logistic values until each leaf of the tree consists of objects with the same class designation. Because of the inherent fuzziness of the FuRES classifier, it can be used when classes are overlapped or contain outliers [146] The PLS-DA classifier PLS-DA is commonly used for pattern recognition. The latent variables which are transformed from the independent X-data are used for regression with a dependent variable Y. The latent variable which has maximum covariance between the X- and the Y-scores is selected and then the variance of this latent variable is removed by deflation. From the residual matrix, the next latent variable is derived and the

47 47 variance is removed in the same way. This procedure is continued until the best prediction rates are achieved using an internal bootstrapped Latin partition [149]. The PLS-DA classifier is used as a reference method in this study The FOAM classifier FOAM is an optimal associative memory (OAM), except fuzzy encoding of the data is used. The data are first encoded as a binary image as opposed to a vector using grid encoding. A fuzzy function is then applied to the binary grid. The FOAM method has three steps, first binary encoding by a gridding function, reconstruction with an orthogonal basis, and then decoding to vector format by reversing the gridding procedure. SVD is used to form the orthogonal basis set from fuzzy encoded data. The basis is built for each class. By comparing the least-squared error between the reconstruction with each basis and the original data an object can be assigned to the best fitting class [150]. The standard grid size is 100 grids between the maximum and minimum intensities of the mean-centered calibration set. The fuzzy function was a triangular membership function with 19 elements that ranged between 0.1 and 1.0 with a 0.1 increment. This function is applied to each variable with respect to intensity although other

48 48 applications might apply across the variables or use a 2-dimensional function. For forming the basis of grid values all the components are used to fully characterize the data. For the inverse grid process, a data point is assigned to the intensity of the maximum grid value for each variable Chemometric methods for Aroclor quantification The partial least squares (PLS) regression method which is accomplished via a small set of latent variables (the components) is a powerful chemometric technique for multivariate calibration when the data are highly correlated [151]. As a full spectrum multivariate calibration method, PLS uses both spectral and concentration data to construct a calibration model. Only the concentrations for the analytes of interest (e.g., Aroclor 1254 or 1260) are necessary for model construction and other interferences or uncorrelated analytes (e.g., uncorrelated PCBs from other sources) have little effect on the calibration sets [152]. This advantage is important for the quantification of PCBs as Aroclors because it enables the rapid and simultaneous determination of different Aroclors in the mixture without time-consuming separation to resolve the overlapped PCB congener peaks and with minimum sample preparation. For the same reason, experience based peak selection and tedious calculation of individual

49 49 peak areas can be avoided. Another advantage of PLS is that the complexity of the model can be controlled by the number of components which is convenient to approach the maximum prediction performance and avoid overfitting [152] Partial least-squares (PLS) regression PLS is one of the most widely used methods in chemometrics for multivariate calibration which was developed around 1975 by the statistician Herman Wold and then introduced into chemometrics by Svante Wold [ ]. PLS is a method to relate a matrix X to a vector Y or to a matrix Y. Generally, there are three steps: (1) calculate the mutual latent variables for both X matrix (e.g., two-way GC/MS data) and Y matrix (e.g., concentrations for Aroclors) based on maximum covariance; (2) generate the diagonal regression matrix for all PLS components; (3) predict the concentrations of calibration samples (cross validation) or/and unknown samples (test set) by retrospective calculation using the diagonal regression matrix and new X matrix [156]. Figure 1-5 is the flowchart for the PLS method.

50 50 Figure 1-5. Flowchart of PLS as a multiple linear regression method. T is the latent two-way GC/MS data matrix, W is the two-way GC/MS data loading matrix, U is the latent concentration matrix, Q represents the concentration loading matrix, Ex, Ey, and Ed are residual error matrices Root mean square of prediction error (RMSPE) The quality of calibration by PLS can be determined by RMSPE. A cross-validation method was used for the calibration data set. The prediction set contained samples that were never used for building the PLS calibration model. The RMSPE is given as,, / 1 1 for which m is the number of data objects, ny is the number of species,, is the jth concentration for the ith data object, and, is the corresponding predicted concentration Coefficient of determination The coefficient of determination ( ) is calculated as follows:

51 for is the mean value of the concentration matrix y Relative predictive errors (RPEs) The RPEs can be calculated as: % The standard deviation of y is given by Chemometric methods for TCE quantification CLS and ILS have been widely used for the qualitative and quantitative analysis of spectra from mixtures with overlapping features such as near-infrared, UV, and Raman measurements since 1970s [157]. Both CLS and ILS methods are multivariate linear regression techniques. The algorithms for CLS and ILS are listed in Table 1-1. CLS minimizes the squared errors with respect to the spectra and ILS minimizes the squared errors with respect to the concentrations during calibration Data preprocessing methods Appropriately preprocessing the data before applying the pattern recognition or linear regression methods is equal or even more important because it can eliminate non-relevant variations, decrease

52 model complexity, and provide appropriate data sets for further data processing [158]. 52 Table 1-1. Comparison of CLS and ILS algorithms for calibration * CLS algorithm ILS algorithm Step 1. Expression between concentration matrix and spectra matrix : + Step 2. Calculate the leastsquares estimate of : Step 1. Expression between concentration matrix and spectra matrix : + Step 2. Calculate the leastsquares estimate of : Step 3. Estimate the unknown sample concentrations: Step 3. Estimate the unknown sample concentrations: * is the coefficients for CLS; is the coefficients for ILS; is the unknown spectra matrix; and are the spectral and concentration errors or residuals not fit by the model Baseline correction Two baseline correction methods were available from previous studies. Both methods reconstruct a best fitting background mass

53 53 spectrum by using orthogonal bases constructed from mass spectra of the baselines (i.e., regions where there are no analytical chromatographic peaks). The earlier approach used mass spectra collected at the end of each chromatogram where no chromatographic peaks had eluted and column bleed was the highest [147, 159]. A later approach used the mass spectra to construct the basis from the entire chromatogram of a solvent blank [149]. The current project employed the approach of modeling the blank measurements in this case the sampling of an empty vial or a blank soil sample. This approach has the advantage of removing artifact peaks that arose from the SPME fiber, column bleed, and from the matrices (i.e., soil). Both approaches use singular value decomposition (SVD) to obtain an orthonormal basis that is used to reconstruct the GC/MS baseline. The number of components selected for the basis can significantly affect the baseline correction results especially when the blank GC/MS data sets contain artifact peaks. In the study of the classification and quantitation of Aroclors, several PDMS chromatographic peaks existed in both the blank and sample GC/MS data sets, so this baseline correction method was used for all the sample data before any other data preprocessing or processing. To investigate the effect of the number of components on baseline

54 54 correction for FuRES and PLS-DA classification, different numbers of components for baseline correction were selected and evaluated. The mass spectra from the chromatogram of a sample were projected onto the bases that were constructed from the mass spectra of the blanks and the basis that had the best fit (i.e., lowest total sum of square error) was selected to correct that sample. Each mass spectrum was projected onto the basis and used to reconstruct a background mass spectrum that was then subtracted from the sample spectrum to correct the background contribution. This procedure was applied to each mass spectrum in the chromatogram. The baseline correction method used SVD to create the bases. The number of components (vectors) of the basis can affect the baseline correction results. Subtraction of a reconstructed background spectrum from too many components can to a large extent remove background features but also increases the risk of overfitting the data and generating negative peaks after baseline correction Normalization Normalization is used to remove or minimize the effect of variable concentration of the samples and the systematic variations due for instance to varying amounts of analytes or variation in the

55 55 mass spectrometer sensitivity. All data were normalized to unit vector length [149] Modulo compression For mass spectral features, a modulo compression method, mod- 14 feature, was developed and applied for the prediction of molecular substructures in 1968 [160]. Condensed mass spectra so-called by the authors comprised 14 features which indicates the class/type of molecule. These features are characteristic for classes of homologous compounds whose masses differ by multiples of 14 Th accounting for the loss of methylene fragments from the ion. Modulo compression was applied after baseline compression in classification of Aroclors study. The number and position of Chlorine atoms on the benzene rings are the most important characteristic features for PCB congeners. So the molecular masses of the Chlorine isotopes (35 or 37 Th) would be reasonable to characterize these chlorinated compounds. For two-way GC/MS data sets, 411 mass spectral peaks were compressed to 35 or 37 features using this compression. The features are defined as follows: 1 4 0,1,2, ;,..,35 37

56 56 for which is the intensity of m/z ( +35 ), is the divisor and values with same remainder (i.e., modulus) are added together. Different divisors were evaluated by the PDR method, and selected divisors were used to evaluate their effects on classification rates. Because the modulo method sums the ion intensities, a signal noise improvement is obtained as with any other signal averaging approach. The number of features or divisor in the modulo compression was evaluated with the projected difference resolution (PDR) method. The effects of modulo compression preprocessing on classification rates obtained from the FuRES, PLS-DA, and FOAM classifiers and on several data set configurations were evaluated Projected difference resolution method Projected difference resolution (PDR) had been used successfully for selecting the optimal parameters for baseline correction [159], comparing the performance of GC/MS, gas chromatography-differential mobility spectrometry (GC-DMS) data [161], and optimizing wavelet filter types and the compression level for the discrete wavelet transform [162]. The PDR method measures the separation of two classes in multivariate data space and gives a figure or merit that resembles chromatographic resolution. First, objects are converted from vectors

57 57 to scalars by projecting the objects (i.e., vectors) of the two classes onto the difference vector of the two class averages. Then, the absolute value of the difference between the averages of the scalar projections is divided by two times the sum of the standard deviations of the scalar projections [159]. The larger the PDR value, the more separated the two classes are in the multivariate data space. In a previous study, the geometric mean was used for the assessment of overall difference of multiple classes [159]. However in many cases, the two most similar classes among all classes are the most difficult to differentiate and their separations are crucial for the construction of the classifiers rather than classes that are well separated in the data space. Thus, the minimum PDR value of all the pairwise combinations was used in the study of classifying Aroclors to evaluate the divisors used in modulo compression.

58 58 CHAPTER 2: DETERMINATION OF AROCLOR 1260 IN SOIL SAMPLES BY GC/MS WITH SOLID PHASE MICROEXTRACTION In this chapter, a fast method for the determination of Aroclor 1260 in soil matrices using headspace SPME-GC/MS was developed. The optimization of headspace SPME is discussed. The sample preparation and analysis was accomplished within 35 min. The total peak areas of tetra-chlorinated biphenyls (tetra-cb, m/z 292), pentachlorinated biphenyls (penta-cb, m/z 326), hexa-chlorinated biphenyls (hexa-cb, m/z 360), hepta-chlorinated biphenyls (hepta-cb, m/z 394), and octa-chlorinated biphenyls (octa-cb, m/z 430) were used to construct the calibration models for the quantification of Aroclor The method was then validated with certified soil samples. 2.1 Experimental Reagents Aroclor 1260 standard at a concentration of 100 µg/ml in methanol as stock solution was obtained from AccuStandard, Inc. (New Haven, CT). The standard solutions of Aroclor 1260 with concentrations of 0.1, 0.3, 1, 3, and 10 µg ml -1 were prepared by dilutions of aliquots of the stock solutions with methanol. A mixture in hexane containing 1 mg ml -1 of decachlorobiphenyl (deca-cb) and of tetrachloro-m-xylene (TCMX) was also obtained from AccuStandard,

59 59 Inc. (New Haven, CT). Potassium permanganate, potassium dichromate, the SPME fibers coated with polydimethylsiloxane (PDMS, 7 µm or 100 µm film thickness), 20-mL headspace glass vials and crimp seals with PTFE/silicone septa were purchased from Sigma- Aldrich Co. LLC. (St. Louis, MO). The clean soil and certified Aroclor 1260 soil samples were purchased from RT Corp (Laramie, WY). The standard soil samples were prepared by thoroughly mixing 50 µl standard solutions with 0.5 g clean soil and completely evaporating the solvent in a hood at room temperature. The internal standard solution containing 10 µg ml -1 of deca-cb and TCMX was prepared by dilution with hexane from the 1 mg ml -1 stock solution, but only deca-cb was used as an internal standard for the Aroclor quantification. A saturated potassium dichromate solution was prepared by dissolving an excess of potassium dichromate in 6.0 M sulfuric acid Instruments All the experimental data were collected on a Thermo Finnigan PolarisQ quadrupole ion trap mass spectrometer/trace GC system with a Triplus AS2000 autosampler (San Francisco, CA, USA). The GC/MS system was controlled by the XCalibur software version provided by Thermo. The GC separation was accomplished on a SHRXI-5MS

60 60 capillary column (5% diphenyl/95% dimethylpolysiloxane cross-linked, 30 m 0.25 mm i.d., 0.1 µm film thickness) from Shimadzu Scientific Instruments Inc. (Columbia, MD). MATLAB R2012b (MathWorks Inc., Natick, MA) was used to process data. The RAW files of two-way GC/MS data sets initially acquired were converted to the network common document format (CDF) with the File Converter Tool in the XCalibur Software. The CDF files were read directly into MATLAB using netcdf tools Sample preparation Soil samples of 0.5 g were added to the 20-mL SPME vial and spiked with 20 µl of internal standard solution. The samples were left in the hood at room temperature to evaporate the solvent. Then 2 ml of saturated potassium dichromate solution was added to the vial and the vial was sealed with PTFE/silicone septa using crimp seals. After 30 s of vortexing, the mixtures were placed in the autosampler tray for analysis. The sample vial was incubated at 100 C for 0.5 min. A PDMS fiber was then exposed to the headspace for 30 min. The agitation was sequentially pulsed on for 10 s and then off for 10 s for the 30 min exposure. The fiber was thermally desorbed in the GC injector at 280 C for 5 min to prevent carryover. The analytes were separated using the

61 61 following oven temperature program at a constant helium flow of 1 ml min -1 : 50 C, hold for 1 min, ramp at 20 C min -1 to 280 C, hold for 10 min. The transfer line and ion source temperatures were both maintained at 280 C. The mass spectrometer was operated in positive ion electron ionization (EI) mode at 40 ev and mass spectra were collected after a 4 min solvent delay. Full scan mode was selected for the mass spectrometer and the scan range was from mass-to-charge ratio (m/z) 140 to 550. Five blank soil samples without any Aroclor and internal standard were treated in the same way. The blank soil sample data were used for correcting the baselines of the Aroclor soil samples. 2.2 Results and discussion Aroclor 1260 soil samples at the concentration of 30 ng g -1 (ppb) were used to evaluate the optimization of SPME and instrument conditions. The peak areas of hexa-chlorinated biphenyls (hexa-cbs) were selected as references to compare the effects of different conditions because hexa-cbs are one of the major PCBs in Aroclor 1260 (46.9 weight %) [163]. Extracted ion monitoring at m/z 360 was used to quantify the hexa-cbs by integrating the peak areas in the extracted ion chromatogram from to min.

62 62 Figure 2-1. The extraction efficiency in different extraction conditions (n = 3) Optimization of SPME conditions Directly exposing a PDMS fiber to the headspace of a vial containing 0.5 g 30 ppb Aroclor 1260 soil sample demonstrated that PCBs were unable to transfer to the headspace efficiently (Figure 2-1). The low efficiency was attributed to the low boiling point and lipophilicity of PCBs, which caused sorption of the PCBs to the surfaces of the soil particles. Elemental sulfur (S6 and S8) is another common

63 63 interference in the soil matrix which could significantly decrease the extraction efficiency of PCBs [118]. R. Montes et al. have demonstrated that the employment of strong oxidative conditions such as the addition of potassium permanganate solution (0.1 M in 6 M sulfuric acid) to the soil assists in the release of the PCBs from the soil and the removal of organic matter and sulfur interferences [118]. In this study, another two strong oxidants, potassium dichromate and chromium trioxide in 6 M sulfuric acid, were compared with potassium permanganate in 6 M sulfuric acid. All the parameters in these initial studies were the same as section except that the EI energy was set to 70 ev instead of 40 ev, and internal standards were not added to the samples. Three extraction solution systems were compared: 1) 2 ml KMnO4 (0.1 M), 0.5 ml H2SO4 (6 M); 2) 2 ml 0.2 M CrO3 in 6 M H2SO4; 3) 2 ml 0.2 M K2Cr2O7 in 6 M H2SO4. The extraction efficiency of the KMnO4 system was significantly higher (average about 2.5 times higher) than the other two extraction systems, but the repeatability was significantly worse based on four replicate extractions (Figure 2-2 A). These preliminary SPME experiments were accomplished with 10-mL SPME vials and it was found that the fiber was fouled by the oxidative conditions, which may have accounted for the poor repeatability. In an

64 64 attempt to mitigate the oxidative fouling problem, 20-mL vials were used to create a larger volume for headspace for the extractions. However, after approximately 20 analyses, the PDMS fiber still turned black, which indicated that even the larger headspace volume was not able to prevent the fiber from being contaminated by the KMnO4 (Figure 2-2 B). The use of CrO3 and K2Cr2O7 offered stable extraction efficiencies and less degradation of the SPME fibers. Other treatments such as with a strong basic solution (10 M NaOH) or a strong acidic solution (10 M H2SO4) or single addition of water were investigated, but failed to effectively release the PCBs from the soil to the headspace (Figure 2-1). The effect of CrO3 vs K2Cr2O7, effect of concentration of K2Cr2O7 in 6 M H2SO4, and the effect of solution volume on absolute recoveries were also studied. The responses obtained from different solutions with a 0.5 g soil sample are given in Figure 2-3. There was no significant difference between different extraction systems (p = 0.15 by one-way analysis of variance). Therefore, K2Cr2O7 was chosen as the extraction solution because of its availability. The concentrations and amounts of K2Cr2O7 solution added to the sample had no

65 65 Figure 2-2. The comparison of KMnO4 solution, CrO3 solution and K2Cr2O7 solution on extraction efficiency of PCBs from Aroclor 1260 in soil (A) (n = 4) and fiber contamination (B).

66 66 Figure 2-3. The effect of concentration and addition volume of K2Cr2O7 and CrO3 solution on extraction efficiency of PCBs from Aroclor 1260 in soil. (n = 3) significant effect on the extraction efficiency. Saturated K2Cr2O7 in 6 M H2SO4 was chosen to oxidize organic matter to the largest extent and the amount of solution was set to 2 ml instead of 4 ml to create more headspace and prevent SPME fiber degradation by the extraction solution. The coatings of SPME fiber were selected between 7 µm PDMS and 100 µm PDMS because PDMS has a higher affinity for PCBs than the other coatings in previous studies [117, 118, 164, 165]. To evaluate the influence of the fiber thickness, both fibers were exposed to the headspace at 100 C for 30 min, and the 100 µm PDMS fiber

67 was chosen for further study because the signals were approximately three times better (Figure 2-4). 67 Figure 2-4. The effect of different SPME fiber on extraction efficiency (n = 3).

68 CResponse (area counts) 68 A Response (area counts) B Response (area counts) Extraction temperature ( C) Extraction time (min) EI energy (ev) Figure 2-5. The effect of extraction temperature (A), extraction time (B), and electron ionization energy (C) on the extraction efficiency of PCBs from soil samples spiked with Aroclor (n = 3)

69 69 After SPME extraction, the desorption of the fiber was accomplished in the GC injection port at 280 C (the maximum operation temperature for 100 µm PDMS fiber) for 5 min to avoid the carry-over effect. The absence of carryover was also validated by a system blank injection after each sample analysis. The effects of extraction temperature and extraction time on the hexa-cbs extracted by headspace SPME with 100 µm PDMS fiber were investigated. Soil samples of 0.5 g were extracted by 2 ml saturated K2Cr2O7 in 6 M H2SO4 for 30 min at 25, 60, and 100 C and responses obtained plotted with respect to temperature as given in Figure 2-5 A. The mobility of the PCBs through liquid and gas phases was significantly improved with the increase in extraction temperature, so the responses obtained at 100 C were much larger than the responses at the other two lower temperatures. Finally, 100 C was selected as the extraction temperature. Equilibrium was not achieved, so even higher temperatures may increase mass transport, but could exceed the pressure safety limits of the SPME vial. The extraction time profiles for 5, 15, 30, and 60 min at 100 C are given in Figure 2-5 B. The adsorption of PCBs to the fiber had not equilibrated after a 30 min extraction. To keep the analysis within a reasonable time, the extraction time was fixed at 30 min.

70 GC/MS analysis To develop an efficient method, a full separation of all 209 possible PCB congeners was not attempted. A 22 min GC temperature program was used in this study which was reported earlier [166]. About 40 TIC peaks can be separated for Aroclor 1260 (Figure 2-6). Each chromatographic peak may contain multiple co-eluted PCBs. Full scan mode was used for MS and the mass scan range was from m/z 140 to 550 because most of the ions for the PCB mass spectra are larger than m/z 145. The effects of EI energy on signal response were evaluated at electron energies of 15, 40, and 70 ev. Very low responses were obtained by using an EI energy of 15 ev, and an EI energy at 40 ev gave a factor of two times better response than using an EI energy of 70 ev (Figure 2-5 C). Therefore, the EI energy was set at 40 ev for further work in this study.

71 71 Figure 2-6. GC/MS total ion current (TIC) chromatograms of Aroclor 1260 before (A) and after (B) baseline/background correction for 30 ng g -1 soil sample after headspace SPME extraction. On the right (C) are EICs for tetra-, penta-, hexa-, hepta-, and octa-cbs at zoom-in retention time window Analytical method performance The data sets were pretreated by orthogonal baseline correction (using bases of 10 components) for which the GC/MS baseline/background was reconstructed from orthonormal bases constructed from the blank SPME runs. The full details of baseline correction are described in a previous study [166]. Using this approach, the artifact peaks (e.g., PDMS peaks) and baseline were thereby significantly reduced in the TIC chromatograms (Figure 2-6 B). The correction method was less effective for extracted ion

72 72 chromatograms (EICs), because the EICs are relatively indifferent to PDMS fragment ions from column bleed and the SPME fiber. Although the internal standard solution contained both TCMX at ~9.5 min and deca-cb at ~19.5 min, only deca-cb was used as internal standard because of its closer structure and chemistry to the PCBs of interest, and because TCMX eluted early and overlapped with some of the matrix peaks. Each chromatogram was normalized to the area of the deca-cb peak. To determine PCBs as the Aroclor, the sum total peak areas of multiple PCBs were used. EICs at retention time windows between 11.0 to 16.0 min were used to construct smaller two-way GC/MS data sets which were selective for the PCBs. The molecular ions of tetrachlorinated biphenyls (tetra-cb, m/z 292), penta-chlorinated biphenyls (penta-cb, m/z 326), hexa-cb (m/z 360), hepta-chlorinated biphenyls (hepta-cb, m/z 394), and octa-chlorinated biphenyls (octa-cb, m/z 430) were used to create EIC two-way data (Figure 2-6 C). All five PCBs mentioned above represent more than 99% of the PCBs in Aroclor 1260 [163]. The proposed method resulted in a linear dynamic range of ng g -1 of Aroclor 1260 with a regression equation y = 73.38x (R 2 = ). The accuracy and precision of the method were

73 73 evaluated by the prepared soil samples at three different concentrations with three replicates at each concentration. As reported in Table 2-1, the accuracy is in the range of 0 to 0.9% expressed by the relative error (RE, %) and the precision is in the range of 4.6 to 12.6%, expressed by relative standard deviation (RSD, %). In this study, the limit of detection (LOD) was calculated from three times the standard deviation of the blank signal [167]. Five blank soil samples with internal standard were treated the same as described in section The predicted concentrations for blank samples were calculated. Then three times the standard deviation of predicted concentrations was taken as LOD, which yielded a value of 5.2 ng g -1.

74 74 Table 2-1. Accuracy and precision of developed method Aroclor 1260 Measured Mean Accuracy concentration concentration concentration (RE, %) (ppb) (ppb) (ppb) Precision (RSD, %) Recovery evaluation of SPME-GC/MS method To evaluate the recovery of the SPME method, another calibration data set using standard liquid injection was collected. All the instrumental parameters were the same as those described in section 2.2.2, except that the injection mode was changed from SPME mode to splitless liquid injection mode. A set of 0.5 g Aroclor 1260 soil samples at the concentrations of 10, 30, 100, and 300 ng g -1 in five replicates were collected using the SPME-GC/MS method. The

75 75 calibration mode was constructed using EIC data and the mass of Aroclor 1260 extracted by SPME was determined. The SPME peak areas were compared to those of the liquid injection calibration curve to assess the gram quantity on column. The percent recovery was calculated using the calculated mass on-column of the SPME-extracted sample relative to the absolute mass contained within the vial before SPME extraction. The results are listed in Table 2-2. The recoveries ranged between 55-66% for the SPME samples at the four concentrations studied. The results are not surprising if one considers that the adsorption equilibrium between PCBs and the SPME fiber was not properly established within 30 min (Figure 2-5 B). However, the low recoveries did not affect the accuracy of the method Validation of method by certified soil samples Certified soil samples originally contaminated with Aroclor 1260 at 1.50 µg g -1 (prediction interval µg g -1 ) were measured using the previously optimized conditions. The certified soil samples were diluted with certified clean soils to the concentrations at 50 ng g -1 and 500 ng g -1. Each soil sample was analyzed by SPME-GC/MS in four replicate trials. As given in Table 2-3, the estimated concentrations are inside of the certified prediction intervals.

76 Table 2-2. The percentage recoveries of Aroclor 1260 by SPME-GC- MS. Aroclor1260 Aroclor1260 Aroclor1260 Mean RSD concentration in the vial on column recovery (%) (ppb) (ng) (ng) (%)

77 77 Table 2-3. Application of the method to certified soil samples (n = 4). Prediction interval (certified reference value) Concentration found (ppb) (ppb) (500) 550± (50) 60± Conclusions A fast Aroclor-based quantitative method for PCBs in soil samples by SPME-GC/MS has been proposed in this study. The combination of potassium dichromate and sulfuric acid solution was used to extract PCB from soil for the first time, and the parameters for the non-equilibrium headspace SPME were optimized. The extracted ion two-way (EIC) data sets were used to construct calibration curves and the method has been validated by commercial certified soil samples. The predicted concentrations of Aroclor 1260 were all in the prediction intervals for the certified soil samples. The proposed method has the advantage of the high sample throughput, with a soil sample being prepared and analyzed about every 35 min. The headspace SPME method is easy to perform and has the potential to be adapted for onsite analysis. Other preliminary studies has demonstrated its application to the field study combined with a portable GC/MS instrument [168]. The method required a low sample

78 78 amount (0.5 g) which can benefit applications for which sample availability is a limiting factor.

79 79 CHAPTER 3: AUTOMATED PIPELINE FOR CLASSIFYING AROCLORS IN SOIL BY GAS CHROMATOGRAPHY/MASS SPECTROMETRY USING MODULO COMPRESSED TWO-WAY DATA OBJECTS In this chapter, a classification pipeline for identifying Aroclor 1016, 1221, 1232, 1242, 1248, 1254, and 1260 samples was developed. The Aroclors are complex mixtures comprising many of the same PCB congeners so identifying these mixtures is a challenging problem, however GC/MS has exceptional informing power to meet the demands of this problem. The Aroclor standards were sampled by solid phase microextraction (SPME) of the headspace in the vials. The Aroclor constituents were separated by GC using a 22-min temperature program and detected with an ion trap mass spectrometer. With the application of modulo compression to the twoway GC/MS data, three classifiers (FuRES, PLS-DA, and FOAM) were constructed and used to classify the GC/MS data into one of 7 Aroclors categories. These classifiers were then applied to data that were collected by measuring soil samples that were spiked with the Aroclors.

80 Materials and methods Reagents Aroclor 1016, 1221, 1232, 1242, 1248, 1254, and 1260 standards at a concentration of 100 µg ml -1 in methanol were purchased from AccuStandard, Inc. (New Haven, CT). The SPME fiber was coated with polydimethylsiloxane (PDMS, 100 µm film thickness) and used with 20-mL headspace glass vials and crimp seals with PTFE/silicone septa. The SPME fibers, vials, and seals were all obtained from Sigma-Aldrich Co. LLC. (St. Louis, MO). The clean soil was purchased from RT Corp (Laramie, WY) Instruments A Thermo Finnigan PolarisQ quadrupole ion trap mass spectrometer/trace GC system with a Triplus AS2000 autosampler (San Francisco, CA, USA) was used to collect all the experimental data. The GC/MS system was controlled using the XCalibur software version provided by Thermo. Analytes were separated using a SHRXI-5MS capillary column (5% diphenyl/95% dimethylpolysiloxane cross-linked, 30 m 0.25 mm i.d., 0.1 µm film thickness) from Shimadzu Scientific Instruments Inc. (Columbia, MD). All the data were processed using the MATLAB R2012b (MathWorks Inc., Natick, MA).

81 Data collection Aroclor standard solutions at concentrations of 0.3, 1, and 3 µg ml -1 in duplicate were prepared by dilution with methanol from a 100 µg ml -1 stock solution. A 50 µl aliquot of each standard solution was added to a 20-mL headspace glass vial. After the vial was sealed, it was incubated at 100 C for 5 min. Then a PDMS fiber was exposed to the headspace for 25 min. The fiber was thermally desorbed in the GC injector at 280 C for 5 min. The analytes were separated using the following oven temperature program at a constant flow of 1 ml min -1 : 50 C, hold for 1 min, ramp at 20 C min -1 to 280 C, hold for 10 min. The transfer line and ion source temperatures were both maintained at 280 C. Full scan mode was selected for the mass spectrometer and the scan range was from mass-to-charge ratio (m/z) 140 to 550. Six replicates of an empty vial and each Aroclor sample were prepared and determined individually. A random block experimental design was used so that replicates of each Aroclor sample were distributed with respect to time and would characterize any variation that occurred during the course of the experiment. Two concentrations at 0.3 and 1.0 µg g -1 of soil samples in duplicate for each Aroclor were prepared by using Aroclor standards and blank soil samples (4 7 = 28 soil samples). The PCBs tend to

82 82 bind to the surface of soil particles strongly because of their lipophilicity. A saturated potassium dichromate solution was used to free the PCBs from the soil and prepared by dissolving an excess of potassium dichromate in 6.0 M sulfuric acid. 0.5 g soil samples were added to the SPME vial and two milliliters of saturated potassium dichromate-sulfuric acid solution were added to the SPME vial for the sampling of the PCBs from the soil matrix. After the vial was sealed and vortexed for 10 s, it was incubated at 100 C for 5 min. Afterwards, a PDMS fiber was exposed to the headspace for 25 min. Two-way data sets were collected using the same GC/MS program as previously described. Five blank soil samples without any Aroclor were treated in the same way. The blank soil sample data were used for correcting the baselines of the Aroclor soil samples Data format. The two-way GC/MS data sets were initially acquired as RAW files. The RAW files were converted to the common document format (CDF) with the File Converter Tool in the XCalibur Software. The CDF files were read directly into MATLAB. For further data processing, the data sets were binned by retention time from 4.10 to min with a 0.01 min increment and binned by mass-to-charge ratios from 140 to 550 Th with a 1 Th

83 83 increment. Therefore, each two-way GC/MS object comprised data points in which 1801 rows corresponded to the retention times and 411 columns corresponded to mass-to-charge ratios. A total of 81 two-way GC/MS data (i.e., 42 for Aroclor standard samples and 6 blanks and 28 for Aroclor soil samples and 5 blanks) were collected Data preprocessing The data sets were pretreated by baseline correction (30 components), modulo compression, and normalization. For comparison purposes, one-way total mass spectrum (TMS) and oneway total ion current (TIC) objects were constructed by summing across the complementary way of the two-way GC/MS data object. No retention time alignment was required in this study. A set of experiments demonstrated that no improvement was obtained by aligning the standard samples and the soil samples, and analyses of the TIC did not demonstrate any retention time drift Data processing After data preprocessing, FuRES, PLS-DA, and FOAM classifiers were compared according to their prediction accuracies through validation with the bootstrap Latin partition method. Four Latin partitions and 100 times bootstraps were used to construct the

84 84 models. To evaluate the effectiveness of modulo compression, four data set representations were evaluated: two-way GC/MS, two-way GC/MS after modulo compression, one-way TIC, and one-way TMS. These different data configurations were used to construct FuRES, PLS-DA, and FOAM classifiers in each Latin partition. The four Latin partitions were randomly selected from the data so that each object was used one time for prediction and three times for calibration for each of the 100 bootstraps. In addition, the partition is constrained so that equal class distributions are maintained between the calibration and prediction data sets. Because each object is used once and only once for prediction, the results of the four partitions were pooled. The pooled prediction results from FuRES, PLS-DA, and FOAM classifiers were averaged across the 100 bootstraps and were reported with 95% confidence intervals. 3.2 Results and discussion SPME-GC/MS analysis The primary purpose of this study was to develop an automatic method or pipeline that does not require human expertise to identify different Aroclors. SPME is an efficient technique to extract PCBs from matrices [118] and it integrates sampling, extraction, concentration, and sample introduction to an instrument into one step. The

85 85 optimization of SPME conditions will be described in another publication. To summarize, a 100 µm PDMS coating the SPME fiber was selected because of its relatively high extraction efficiency; for a reasonable extraction time and acceptable response, 100 C as the extraction temperature and 25 min as extraction time were used; to avoid cross-contamination, the SPME fiber was desorbed at 280 C for 5 min which is the maximum operation temperature for 100 µm PDMS SPME fiber. Figure 3-1 gives the superimposed TIC chromatograms of the blank and Aroclor samples. The PDMS peaks appear at retention times of 4.83 min, 6.10 min, 7.36 min, 8.51 min, 9.53 min, min, min, min, min, min, min, min, and min. These mass spectra conformed to the PDMS spectra in the National Institute of Standards and Technology (NIST) database (Version 2.0). The size of the PDMS interference peaks and their effect on classification were reduced after baseline correction.

86 86 Figure 3-1. Total ion current (TIC) chromatograms of solvent blank samples (A), Aroclor samples before (B), and after (C) baseline correction. PDMS peaks are designated with a label Baseline correction The baseline variations may cause errors in classification. The baseline chromatograms are not always constant so the baseline correction simply by subtracting an average blank spectrum does not work very well, nor does it work well for two-way GC/MS data objects because of column bleed [159]. In the present study, a blank object from each block of experiments was collected. The blank samples

87 87 were either the SPME of the empty headspace vials or the SPME sample of soils without any Aroclors added. The mass spectra from each blank run were used to construct an orthonormal basis set. Therefore, the bases characterized variations from the septum peaks, the SPME fiber peaks and column bleed, and for the soil the same experimental artifacts along with any peaks that arose from the soil were corrected. The soil and standards were corrected separately so a soil basis was never used to correct the standard and vice versa. Different numbers of components for the baseline correction were systematically investigated and their effects on classification rates with FuRES and PLS-DA are reported in Figure 3-2. The models were constructed from four Latin partition and ten bootstraps. Increased numbers of components improved classification rates for both FuRES and PLS-DA. The classification rates with FuRES using modulo compressed data sets was 100% after 30 or more components were used. The quality of the data was significantly improved after baseline correction. Some negative mass spectral peaks arose from the baseline correction which when calculating the total ion current results in a decrease in chromatographic peak intensities. However, when a baseline is properly modeled there should be an equal number of positive and negative residuals about the baseline. For the

88 88 unprocessed data, all the mass spectral peaks are positive so that when the mass spectra are totaled the total ion current produces larger chromatographic peak intensities. The benefit of baseline correction can also be visualized by the comparison of TIC chromatograms of Aroclor samples before and after baseline correction in Figure 3-1. The PDMS peaks were reduced or even eliminated and the baselines for all the TIC chromatograms were corrected. As a result of this evaluation, 30 components were chosen for baseline correction to treat all the Aroclor data sets before any further data preprocessing or processing.

89 89 Figure 3-2. FuRES and PLS-DA prediction rates with respect to the numbers of components for baseline correction using different data sets representations. The 95% confidence intervals are given in dashed lines Modulo compression After baseline correction of the two-way GC/MS data sets, the mass spectral features were compressed with different divisors or numbers of features. The PDR method was used to study the effect of the number of features on modulo compression. For evaluating the choice of divisor for the modulo compression, 21 PDR values among the seven classes/aroclors (21 pairwise

90 90 combinations) were obtained for each divisor. The pair that had the minimum PDR value represented the most similar pair of classes. The modulo divisor was evaluated from 1 to 60 and the minimum PDR values are plotted with respect to divisor in Figure 3-3. The maximum PDR value was observed when modulo number 35 which indicated that the best separations were achieved for the most abundant isotope of Chlorine. The PDR method is very efficient, and it required about 12 min to complete the 60 evaluations. Figure 3-3. The minimum PDR value among 7 classes in modulo compression data sets varies with the modulo numbers.

91 91 The average classification rates with FuRES, PLS-DA, and FOAM by using two-way modulo compression data with different modulo numbers were also evaluated for 100 bootstrapped Latin partitions. Divisors of 35 and 37 were selected to model the two most abundant isotopes of Chlorine. Smaller and larger divisors 19 and 60 were selected arbitrarily. A divisor of 18 was also chosen because it was half of the average of 35 and 37. The PDR value for the divisor 18 is relatively high compared to the other values in Figure 3-3. Smaller divisors achieve greater compressions and a consequent improvement in signal to noise ratio although with a loss of informing power. The choice of divisor significantly affected the classification rates of all three classifiers. See Table 3-1. Using a modulo divisor of 35 (mod- 35) to compress the mass spectra provided the best classification rates, which was expected from the PDR study. When the mass spectra were compressed with mod-19 or mod-60, the classification rates were worse. Among these tests, using mod-35 yielded the best classification performance which is in accord with the results from the PDR evaluation. Henceforth, mod-35 compression was used. Using classification rates to optimize modulo numbers is clear and straightforward but requires longer evaluation times. At least 1.5 h was needed for just one bootstrap evaluation which would be enough

92 92 time to complete 450 evaluations using PDR. However, by reducing the number of bootstraps would also decrease the evaluation time. The benefit of modulo compression for the data was evaluated by the classification rates with FuRES and PLS-DA for different data configurations. Among them, the classification rates of FuRES and PLS-DA using mod-35 compressed data were better than the other data representations. Table 3-1. FuRES, PLS-DA and FOAM Classification Rates with 95% Confidence Intervals Obtained by Using Different Numbers of Modulo Features in Modulo Compression with Baseline Correction of Training Data Using 3 Components. Average classification rates a (%) Modulo numbers FuRES PLS-DA FOAM ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± 0.8 a Averages were calculated from bootstrapped Latin partitions.

93 93 All the data representations were evaluated by the classification rates after baseline correction for which 3 components were used. With 3 components selected for the baseline correction, so that the prediction results were good but not perfect and the classifiers would be more sensitive to different data representations, e.g., 2-way versus 1-way. The average classification rates of FuRES and PLS-DA for the two-way modulo compressed data (mod-35) were 92.8 ± 0.7% and 86.0 ± 0.7%; for two-way data sets were 86.3 ± 0.7% and 84.0 ± 0.9%; for one-way TIC were 88.4 ± 0.7% and 83.7 ± 0.8%; for oneway TMS were 78.7 ± 0.8% and 74.3 ± 1.3%, respectively, for bootstrapped Latin partitions. See Table 3-2. Moreover, the classification results for FuRES were always better than those of PLS-DA when comparing the classification rates by using the same data sets. However, PLS-DA was implemented an automatic fashion so it would function similar to FuRES which is parameter free. The training data was partitioned into 2 and bootstrapped 10 times. The number of components with the lowest pooled prediction error was selected for building a model that was then used for the prediction data. By tweaking the number of PLS-DA components for each training prediction set pair might improve the prediction results would,

94 but that would be a prodigious task when implemented in a bootstrap evaluation. 94 Table 3-2. FuRES and PLS-DA Classification Rates with 95% Confidence Intervals Obtained by Using Different Data Representations with Baseline Correction of Training Data. Average classification rates a (%) Classifiers Two-way mod-35 b Two-way One-way TIC One-way TMS FuRES 92.8 ± ± ± ± 0.8 PLS-DA 86.0 ± ± ± ± 1.3 a Two-way mod denotes two-way data sets after modulo compression; Two-way denotes two-way data sets; One-way TIC denotes total ion current data sets; One-way TMS denotes total mass spectra data sets. The number of components in baseline correction was 3. Averages were calculated from bootstrapped Latin partitions. b The mass spectrum were compressed to 35 features in modulo compression FuRES, PLS-DA, and FOAM classification Three classifiers, FuRES, PLS-DA, and FOAM, were constructed and evaluated with four Latin partitions and 100 bootstraps for the classification of seven standard Aroclor sample data sets using the optimized parameters for data preprocessing as previously described. All the four data representations were pretreated by baseline

95 95 correction using 30 components, and normalization. For two-way modulo compressed data, the mass spectra were compressed using a divisor of 35. The classification rates of the three classifiers for the four different data sets are listed in Table 3-3. The FuRES, PLS-DA, and FOAM predictions were averaged across the 100 bootstraps for the different data representations. Therefore all three classifiers using two-way mod-35 and one-way TIC representations worked well while the classification of two-way modulo compressed data sets by FuRES gave the best classification rates.

96 Table 3-3. FuRES, PLS-DA and FOAM Classification Rates with 95% Confidence Intervals Obtained by Using Different Data Representations with Baseline Correction Using 30 Components of the Calibration Data. Average classification rates a (%) Classifiers Two-way mod-35 b Two-way One-way TIC One-way TMS FuRES 100 ± ± ± ± 0.3 PLS-DA 94.6 ± ± ± ± 0.8 * FOAM 96.1 ± ± 0.9 * 97.0 ± ± 0.5 * a Two-way mod denotes two-way data sets after modulo compression; Two-way denotes two-way data sets; One-way TIC denotes total ion current data sets; One-way TMS denotes total mass spectra data sets. Averages were calculated from bootstrapped Latin partitions. b The mass spectrum were compressed to 35 features in modulo compression. * indicates that the classification rates smaller than 95%. 96

97 97 Figure 3-4. Principal component analysis score plot for two-way modulo compressed (35 features) data sets of seven Aroclor standard samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 64%. The second number in the parentheses denotes the absolute variance. Figure 3-4 is a principal component analysis (PCA) score plot for the two-way mod-35 data of the 7 Aroclor standard samples. After the data preprocessing of baseline correction, modulo compression, and normalization; Aroclor 1260, 1254, 1221 and 1248 were individually separated very well from the other Aroclors. The similarity of the data among Aroclor 1016, 1232, and 1242 lead to overlapping 95% confidence ellipses in the PCA score plot. By comparing the complete PCB congener distributions for Aroclor 1016, 1232, and

98 , it was found that the compositions of PCB congeners for these Aroclors were very similar [18]. Actually Aroclor 1016 was manufactured by the fractional distillation of Aroclor 1242, which excluded the more highly chlorinated congeners and Aroclor 1232 was a 50:50 blend of Aroclor 1242 and 1221 [18]. Nevertheless, the FuRES classifier successfully classified all 7 Aroclors for 100 bootstraps without a single error and the PLS-DA and FOAM classifiers were able to classify the seven Aroclors with classification rates above 95% Classification of Aroclor soil samples FuRES, PLS-DA, and FOAM classifiers were constructed using the initial 7 Aroclor standard sample data sets after baseline correction (30 components), normalization, modulo compression (mod-35). The constructed FuRES, PLS-DA, and FOAM classifiers were used on the new data sets of Aroclor soil samples. The prediction rates of 28 Aroclor soil samples with FuRES, PLS-DA, and FOAM classifiers using four data set representations are listed in Table 3-4. One Aroclor 1232 was misclassified to Aroclor 1242 by the FuRES classifier built from the mod-35 data (Figure 3-5 A) and all Aroclor soil samples were accurately classified by the FOAM classifier established by using the mod-35 data (Figure 3-5 B). Three misclassifications were found when Aroclor soil samples were classified with PLS-DA by using one-way TIC

99 99 data, and the results were worse for the other data representations. The PCA score plot using the two-way mod-35 data of 7 Aroclor standard samples as training sets and seven Aroclor soil samples as prediction sets are given in Figure 3-6. Although the prediction rates for the 7 groups of Aroclor soil samples were 100%, the prediction data sets did not perfectly match with the training data sets in the PCA score plot (Figure 3-6). This disparity could be ascribed to the variations of PCB extraction efficiency and matrix effects differences for the Aroclor standard samples and Aroclor soil samples. However this study has demonstrated the capability to the classification of Aroclor soil samples using Aroclor standard samples and thereby eliminating the requirement to collect a variety of soil samples. The PCA score plot using two-way, one-way TIC, and one-way TMS data representations of training sets and prediction sets are given in Figure 3-7, 3-8, and 3-9, respectively. In the PCA score plots by using twoway and one-way TMS data sets, the standard samples and soil samples for several Aroclors are separated [e.g., the standard samples and soil samples for Aroclor 1260 and 1254 by two-way data set (Figure 3-7), for Aroclor 1016, 1221, 1232 and 1248 by one-way TMS data set (Figure 3-9)]. The classes of these soil samples will be misclassified. Although in the PCA plot by using one-way TIC, the

100 100 standard samples of all the Aroclors overlap with the respective soil samples in some extent (Figure 3-8), the separation between different Aroclors is poorer comparing the PCA plot by using mod-35 data (Figure 3-6). This result agrees with the prediction result in Table 3-4, from which we observed that the prediction rates by using FuRES and FOAM with one-way TIC data set are much better than the rates by using two-way and one-way TMS data representations, but worse than the rates by using two-way mod-35 data set. Table 3-4. Classifier Prediction Rates (%) of Standard Aroclor Soil Samples with No Parametric Changes. Classification rates a Classifiers Two-way Two- One-way One-way Two-way mod-35 b way TIC TMS mod-18 c FuRES PLS-DA FOAM a Two-way mod denotes two-way data sets after modulo compression; Two-way denotes two-way data sets; One-way TIC denotes total ion current data sets; One-way TMS denotes total mass spectra data sets. The number of components in baseline correction was 30. Averages were calculated from bootstrapped Latin partitions. b The mass spectrum were compressed to 35 features in modulo compression. c The mass spectrum were compressed to 18 features in modulo compression.

101 Figure 3-5. Prediction plot of Aroclor soil samples using FuRES (A) and FOAM (B) classifiers constructed by modulo compressed (35 features) Aroclor standard data sets. Each spot represents an Aroclor soil sample. The correct classified samples are colored in green, and misclassified samples are colored in red. 101

102 Figure 3-6. Principal component analysis score plot for two-way modulo compressed data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 57%. The second number in the parentheses denotes the absolute variance. 102

103 103 1 A: Aroclor 1260 standard samples; B: Aroclor 1254 standard samples; C: Aroclor 1016 standard samples; D: Aroclor 1221 standard samples; E: Aroclor 1232 standard samples; F: Aroclor 1242 standard samples; G: Aroclor 1248 standard samples; a: Aroclor 1260 soil samples; b: Aroclor 1254 soil samples; c: Aroclor 1016 soil samples; d: Aroclor 1221 soil samples; e: Aroclor 1232 soil samples; f: Aroclor 1242 soil samples; g: Aroclor 1248 soil samples dd d D d D D D D D PC #2 (17%, ) ee e e E E c C F F c c ff F C f F E C E C f F F E G G g G g G G b b B a a B a B B B B A A A A A PC #1 (30%, 0.125) Figure 3-7. Principal component analysis score plot for two-way data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 47%. The second number in the parentheses denotes the absolute variance.

104 A: Aroclor 1260 standard samples; B: Aroclor 1254 standard samples; C: Aroclor 1016 standard samples; D: Aroclor 1221 standard samples; E: Aroclor 1232 standard samples; F: Aroclor 1242 standard samples; G: Aroclor 1248 standard samples; a: Aroclor 1260 soil samples; b: Aroclor 1254 soil samples; c: Aroclor 1016 soil samples; d: Aroclor 1221 soil samples; e: Aroclor 1232 soil samples; f: Aroclor 1242 soil samples; g: Aroclor 1248 soil samples. 0.6 d d D d D D D PC #2 (19%, ) CC e e e c cc c C f C f f C f F E E F E F E E F F G G E F D D A A A A B B B B B B a aa a -0.4 g g g GG G b b bb PC #1 (29%, 0.101) Figure 3-8. Principal component analysis score plot for one-way TIC data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 48%. The second number in the parentheses denotes the absolute variance.

105 A: Aroclor 1260 standard samples; B: Aroclor 1254 standard samples; C: Aroclor 1016 standard samples; D: Aroclor 1221 standard samples; E: Aroclor 1232 standard samples; F: Aroclor 1242 standard samples; G: Aroclor 1248 standard samples; a: Aroclor 1260 soil samples; b: Aroclor 1254 soil samples; c: Aroclor 1016 soil samples; d: Aroclor 1221 soil samples; e: Aroclor 1232 soil samples; f: Aroclor 1242 soil samples; g: Aroclor 1248 soil samples d d d d DDD D PC #2 (23%, ) c c c e e e e f f fc f C C gg g F E E F F F E E F F E G G G bb b D D a B a a a BB B B B A A A A PC #1 (38%, 0.102) Figure 3-9. Principal component analysis score plot for one-way TMS data sets of seven Aroclor standard samples and seven Aroclor soil samples. The 95% confidence intervals are represented by the ellipses. The percent variance of the first two principal components given in parentheses is 61%. The second number in the parentheses denotes the absolute variance. 3.3 Conclusions An automatic classification method for 7 Aroclors using SPME- GC/MS and chemometrics was developed in this study. The modulo compression was introduced and evaluated for the classification of complex mixtures of PCBs by GC/MS for the first time. The performance of the classifiers was improved after mod-35 compression

106 106 was applied. The effect of numbers of components in the baseline correction was investigated. The application of a larger number of components in the baseline correction successfully removed the baseline and significantly reduced the influence of artifact peaks that arose from the SPME fiber or the septa (i.e., PDMS peaks). After the Aroclor data sets were treated by baseline correction using 30 components and the mass spectral features were compressed from 411 to 35 by modulo compression, the average classification rates for bootstrapped Latin partitions were 100 ± 0% with FuRES, 94.6 ± 0.7% with PLS-DA and 96.1 ± 0.6% with FOAM. This method compressed the data to 8.5% of its original size. Compressed GC/MS data may be beneficial for miniaturized and portable GC/MS units that lack sensitivity and the capacity to store large volumes of data. Modulo peak integration may improve signal-to-noise ratios and sensitivity for miniaturized instruments while maintain selectivity. With an appropriate baseline correction method, all the other data representations, two-way data, one-way TIC data, and one-way TMS data, can be used to establish classifiers with satisfactory classification rates without the need to run standards in soil or other complex matrices. Twenty-eight Aroclor soil samples were classified by these classifiers with the classification rates of 100% for FOAM and 96.4%

107 107 for FuRES. This study proposed a method for the classification of 7 Aroclors at low concentration (1 ppm and below) in soil without using a long GC program for full separation of PCB congeners nor were any soil samples used in building the classifiers or configuring the automatic classification system.

108 108 CHAPTER 4: SIMULTANEOUS QUANTIFICATION OF AROCLOR MIXTURES IN SOIL SAMPLES BY GAS CHROMATOGRAPHY/MASS SPECTROMETRY WITH SOLID PHASE MICROEXTRACTION USING PARTIAL LEAST-SQUARES REGRESSION Multivariate partial least-squares (PLS) method was applied to the quantification of two complex polychlorinated biphenyls (PCBs) commercial mixtures, Aroclor 1254 and 1260, in a soil matrix. PCBs in soil samples were extracted by headspace solid phase microextraction (SPME) and determined by gas chromatography/mass spectrometry (GC/MS). Decachlorinated biphenyl (deca-cb) was used as internal standard. After the baseline correction was applied, four data representations including extracted ion chromatograms (EIC) for Aroclor 1254, EIC for Aroclor 1260, EIC for both Aroclors and two-way data sets were constructed for PLS-1 and PLS-2 calibrations and evaluated with respect to quantitative prediction accuracy. The PLS model was optimized with respect to the number of latent variables using cross validation of the calibration data set. The validation of the method was performed with certified soil samples and real field soil samples and the predicted concentrations for both Aroclors using EIC data sets agreed with the certified values. The linear range of the method was from 10 ng g -1 to 1000 ng g -1 for both Aroclor 1254 and

109 in soil matrices and the detection limit was 4 ng g -1 for Aroclor 1254 and 6 ng g -1 for Aroclor Experimental Reagents A mixture containing 1 mg ml -1 of decachlorobiphenyl (deca-cb) and of tetrachloro-m-xylene (TCMX), 100 µg ml -1 Aroclor 1254 and 1260 standards solutions were purchased from AccuStandard, Inc. (New Haven, CT). Potassium permanganate, potassium dichromate, the SPME fibers coated with polydimethylsiloxane (100 µm film thickness), 20-mL headspace glass vials, and crimp seals with PTFE/silicone septa were obtained from Sigma-Aldrich Co. LLC. (St. Louis, MO). The clean soil and certified Aroclor 1254 and 1260 soil samples were purchased from RT Corp (Laramie, WY). The standard solutions for Aroclor 1254 and 1260 at the concentrations of 0.1, 0.3, 1, 3 and 10 µg ml -1 were prepared by dilutions from 100 µg ml -1 Aroclor 1254 and 1260 stock solutions with methanol. The internal standard solution containing 10 µg ml -1 deca- CB and TCMX was prepared by dilution of aliquots of the 1 mg ml -1 stock solution with hexane. The standard soil samples for calibration set were prepared by thoroughly mixing 50 µl standard solutions with 0.5 g clean soil and evaporating the solvent in a hood under the room

110 110 temperature until the soil became free-flowing and powdery. The concentrations of Aroclors in prepared soil samples were calculated as follows: solution 4 1) for which C is the concentration of Aroclor standard solutions; is 50 µl and is the mass of the soil which was 0.5 g in our experiment. To determine the two Aroclors simultaneously by PLS, the calibration matrix for 6 6 mixture Aroclor 1254 and Aroclor 1260 was constructed. A saturated potassium dichromate solution was prepared by dissolving an excess of potassium dichromate in 6.0 M sulfuric acid Instruments Analyses were conducted on a Thermo Finnigan PolarisQ quadrupole ion trap mass spectrometer/trace GC system with a Triplus AS2000 autosampler (San Francisco, CA, USA) which was controlled using the XCalibur software version provided by Thermo. Analytes were separated using a 30 m 0.25 mm 0.1 µm 5% diphenyl/95% dimethylpolysiloxane cross-linked capillary column (SHRXI-5MS, Shimadzu Scientific Instruments Inc. Columbia, MD). The temperature program was; 50 C, hold for 1 min, ramp at 20 C min -1 to 280 C, hold for 10 min. Helium was used as the carrier gas, with a constant flow rate of 1 ml min -1. The transfer line and ion

111 111 source temperatures were both maintained at 280 C. Mass spectra were obtained in the electron ionization mode (40 ev) in the range from mass-to-charge ratio (m/z) 140 to 550 starting after a 4 min solvent delay. All the data were processed using MATLAB R2012b (MathWorks Inc., Natick, MA) Sample preparation Extraction of PCBs from soil samples under optimized conditions were carried out using headspace SPME with 2 ml saturated potassium dichromate solution and a 100 µm PDMS fiber at 100 C for 30 min [169]. An aliquot of 0.5 g of soil sample was placed in a 20- ml SPME vial with the addition of 20 µl of internal standard solution. After the solvent was evaporated in the hood at room temperature, 2 ml of the saturated potassium dichromate solution was added to the vial and the vial was sealed by a headspace aluminum cap with PTFE/silicone septa. After the sample vial was incubated at 100 C for 0.5 min, a PDMS fiber was exposed to the headspace for 30 min with the agitation setting on for 10 s and then off for 10 s sequentially. The fiber was then immediately inserted into the GC injector for desorption at 280 C for 5 min, and then the analytes were analyzed by GC/MS.

112 Results and discussion The optimization of the SPME-GC/MS method was discussed in another publication [169]. Aroclor 1260 standard was used to evaluate the optimization of the SPME and instrument conditions. The peak areas of hexa-chlorinated biphenyls (hexa-cbs) were representative for the comparison of the effects of different conditions because hexa-cbs are one of the major PCBs in both Aroclor 1254 (26.9 weight %) and Aroclor 1260 (46.9 weight %) [163]. Figure 4-1 comprises the chromatograms obtained from Aroclor 1254 and Because of the high similarity of both Aroclors, none of the PCB peaks can be used as characteristic peaks for either Aroclor. However the ratios of different PCBs in each Aroclor vary. For this case, the multivariate PLS (PLS-2) method is suitable.

113 Figure 4-1. GC/MS total ion current (TIC) chromatograms of Aroclor 1254 (A) and Aroclor 1260 (B) for 300 ng g -1 soil sample after headspace SPME extraction. On the right of each TIC chromatogram is zoom-in retention time window of PCB peaks for each Aroclor. 113

114 114 Table 4-1. Concentration matrix of calibration data set Sample ID Concentration of Concentration of Aroclor1254 (ng g -1 ) Aroclor1254 (ng g -1 )

115 PLS method Orthogonal baseline correction was applied prior to data processing which reduced the artifact peaks (e.g., PDMS peaks) and corrected the baselines for all the TIC chromatograms [166]. Then the intensities at different masses in each chromatogram were normalized to the area of its deca-cb peak. The calibration set comprised 36 standard mixtures for which the concentrations were 10, 30, 100, 300, and 1000 ng g -1 for each Aroclor. The concentration matrix (36 2) for the calibration set for which each column contains the concentrations of the corresponding Aroclor is given in the Table 4-1. Other than the original two-way GC/MS data set, extracted ion chromatograms (EIC) at retention time windows (RTW) between 11.0 to 16.0 min were used to construct smaller two-way GC/MS data sets which were selective for the PCBs. The molecular ions of tetra-chlorinated biphenyls (tetra-cb, m/z 292), penta-chlorinated biphenyls (penta-cb, m/z 326), hexa-cb (m/z 360), hepta-chlorinated biphenyls (hepta-cb, m/z 394), and octa-chlorinated biphenyls (octa-cb, m/z 430) were used to create the EIC two-way data. Three EIC data sets were investigated: 1) the EIC-A54 data set includes m/z 292, 326 and 360 for tetra-, penta-, and hexa-cbs that account for 94% of the PCBs in Aroclor 1254; 2) the EIC-A60 data set

116 116 includes m/z 360, 394 and 430 for tetra-, penta-, and hexa-cbs which account for 90% of the PCBs in Aroclor 1260; 3) the EIC-Both data set includes all five PCBs mentioned above which represent more than 98% of the PCBs in both Aroclor 1254 and Aroclor 1260 [163]. Figure 4-2. Plot of the RMSPE with respect to the number of PLS components for the EIC-Both data. To select the number of components for the PLS-2 model and compare the different data sets, a cross-validation method was used that leaves out one sample at a time (i.e., one design point). Figure 4-2 is a plot of the RMSPE with respect to the number of PLS components for the EIC-Both data. The summed prediction error for

117 117 both Aroclor 1254 and 1260 decreased for the first 8 principal components and increased thereafter. Therefore number of components was set to 8 for PLS-2 model. From Figure 4-2, it is apparent that the optimal component numbers for the two Aroclors are different (e.g., 7 for Aroclor 1254 and 9 for Aroclor 1260). To investigate the effect of the number of PLS components, PLS (PLS-1) was used because it models each Aroclor separately. The results for the optimal component numbers, linear regression coefficients, RMSPE and RPEs for univariate PLS (PLS-1) and multivariate PLS (PLS-2) using different data representations are reported in Table 4-2. For Aroclor 1254, the coefficient of determination ( ), RMSPE, and RPEs results for the EIC-A60 data are the worst among the different data representations. This result is reasonable because hepta-cb (m/z 394) and octa-cb (m/z 430) only account for less than 4% of the PCBs in Aroclor 1254 and those PCB signals can only be detected in calibration samples at the higher concentrations, so they result in larger errors. For similar reasons, the results for Aroclor 1260 using EIC-A54 data sets are slightly worse than the other data representations because tetra-cb (m/z 292) and penta-cb (m/z 326) only account for less than 10% of the PCBs in

118 118 Aroclor 1260 and cannot be detected in calibration samples at the lower concentrations. The optimal component numbers for PLS-2 models using different data representations are consistent but those for PLS-1 models using different data representations varies. Comparing the, RMSPE, and RPEs results between PLS-1 and PLS-2 models, overall the PLS-1 results are better than PLS-2 except that for Aroclor 1254 using the EIC-Both data representation, the prediction errors of the PLS-1 model are slightly larger than the results of the PLS-2 model. It is worth mentioning that any single EIC (i.e., m/z 292 for tetra-cb) data may be used to construct the multivariate calibration. However some masses result in substantial errors (see Table 4-3). It is risky to use only one mass for the whole mixture because the selected PCBs could be relatively low in intensity compared with other PCBs and they could be highly overlapped between different Aroclors. Moreover it would be difficult to predict which mass is the best to use.

119 119 Table 4-2. Comparison of PLS-1 and PLS-2 using calibration set by cross validation EIC-Both a EIC-A60 A1254 A1260 A1254 A1260 PLS-1 PLS-2 PLS-1 PLS-2 PLS-1 PLS-2 PLS-1 PLS-2 Num_comp b (μ ) (%) EIC-A54 Two-way A1254 A1260 A1254 A1260 PLS-1 PLS-2 PLS-1 PLS-2 PLS-1 PLS-2 PLS-1 PLS-2 Num_comp (μ ) (%) a EIC-Both denotes extracted ion chromatogram (EIC) data sets which includes m/z 292, 326, 360, 394, and 430; EIC-A60 denotes EIC data sets which includes m/z 360, 394, and 430; EIC-A54 denotes EIC data sets which includes m/z 292, 326, and 360; Two-way denotes two-way data sets. b Num_comp denotes number of components used in PLS models.

120 120 Table 4-3. Linear regression coefficient ( ), RMSPE and RPEs results using different data representations. μ % Data A125 A126 representation A1254 A1260 A1254 A EIC-m/z EIC-m/z EIC-m/z EIC- m/z EIC- m/z Validation of method Certified soil samples that contained Aroclor 1254 at 5.93 µg g -1 (prediction interval µg g -1 ) and Aroclor 1260 at 1.50 µg g -1 (prediction interval µg g -1 ) were analyzed. The certified soil samples were diluted with certified clean soils in accord with the calibration range and mixed to furnish 8 soil samples with different concentrations of Aroclor 1254 and Each soil sample was analyzed by SPME-GC/MS in five replicates. The same PLS-1 and PLS- 2 models using their optimal component numbers for the 4 different data representations were applied with a validation set of certified Aroclor soil samples. The prediction rates are given in Table 4-4 and

121 the full details of samples and prediction results are shown in Table Table 4-4. Comparison of prediction rates between PLS-1 and PLS-2 for certified soil samples (number of samples = 8). Prediction Rate (%) Data representation a PLS-1 PLS-2 EIC-Both EIC-A60 EIC-A54 Two-way A A A A A A A A a EIC-Both denotes extracted ion chromatogram (EIC) data sets which includes m/z 292, 326, 360, 394, and 430; EIC-A60 denotes EIC data sets which includes m/z 360, 394, and 430; EIC-A54 denotes EIC data sets which includes m/z 292, 326, and 360; Two-way denotes two-way data sets.

122 122 Table 4-5. Prediction results of certified soil samples using different data representations a by PLS-2. ID 1 Prediction interval (certified reference value) (μ ) Aroclor (494) (49) (494) (49) (49) (494) Aroclor 1260 Aroclor 1254 EIC-A54 Concentration b (μ ) Aroclor 1260 EIC-A60 Concentration b (μ ) Aroclor 1254 Aroclor 1260 Aroclor 1254 EIC-Both Concentration b (μ ) Aroclor 1260 Two-way Concentration b (μ ) Aroclor ± ± ± ± (500) Aroclor ± ±30* 390± ± ± ±6 0 40± ±6 0 0* (50) (500) (50) (500) (50) 0 70±20 40±20* 50± ± ±60* 270±50 600± ± ± ±50 500± ±70 500±100 40±20 80±20* 70±10 60±10 40±20 70±20 0* 0* 50±30 650±50 340±60* 450±20 40±30 540±30 120±20* 560±30 340±30 40±20 360±40 33±8 340±30 30±10 390±60 50±20 a EIC-A54 denotes extracted ion chromatogram (EIC) data sets which includes m/z 292, 326 and 360; EIC-A60 denotes EIC data sets which includes m/z 360, 394, and 430; EIC-Both denotes EIC data sets which includes m/z 292, 326, 360, 394 and 430; two-way denotes two-way data sets. b Average concentration for 5 replicates with concentration interval at a confidence interval of 95%. *The prediction results are out of prediction interval range.

123 123 For the PLS-1 model, 100% prediction rates were achieved for both Aroclor 1254 and 1260 by using EIC-Both and EIC-A54 data representations, and for Aroclor 1260 by using EIC-A60 and two-way data representations. The predicted concentrations for three samples of Aroclor 1254 using EIC-A60 using two-way data are outside of prediction intervals. The large prediction error for Aroclor 1254 using EIC-A60 data is also found in Table 4-2 (RPEs = 29.96%). Both results indicate that the EIC-A60 data representation is not suitable for predicting the concentration of Aroclor The concentrations for two samples of Aroclor 1254 using two-way data did not fall into the prediction interval ranges. The reason could be that the two-way data is more sensitive to background variations that arose from the soil samples. Although the baseline correction was applied to all the data sets prior to data processing, the soils used for the background correction may not have matched well with the soil from the real and certified soil samples. However, for the other three EIC data sets, the selected PCB ions were extracted from two-way data sets, so they are more selective and more resistant to background peaks. For the PLS-2 model, all the concentrations obtained for Aroclor 1254 using EIC-A54 data sets, for Aroclor 1260 using EIC-A60 data sets and for both Aroclors using EIC-Both data sets were in agreement

124 124 with certified reference values and fell inside the certified range. Predicted concentrations for the EIC-A60 data for three samples of Aroclor 1254 were outside the prediction intervals and one sample for the EIC-A54 data for Aroclor 1260 was outside of the prediction interval range, which indicates that EIC-A60 and EIC-A54 data sets may not be suitable for Aroclor 1254 and Aroclor 1260, respectively. These results agree with the PLS cross validation results (Table 4-2). All the soil samples at low concentrations (50 ppb for either Aroclor 1254 or 1260 or both) failed to predict accurately using the two-way data sets which might be caused by background variation as well. Overall, the prediction results for certified soil samples from the PLS-1 models are better than the results from the PLS-2 models. However, PLS-2 models built from EIC-Both, EIC-A54 and EIC-A60 data representations have been demonstrated the capability to predict the concentrations of Aroclor 1254 or/and Aroclor 1260.

125 125 Table 4-6. Comparison of prediction rates between PLS-1 and PLS-2 for real soil samples (number of samples = 3). Data Prediction Rate (%) representation a PLS-1 PLS-2 EIC-Both EIC-A60 EIC-A54 Two-way A A A A A A A A a EIC-Both denotes extracted ion chromatogram (EIC) data sets which includes m/z 292, 326, 360, 394, and 430; EIC-A60 denotes EIC data sets which includes m/z 360, 394, and 430; EIC-A54 denotes EIC data sets which includes m/z 292, 326, and 360; Two-way denotes two-way data sets. The real soil samples were collected at the Portsmouth Gaseous Diffusion Plant, Ohio. The samples were split for analysis between our laboratory and GEL Laboratories, LLC (GEL), Charleston, South Carolina. Five replicates were performed on each soil sample and the same soil samples were analyzed by GEL using EPA 8082 method. As given in Table 4-6, the estimated concentrations for both Aroclor 1254 and 1260 using EIC-A54, EIC-A60, and EIC-Both data sets are in

126 126 agreement with the results from GEL using EPA 8082 method by PLS-2 models. All samples were outside the prediction intervals using the two-way data by PLS-2 model, which again might be caused by novel peaks arising from the soil background. PLS-1 models had better predictions for both Aroclors using two-way data but worse performance using EIC-Both and EIC-A60 data. The full prediction results are given in Table 4-7. Together with the prediction results for certified soil samples (Table 4-4), PLS-1 models were better for the EIC-A54 and the two-way data representations and PLS-2 models were better for the EIC-Both and the EIC-A60 data representations. Therefore, the EIC-A54 data sets predicted well for the Aroclor 1254 with PLS-2 regression; EIC-A60 data sets predicted well for the Aroclor 1260 with both PLS-1 and PLS-2 regression; the EIC-Both data predicted the concentrations of both Aroclors well with PLS-2 models; and the comprehensive two-way data predicted both Aroclors poorly with the PLS-2 regression but predicted Aroclor1260 well with the PLS- 1 regression. In practice, PLS-2 regression is preferable because only one calibration model is used Detection limit determination Several approaches have been reported for the determination of detection limit in multivariate calibration procedures [170]. In this

127 127 study, the detection limit has been calculated based three times the standard deviation of the blank signal. Four blank soil samples with internal standard were treated the same as those described in section Sample preparation. From the PLS-2 modeling for each Aroclor, the predicted concentrations for blank samples were calculated. Then three times the standard deviation of the predicted concentrations for each Aroclor was taken as detection limit. The detection limits of our method for Aroclor 1254 and 1260 are 4 ppb and 6 ppb, respectively. 4.3 Conclusion In this study, a multivariate linear regression method, PLS, has been developed and used to determine Aroclor 1254 and 1260 mixtures in soil samples. The use of PLS method enabled the analysis of multiple Aroclors in soil without the complex sample preparation and tedious instrumental separation. Four data representations, SIC-A54, SIC-A60, SIC-Both and the two-way data were evaluated for predicting the Aroclor concentrations by PLS-1 and PLS-2 models. The selected ion two-way (SIC) data sets were found to be more selective and resistant to background variance. The method has been validated by commercial certified soil samples and real on-site soil samples. The predicted concentrations of Aroclor 1254 by SIC-A54 data sets, the predicted concentrations of Aroclor 1260 by SIC-A60 data sets and the

128 128 predicted concentrations of both Aroclors by SIC-Both data sets were all in the prediction intervals for the certified soil and real soil samples. The main advantage of our method over previous reported approaches is the high sample throughput, with a soil sample being processed about every 35 min. Because the SPME method requires only 0.5 g soil, the difficulties of sampling, homogenizing, shipping, and storing the soil samples for laboratory analysis are greatly reduced. The uses of nonorganic solvents and short analysis time also increase the applicability of the developed method for on-site analysis [168].

129 129 Table 4-7. Prediction results of real soil samples using different data representations a by PLS-2 and the comparison with prediction results using EPA 8082 method. Prediction interval (mean value) (μ ) b ID Aroclor A (25) 7A 0 9A (71) Aroclor (30) (49) (154) EIC-A54 Concentration c Aroclor 1254 (μ ) Aroclor 1260 EIC-A60 Concentration c Aroclor 1254 (μ ) Aroclor 1260 EIC-Both Concentration c Aroclor 1254 (μ ) Aroclor 1260 Two-way Concentration c Aroclor 1254 (μ ) Aroclor ±10 40±10 45±9 35±6 40±10 30±7 0* 0* 0 40± ± ± ±8* 60±20 190±30 70±60 160±30 50±20 170±20 0* 70±50* a EIC-A54 denotes extracted ion chromatogram (EIC) data sets which includes m/z 292, 326, and 360; EIC-A60 denotes EIC data sets which includes m/z 360, 394, and 430; EIC-Both denotes EIC data sets which includes m/z 292, 326, 360, 394 and 430; two-way denotes twoway data sets. b The prediction results were accomplished by GEL lab using EPA 8082 method. c Average concentration for 5 replicates with concentration interval at a confidence interval of 95%. *The prediction results are out of prediction interval range.

130 130 CHAPTER 5: EXPEDITED FIELD SURVEY AND SAMPLING METHOD FOR PCBS, PCDDS AND PCDFS USING PORTABLE GAS CHROMATOGRAPHY/MASS SPECTROMETRY Currently, the analysis of PCBs and dioxins requires that the samples be collected, then packaged and transported to a laboratory for preservation and laboratory analysis. This is a time consuming process and adds sources of variance and complexity to each sample s history. Significant attention needs to be paid to method blanks and controls to account for the changing environmental conditions during handling, transportation and storage. To overcome these limitations, on-site analysis is cost effective since up to 70% cost can be reduced from on-site analysis compared with those made using a stationary laboratory as reported [171]. Described in this chapter is a sub-40 minute analysis method for the determination of PCBs in soil samples using a portable GC/MS instrument with headspace SPME. The method used a batterypowered portable GC/MS, a SPME sampling device, a 1000-W portable generator, a portable heating block and a portable scale to sample and analyze PCBs in soil samples on site. The performance of the portable instrument was discussed herein. The sample preparation was accomplished within 30 min and the instrumental analysis required

131 131 less than 7 min. The analysis of PCDDs and PCDFs using the portable SPME-GC/MS system was attempted, but the extraction efficiency from wet soil or dry soil was extremely low. 5.1 Experimental Portable GC/MS The Guardion 8 GC/MS (Torion Technologies, American Fork, Utah, USA) consists of a low thermal mass GC and a miniature toroidal ion trap mass analyzer (TMS). The system uses a fast heating low thermal mass injector and a miniature vacuum system, dual-stage diaphragm roughing pump and a turbo-molecular pump. The instrument can be ready for an injection within 3 minutes of powering on. A 90-cm 3 disposable helium cartridge and a rechargeable battery provide the carrier gas and power to the GC/MS system, which enable the portable stand-alone instrument to be used in the field without any other peripherals. The entire system weighs about 13 kg (28 lb) and is 47 cm 36 cm 18 cm ( in.) (Figure 5-1). The instrument can be operated from the on-board color touch LCD screen, or via a laptop connection. Commercially available LTM GC columns from Supelco can be used. In these studies, the column was an MXT- 5, 5 m 0.1 mm ID capillary column chemically bonded with 5% diphenyl/95% dimethyl polysiloxane, 0.4 µm film thickness.

132 132 Figure 5-1. Photographs of the Guardion -8 GC-TMS showing A) external and B) internal components Devices for SPME extraction The SPME sampling device (Custodion ) provided by Torion is specially designed to be field-portable and easy to operate. The mechanism of the SPME holder is similar to automatic ballpoint pens. The SPME fiber can be extended out of or withdrawn into a protective metal needle just by pushing the plunger on top of the holder. Commercial SPME fibers from Supelco (Bellefonte, PA, USA) were used. In this study, as elsewhere, 100 µm PDMS fibers were found to be the most suitable for PCB analyses. A pocket scale capable of weighing to 0.01 g (No. YA102, Ohaus, Parsippany, NJ, USA) was used to weigh soil samples in the field. 10 ml/20 ml glass headspace vials (SUN-SRI, Rockwood, TN, USA) and crimp top caps (SUN-SRI) were used for the sample collection and extraction steps. A stopwatch was used for monitoring sampling

133 133 times. A portable 1000-W generator was used as backup power for instrument and power supply for the heating block. The portable heating block, which is capable of heating up to 100 C within 20 min was designed and assembled in the lab. For this heating block, the resistive heating elements were inserted into the aluminum block and a variable voltage controller was used to control the block temperature. A transparent Perspex cover was made to minimize convective heating losses, and the effect of wind or rain when operating outdoors Reagents Standards used in this method include Aroclors 1016, 1221, 1232, 1242, 1248, 1254 and 1260, and an EPA 8082A PCB standard, which contains 19 PCB congeners, and were purchased from AccuStandard, Inc. (New Haven, CT, USA). Commercially available blank soil and PCB contaminated soil used to simulate the real soil samples were purchased from RT Corp (Laramie, WY, USA). A solution of 6 M H2SO4 was prepared from a stock solution of 95% H2SO4 (Sigma Aldrich, St. Louis, MO, USA). A solution of 0.2 M KMnO4 was prepared from a primary solid standards (Sigma Aldrich). Blank soil was used to simulate the PCB contaminated soil samples. For example, to simulate 10 ppm Aroclor 1260 contaminated

134 134 soil samples, 0.5 g blank soil was placed into the 10 ml glass vial and spiked with 50 µl of 100 ppm Aroclor 1260 standard solution. The soil was then vortexed for 2 minutes. The soil was dried in the hood at room temperature Sample preparation and analysis Two 0.5 g aliquots of each soil sample are measured in to two 10 ml glass vials. One is used for GC/MS analysis, and the other is used as a back-up or for moisture analysis. To extract PCBs from the soil samples, we added 2.5 ml 0.2 M KMnO4 and 0.25 ml 6 M H2SO4 to the 10 ml vial. After sealing the vial and vortexing for 30 seconds, the samples were extracted using headspace SPME for 30 min at 100 C. For the Aroclor 1260 determination, the GC/MS temperature was programed as follows: 50 C (hold for 60 s), rate 1.5 C s -1 to 290 C (hold for 150 s). The whole program was complete in 380 s (< 7 minutes). The injector was maintained at 280 C and SPME fiber desorption was performed in the injection port for 1 min to prevent carryover. A constant helium flow of 1.0 ml min -1 was used. The compounds were detected by full scan mode with a scan range mass to charge ratio (m/z) 50 to 500. The electron ion (EI) source was operated at 70 ev.

135 Results and discussion Aroclor 1260 was most frequently used for method development because of its relevance in the target application site. Tuning conditions for the portable GC/MS should be performed at least daily or on every start-up using the CALION calibration mixture (Torion). Table 5-1. Paired t test results (the two tailed p-values) for the comparison of peak areas for five different PCB peaks under different conditions during headspace SPME extraction. Condition of soil during headspace SPME sampling 10 vs 30 Agit. vs no KMnO4 vs H2O vs H2O vs H2O vs min agit. 1 H2O Soil H + OH * * * * * 1 Agit. stands for agitation. *p 0.05, difference exists. Bold indicates the condition with the better recoveries Extraction condition optimization During method development, various factors that are known to affect sample recoveries were studied. Recoveries are largely affected by extraction time, fiber type, analyte volatility, solubility, and surface adsorption to particulates. KMnO4 in acid conditions has been proven to be an effective clean-up strategy for PCBs with the advantage of removing most of the co-extracted organic species and elemental sulfur [118]. Figure 5-2 and Table 5-1 summarize the results of

136 136 extraction time and the addition of wet chemicals on recoveries of selected PCBs. These results indicate that 30-minute extractions with acidified KMnO4 provide significantly better extraction recoveries for soil than the other conditions studied. Agitation may have a weak benefit. Preliminary optimization of SPME parameters were performed on a bench-top Thermo PolarisQ GC/MS [172]. Figure 5-2. Bar charts showing the influence of (A) SPME sorption time, (B) agitation, (C) addition of KMnO4 and H2O and (D) addition of acid and base on extraction efficiency of PCB 66, PCB 153, PCB 138, PCB 180 and PCB 170 from soil, as measured on the bench-top GC/MS instrument. Error bars show ±1 standard deviation (n = 3). Significance tests are shown in Table 5-1.

137 Peak identities and general performance The identification of specific Aroclors is based on the GC peak patterns and relative mass spectra. EPA 8082A standards, which contain 19 specific PCBs can be used to predict the general retention time windows of PCB homologs in Aroclors. Overall, the heavier PCBs have larger retention indices and longer retention times, although some exceptions exist. On the other hand, the comparison of mass spectra with National Institute of Standards and Technology (NIST) database can be important information for PCB homologs. Other databases may be used, such as a laboratory self-established compound library using PCB standards. A headspace SPME GC/MS chromatogram of EPA 8082A standard containing 19 PCB congeners was compared with a similar analysis of Aroclor 1260 in Figure 5-3. Retention times and fragmentation pattern similarities between the EPA standard and the Aroclor standard enable peak assignments to be made in the Aroclor mix. Peak assignment is made with the caveat that partial or complete co-elution of different PCB congeners cannot be excluded. Although it was not confirmed that each peak was a unique PCB congener, the retention times and mass spectra of the difference peaks enabled assignments to be made for the most abundant

138 138 congener present in each chromatographic peak. The tentative assignments for several of the marked peaks in Figure 2 are PCB 66, 153, 138, 180 and 170. PCB 153 PCB 66 PCB 138 PCB 170 PCB 180 Figure 5-3. Total ion current (TIC) chromatogram of headspace SPME of 40 µl of 100 ppm Aroclor 1260 (red) and EPA 8082A mix (blue) in the absence of soil matrix on the Torion Guardion -8 GC/MS. Peak assignments for the Aroclor mix (red) were made from the EI mass spectra and retention times but cannot exclude the possibility of congener co-elution.

139 139 Figure 5-4. Example of mass spectra comparison for pentachlorobiphenyl between Torion Guardion -8 GC/MS data (top) with NIST database (bottom). The chlorine isotope distributions are identifiable in both spectra, especially around m/z 254 and 326. To identify different Aroclors, all the samples must be collected, extracted and detected in the same condition and system. Figure 5-4 shows a Guardion GC/MS spectrum collected using this method and the comparison with NIST spectrum for the expected PCB. The combination of retention time, fragment ion masses and isotope

140 140 envelope all provide evidence for the peak assignment. Similar comparisons were completed for each tentatively assigned peak in the different Aroclor mixes. According to Table 5-2, the PCB distributions in different Aroclors are different and will therefore show characteristic peak patterns in resulting GC chromatograms. Light PCBs, which have a dominant proportion of 1-3 chlorine substituents in their structures, are the major PCBs in Aroclor 1221, Aroclor 1016 and Aroclor Aroclors 1254 and 1260 contain relatively more chlorinated PCBs such as penta-, hexa- and hepta- chlorobiphenyls (CBs). Tri-CBs and tetra- CBs are found to be the most abundant PCBs in Aroclor 1242 and Aroclor To differentiate Aroclor 1016 and 1232, the relative amount between tri-cbs and tetra-cbs can be used; the tri-cbs are relatively more abundant in Aroclor Similarly, to compare Aroclor 1242 and 1248, tri-cbs are more abundant in Aroclor 1242 but tetra-cbs are more in Aroclor The hepta- and octa- CBs can be used as characteristic patterns for Aroclor The differences between Aroclors 1016 and 1232, 1242 and 1248 are not very clear, so care must be taken when interpreting the results.

141 141 Table 5-2. Comparison of PCB distributions in different Aroclors [163]. Weight % in Aroclors #Cl atoms

142 Figure 5-5. Portable GC/MS chromatograms (TIC) for headspace SPME analyses of 10 µg spikes of (A) Aroclor 1016, (B) Aroclor 1232, (C) Aroclor 1242, (D) Aroclor 1248, (E) Aroclor 1254 and (F) Aroclor 1260 in the absence of soil matrix. The retention time windows of chromatograms for each Aroclor (upper chromatogram of each Aroclor) are shown from 3.1 min to 4.9 min. The lower chromatograms of each Aroclor show the same data in larger scale. (2CB: Dichlorobiphenyl; 3CB: Trichlorobiphenyl; 4CB: Tetrachlorobiphenyl; 5CB: Pentachlorobiphenyl; 6CB: Hexachlorobiphenyl; 7CB: Heptachlorobiphenyl; 8CB: Octachlorobiphenyl). 142

143 143 The method established for Aroclor 1260 was tested on the other common Aroclors to provide evidence that the method can distinguish between them. Preliminary tests were performed on Aroclor spikes added to empty vials, in the absence of soil. Recoveries from soil were shown to be ~30% those in the absence of a soil matrix, so it is expected to identify different Aroclors from soil samples at higher concentrations. The chromatograms shown in Figure 5-5 show obvious differences between the TIC patterns of the different Aroclors in the resulting chromatograms. Extracted ion chromatograms could be used to help identify specific congeners (or co-eluting structural isomers), which could be used to manually differentiate between the different Aroclors. These chromatograms provide a proof of principal that Aroclor differentiation should be possible at the level of ~10 ppm (PCB in soil) in this portable system Quantitation Calibration curves were collected on the portable GC/MS system to assess the linearity of the response function near the limits of detection of the instrument. Aroclor 1260 standard solutions with different concentrations were prepared in empty glass vials (no soil, water or modifiers) and analyzed. Different volumes of 100 ppm

144 144 Aroclor 1260 solution were spiked into separate 10-mL vials. After the samples were dried in the hood under the room temperature to remove the solvent, they were sealed and then extracted by SPME for 30 min under 100 C. All the samples were analyzed using the same GC program on portable GC/MS. The results of the calibration curves collected on the Portable GC/MS and bench-top GC/MS are shown in Figure 5-6. The five major PCBs identified in Figure 5-6 show linear relationships between the concentration and instrument response (peak height or area) on both instruments. However, the portable GC/MS instrument had significantly higher (worse) detection limits and had poorer correlation values (expressed as R). Whereas the benchtop Polaris Q GC/MS consistently provided R values greater than 0.98, the portable GC/MS gave weaker correlation scores, but still exceeding It was found that the reproducibility of peak heights (or peak areas) of PCB congeners was worse on the portable instrument relative to the bench-top instrument; peak areas had percent relative standard deviations on the order of 20% on the portable system. This variance could contribute to the poorer linearity observed on the portable instrument.

145 Figure 5-6. Comparisons of headspace SPME calibration curves of Aroclor 1260 in the absence of soil matrix for the peak tentatively assigned as PCB 180 collected on (A) Portable Torion Guardion -8 GC/MS (R = 0.96) and (B) Bench-top Thermo Polaris Q GC/MS (R = 0.98). Note that the bench-top calibration curve covers significantly lower quantities. 145

146 Applications The real soil samples were collected at the former Portsmouth Gaseous Diffusion Plant, Portsmouth, Ohio USA. The samples were split for analysis between our laboratory and a commercial service laboratory (GEL Laboratories, LLC (GEL), Charleston, SC, USA). Five replicates were performed on each soil sample and the same soil samples were analyzed by GEL using EPA 8082 method. The concentrations in the real soil samples were too low (< 0.2 ppm) to be determined by our method. Instead of analyzing the authentic samples, simulated soil samples that contained 10 ppm Aroclor 1260 were prepared for the on-site analysis demonstration. The outdoor demonstrations were performed at Endeavor Center, Piketon, OH and at Dairy Lane Park, Athens, OH in July 2012 and April 2013, respectively. Aroclor 1260 was identified and determined by our proposed method in the field. An example of Torion GC/MS chromatograms and mass spectrum for 10 ppm Aroclor 1260 in simulated soil sample is shown in Figure 5-7. Only semi-quantitative analysis of Aroclors was possible on-site because of the difficulty in establishing the moisture content of the soil: concentrations can only meaningfully be reported relative to dry mass of soil, which cannot be assessed on-site.

147 147 Figure 5-7. GC chromatogram of Aroclor 1260 (top) and representative MS spectrum of PCBs (bottom) for 10 ppm Aroclor simulated soil sample. 5.3 Analysis of PCDDs and PCDFs The headspace SPME method used for the various Aroclors was also applied to determine PCDDs and PCDFs. According to EPA method 8280B, PCDDs and PCDFs mix was purchased from AccuStandard Inc, which contained 5 PCDDs and 5 PCDFs. The PCDDs mix standards were tested on the portable GC-TMS system. The GC program was optimized to 10 min. A spike consisting of 50 µl of 5 ppm PCDDs and PCDFs standard solution were added into 10 ml vial and dried in a fume hood. The samples were sealed and then

148 148 extracted by SPME using 100 µm PDMS fiber for 30 min at 100 C. The samples were analyzed using the Torion Guardion portable GC- TMS system. According to the results in Figure 5-8, 5 of the 10 compounds from the PCDDs and PCDFs mix standard can be identified based on their retention times and mass spectra. The other 5 relatively heavy PCDDs and PCDFs mix standards were not detectable. The main reason could be that the sensitivity of the instrument was much lower than bench-top instrument, especially for heavy PCDDs and PCDFs. The other reason should be the extraction efficiency for heavy PCDDs and PCDFs was worse than lighter components. The results of analyzing PCDDs and PCDFs by bench-top Thermo PolarisQ GC-MS system (Figure 5-9) demonstrated that the extraction efficiency of dioxins and furans from wet soil or dry soil with 2.5 ml 0.2 M KMnO4 and 0.25 ml 6 M H2SO4 solution was extremely low. All the results above indicated that it was difficult to extract the dioxins and furans from soil using this method. By reviewing the literature, SPME could be hardly found, which also indicates the difficulty of dioxin extraction from soil by this method.

149 Figure 5-8. Dioxin mix standards tested on portable Torion Guardion 8 GC-TMS system using headspace SPME extraction. Spike consisted of 50 µl of a 5 ppm stock solution of EPA 8280B mix in the absence of soil matrix. 149

150 Figure 5-9. Dioxin mix standards tested on bench-top Thermo Polaris Q GC-MS system using headspace SPME extraction. Spike consisted of 50 µl of a 5 ppm stock solution of EPA 8280B mix with different matrices. The dry dioxin standard (top figure) was measured in the absence of any chemical modifiers or matrix. The dry dioxin-spiked soil (bottom figure), was analyzed with and without treatment with acidified KMnO4. Wet dioxin-spiked soil (bottom figure) was also extracted following treatment with KMnO4 + H

151 Conclusions An on-site analysis method for the determination of PCBs and Aroclors by portable GC/TMS with SPME was developed in this study. Potassium permanganate acid solution was used to assist the PCB releasing from soil matrix. The method has the advantage of the high sample throughput, with a soil sample being prepared and analyzed about every 37 min. By adapting the headspace SPME method with portable scale and heating block, the on-site sampling and sample preparation can be perform on the field. Although the capability of the quantitative analysis for PCBs and Aroclors using this method is understated, this method can be very beneficial and cost-effective for the fast decision of environmental sample investigation. The method is not suitable for the analysis of PCDDs and PCDFs, because the extraction efficiency of PCDDs and PCDFs was extremely low by SPME.

152 152 CHAPTER 6: DETERMINATION OF TRICHLOROETHYLENE IN WATER USING LIQUID-LIQUID MICROEXTRACTION ASSISTED SOLID PHASE MICROEXTRACTION AND CLASSICAL LEAST SQUARES RESOLUTION OF OVERLAPPING TRICHLOROETHYLENE AND ITS ISOTOPIC INTERNAL STANDARD PEAK CLUSTERS A method for the determination of trichloroethylene (TCE) in water using portable gas chromatography/mass spectrometry (GC/MS) was developed. A novel sample preparation method, liquid-liquid microextraction assisted solid phase microextraction (LLME-SPME), is introduced. In this method, 20 µl of hexane was added into 10 ml of TCE contaminated aqueous samples to assist headspace SPME. The extraction efficiency of SPME was significantly improved with the addition of minute amounts of organic solvents (i.e., 20 µl hexane). Two chemometric methods, classical least-squares (CLS) and inverse least-squares (ILS), were applied to resolve overlapping TCE and deuterated TCE (TCE-d) mass spectral signals and evaluated for the determination of TCE. CLS and ILS models were constructed using the data sets from the samples containing TCE and TCE-d at different concentrations, and were used to predict concentration ratios of TCE and TCE-d. Calibration samples were prepared by adding TCE at different concentrations and TCE-d at 300 ng ml -1 as the internal

153 153 standard. A linear calibration model was constructed between predicted concentration ratios of TCE/TCE-d and prepared concentrations of TCE in aqueous samples with a weighting factor that accentuated samples with lower concentrations. Predictions were obtained from a validation data set. Lower prediction errors and higher coefficients of determination (R 2 ) were obtained from TCE/TCEd ratios predicted by the CLS model. The calibration curve was linear over a range of ng ml -1 with a R This paper presents the first application of chemometrics to overcome overlapping peaks between an analyte and its corresponding isotopic internal standard. 6.1 Materials and methods Reagents and materials. Analytical grade trichloroethylene (TCE, 99.5%), benzene, sodium chloride (NaCl), sodium sulfate (Na2SO4), SPME fibers coated with polydimethylsiloxane (PDMS, 100 µm film thickness), carboxen/pdms (75 µm film thickness), or PDMS/divinylbenzene (PDMS/DVB, 65 µm film thickness), 20-mL headspace glass vials, and crimp seals with PTFE/silicone septa were purchased from Sigma- Aldrich Co. LLC. (St. Louis, MO, USA). Deuterated TCE was obtained from C/D/N Isotopes INC. (Pointe-Claire, Quebec, Canada).

154 154 TCE standard solutions for calibration were prepared in acetonitrile at the following concentrations: 5, 15, 50, 150, and 500 µg ml -1. Water samples for calibration and validation were prepared by adding 20 µl of the TCE standard solutions to 10 ml D.I. water or groundwater. This procedure yields the final concentrations of 10, 30, 100, 300, and 1000 ng ml -1. TCE-d solutions at 15, 50, 150, and 500 µg ml -1 were prepared in acetonitrile Instruments. The portable TRIDION-9 GC-TMS instrument (Torion Technologies, American Fork, UT, USA) consists of a low thermal mass (LTM) GC and a miniature toroidal ion trap mass analyzer with a disposable helium cartridge and rechargeable battery. In this study, the column was an MXT-5, 5 m 0.1 mm i.d. capillary column chemically bonded with 5% diphenyl/95% dimethyl polysiloxane and 0.4 µm film thickness. The injection port was held at 270 C and split mode was used with split ratio at 1:10. The oven temperature was programmed as follows: 50 C, hold for 10 s, ramp at 2 C s -1 to 250 C, hold for 10 s. A constant helium flow of 1.0 ml min -1 was used and the total GC run time was 2 min. The transfer line and ion source temperatures were both maintained at 270 C. The mass spectrometer was operated in positive ion electron ionization (EI)

155 155 mode at 70 ev and mass spectra at full scan mode with the scan range from mass-to-charge ratio (m/z) 49 to 527 were collected starting from 0.39 min after injection. All the data were processed using MATLAB R2013b (MathWorks Inc., Natick, MA, USA). The extraction optimization process was performed on a benchtop Thermo Finnigan PolarisQ quadrupole ion trap mass spectrometer/trace GC system with a Triplus AS2000 autosampler (San Francisco, CA, USA). Table 6-1. Concentration matrix of modelling set (n = 3) Concentration of TCE Concentration of TCE-d Sample ID (ng ml -1 ) (ng ml -1 )

156 Data collection. CLS and ILS models were constructed using the modeling data sets from the samples containing TCE and TCE-d at different concentrations in triplicate. The concentration matrix (21 2) TCE and TCE-d is given in Table 6-1. A magnetic stir bar and 10 ml of water sample were placed into a 20-mL headspace glass vial with addition of 20 µl of each TCE standard solution, TCE-d solution, and hexane. After the vial was sealed by an aluminum cap with a PTFE/silicone septum, a PDMS/DVB fiber was exposed to the headspace for 15 min at 25 C. The fiber was then immediately inserted into the GC injector for desorption at 270 C for 5 s of the portable GC/MS. For calibration and validation samples, the sample was prepared in the same way except that concentration of TCE-d IS solution was 300 ng ml -1 for all the samples. The calibration spectra at five standard concentrations were collected in triplicate with a random block design Data format. The two-way GC/MS data sets were initially acquired as network common document format (CDF) which were read directly into MATLAB using the netcdf tools. The data sets were binned by retention time

157 157 from 0 to 2.01 min with a min increment and binned by massto-charge ratios from 49 to 527 Th with a 1 Th increment. Therefore, each two-way GC/MS object comprised data points in which 1006 rows corresponded to the retention times and 479 columns corresponded to mass-to-charge ratios Retention time alignment. Retention time (RT) variation among the chromatograms was observed and ineluctable because of the manual injections. RTs in the modeling set were aligned with an algorithm that uses a polynomial to align the mass spectra with respect to retention time so that the correlation with the mean spectrum is maximized. Each object in all other data sets was aligned to the two-way average of the modeling set [147].

158 158 Step 1: Construct CLS/ILS model using modeling data set to determine TCE and TCE-d ratios (Modeling data set: 7 3 = 21) Concentrations of TCE and TCE-d in modeling data (ng ml -1 ) TCE TCE-d Concentrations of TCE and TCE-d in calibration data (ng ml -1 ) Step 2: Concentrations of TCE and TCE-d in validation data (ng ml -1 ) TCE TCE-d Predict concentrations ratios of TCE and TCE-d in calibration data set Predict concentrations ratios of TCE and TCE-d in validation data set (Calibration data set: 5 3 = 15) (Validation data set: 3 3 = 9) TCE TCE-d TCE/TCE-d Concentration Ratio y = 0.06x 0.21 r 2 = TCE concentration (ng ml -1 ) Step 3: Step 4: Construct calibration curve by plotting TCE/TCE-d ratios with TCE concentrations using calibration data set Predict TCE concentrations for validation data set from the calibration curve TCE Concentrations (ng ml -1 ) 28±3 88±5 330±15 Note: concentration of TCE-d in calibration and validation data sets is fixed to 300 ng ml -1, but concentration of TCE-d in modeling data set varies (Supporting Information Table S1). Figure 6-1. Flowchart of data analysis.

159 Data processing. Besides the original two way data sets (two-way), compressed two-way data sets (c-two-way) that contain data in a retention time window between 0.39 and 0.44 min and a mass range from m/z 130 to 137 were constructed and evaluated for all data sets. The flowchart for data processing is given in Figure 6-1. The modeling data set (7 concentrations in triplicate) were pretreated by normalization to unit vector length and RT alignment. The CLS and ILS models were constructed using the modeling set of data. After the calibration data set (5 concentrations in triplicate) and validation data set (3 concentrations in triplicate) were pretreated by RT alignment, ratios of TCE/TCE-d in the calibration data set and validation data sets were predicted by both CLS and ILS models. A linear calibration of the predicted ratios of TCE/TCE-d with respect to the prepared concentrations of TCE was constructed from the calibration set using a weighting factor of the inverse of the squared concentrations ( 1 ) to give greater emphasis to the lower concentrations [173]. The TCE concentrations of the validation data set were calculated from the calibration models. As a reference method, PLS model was constructed from calibration data set. The concentration matrix (15 2) for the

160 160 calibration set using PLS were the concentrations of TCE and TCE-d. The concentrations of TCE in validation data sets were predicted by the constructed PLS model. For comparison purpose, both calibration and validation data sets using the PLS method were pretreated the same way as mentioned above (e.g., RT alignment, data compression). 6.2 Results and discussion LLME-SPME method optimization. Many factors could affect the LLME-SPME process. Some of them are optimized and discussed in this study including selection of extraction solvent, volume of extraction solvent, extraction time and temperature, SPME fiber coatings, and effects of dispersive solvent, stirring, and salt. The portable GC/MS system can only perform manual SPME mode. The extraction optimization process was performed on a bench-top GC/MS instrument equipped with autosampler because the autosampler can reduce the systematic errors comparing with using manual injection mode. The bench-top instrument is also more stable than the portable instrument, and enable the liquid injection mode which can be used to evaluate the absolute recoveries of SPME and LLME-SPME methods. The peak area was calculated by integration in the TCE retention time window of the extracted molecular ion m/z 132 chromatogram. The peak area was

161 161 used to compare the extraction efficiencies obtained from different extraction conditions Selection of organic solvent. Additional organic solvents are usually avoided when applying SPME because the SPME fiber may become saturated with organic solvents (e.g., hexane) instead of the analyte of interest. In our study, it was found that the extraction efficiency could be significantly improved with the addition of minute amounts (i.e., microliters) of organic solvents. Several commonly used organic solvents in liquidliquid extraction were selected including pentane, hexane, benzene, chloroform, ethyl ether, and ethyl acetate. A series of 10-mL TCE water samples containing 20 µl each of organic solvent were prepared. For ethyl ether and ethyl acetate, another set of samples with 500 µl of each in 10 ml aqueous TCE standards were prepared for comparison because their solubilities in water are relatively high. The extraction temperature at 25 C and extraction time at 15 min were used as LLME-SPME extraction conditions. The responses obtained from the different solutions are given in Figure 6-2A. The response was significantly higher with the addition of 20 µl of hexane.

162 162 The volume effect of hexane was also investigated and results are given in Figure 6-2B. The optimum volume for hexane was 20 µl. Therefore, 20 µl hexane was used in this study. The putative mechanism for the enhanced SPME efficiency by LLME is that the LLME enriches the TCE from aqueous solution into an organic film on the solution surface. When organic solvents with the densities less than 1 g cm -3 (e.g., hexane) are selected, the organic film resides at the solution/headspace interface. The TCE enriched film has a greater mass transfer efficiency to the headspace. Solvents denser than water such as chloroform did not provide an enhancement of the SPME efficiency, because the denser TCE enriched organic film was not in contact with the headspace even with stirring. Compared with hexane, pentane has better volatility but cannot improve the extraction efficiency that much. The reason could be that the fiber was overwhelmed by pentane and left insufficient sites on the fiber for the analyte (TCE). Other than density and volatility, the selection of organic solvent should also consider the retention properties of the organic solvent. The retention index (RI) can be used as a criterion to select an organic solvent. The RI of the analytes should be larger than the RI of the organic solvent so that the solvent delay period will not include any

163 163 analyte peaks. In practice, the RI value of the analyte is recommended to be at least 30 larger than that of the organic solvent. Table 6-2 listed the RIs for TCE and the organic solvents tested in our study. In this study, the solvent peak of hexane can be fully separated from the TCE peak with respect to retention time. Table 6-2. Retention indices and densities of selected organic compounds Compound Name RI* Density (g cm -3 ) TCE Ethyl ether Ethyl acetate Hexane Chloroform Benzene *RI: retention index from NIST database

164 164 A TCE peak area x 10 5 B x 10 5 TCE peak area Volume of hexane (µl) Figure 6-2. Effect of different extraction solvent for LLME-SPME (A) and effect of hexane volume for LLME-SPME with 95% confidence intervals (n = 3) (B). TCE peak area: m/z 132 (molecular ion of TCE) was extracted from total ion current chromatogram and integrated in TCE retention time window.

165 165 Extraction temperature ( C) Two factors central composite design Extraction time (min) Figure 6-3. Data points in the two factors (extraction time and extraction temperature) central composite design. Figure 6-4. Response surface of the second-order polynomial model with the zoom-in window of interested region (A) and response surface model at 15 min (B).

166 166 A x Hex TCE peak area µl ACN & Hex 500 µl MeOH & Hex 500 µl Acetone & Hex 2 0 B x TCE peak area C Volume of Acetonitrile (µl) x 10 5 LLME-SPME with Hexane LLME-SPME with Acetonitrile and Hexane TCE peak area Extraction time (min) Figure 6-5. Effect of dispersive solvents (A), volume of acetonitrile (B) and extraction time (C) on LLME-SPME extraction efficiency. (n = 3) Hex: hexane; ACN: acetonitrile; MeOH: methanol. Note that in Figure 4B, volume of acetonitrile refers to additional volume of acetonitrile as dispersive solvents added to the solution, and the acetonitrile in TCE standard solution was not counted.

167 Effect of extraction temperature and extraction time. The extraction temperature and extraction time are usually interacted factors [174]. The full second-order polynomial models are versatile in many systems over a limited factors, and the central composite designs are very useful for obtaining data to fit the full second-order polynomial models [175]. Figure 6-3 is a central composite design used in our study for two experimental conditions: extraction time and extraction temperature. The model for the response surface is given as equation (6-1). The polynomial model is fit to the response values obtained from the central composite design. = (6 1) for which is the response that is the peak area of the TCE; are the coefficients for the model; is the extraction time and is the extraction temperature. Figure 6-4A is the contour plot of the modeled response surface. The best extraction result according to the model occurs at 15 C and 60 min. For a high throughput method, the extraction time of 60 min is too long. If a 15 min extraction time is used, the optimum extraction temperature is 31 C according to the fitted model (Figure 6-4B). The peak area at the condition of 15 min and 31 C is about 82.4% of the best extraction condition at 60 min and 15 C.

168 168 Therefore 45 minutes (75% of the best extraction time) are saved with an 18% loss of peak area. For the extraction temperature greater than 31 C, the extraction efficiency decreases as the temperature increases, which agreed with the result in a previous study [136]. In our study, the extraction temperature of 31 C and extraction time of 15 min were chosen Effect of dispersive solvents. In regular LLME, a dispersive solvent such as methanol, acetone, or acetonitrile with high miscibility in both extractant and aqueous phases can give rise to the formation of small droplets throughout the aqueous sample. The extraction time can be shortened because of the increased surface area between the extractant and aqueous sample in the cloudy solution, so the equilibrium is achieved quickly [176]. The extraction efficiencies by LLME-SPME with 20 µl hexane and 500 µl of different dispersive solvents including methanol, acetone, and acetonitrile were compared and the results are reported in Figure 6-5A. None of the dispersive solvents improved the extraction efficiency. The volume effect of acetonitrile as a dispersive solvent can be seen in Figure 6-5B. There was no significant difference for extraction efficiency when using 0, 100, 200, or 500 µl acetonitrile (p-value of 0.2 by one-way analysis of variance).

169 169 Different extraction times in the range of 5-90 min for SPME with 20 µl hexane or with 20 µl hexane and 100 µl acetonitrile were evaluated (Figure 6-5C). The maximum response was achieved at 60 min for LLME-SPME with hexane, and at 30 min for LLME-SPME with hexane and acetonitrile. No advantage to using acetonitrile was achieved, especially at the pre-selected extraction time of 15 min Other factors: SPME fiber, stirring, and salting out. Coatings of SPME fiber were selected among PDMS (100 µm film thickness), carboxen/pdms (75 µm film thickness) and PDMS/DVB (65 µm film thickness). The PDMS/DVB fiber was chosen because better recovery of TCE were achieved (Figure 6-5 A). Responses for TCE in non-stirred samples was about 50% of those obtained in stirred samples (Figure 6-5 B), so stirring was used. Increasing the ionic strength by adding 3 g NaCl or Na2SO4 did not influence the efficiency of the extraction (Figure 6-5 C), therefore the addition of salt was not considered in the experiments.

170 170 A 15 x µm PDMS/DVB TCE peak area µm Carboxen/PDMS 100 µm PDMS B x 10 5 TCE peak area C 0 18 x 10 5 Stir Non-stir 15 TCE peak area No salt NaCl Na 2 SO 4 Figure 6-6. Effects of fiber coatings (A), stirring (B) and salting out (C) with 95% confidence intervals. (n = 3).

171 Recoveries and enrichment factors To evaluate the absolute recovery of the LLME-SPME method, another calibration data set using standard liquid injection was collected across the range of ppm. The recovery was calculated using the calculated TCE mass on-column of the LLME- SPME-extracted sample relative to the absolute TCE mass contained within the vial before extraction. The enrichment factor of LLME-SPME method was defined as the ratio of the calculated TCE masses oncolumn from the samples extracted by LLME-SPME with 20 µl hexane and without hexane (SPME). The recovery and enrichment factor results are listed in Table 6-3. The absolute recoveries are in the range of 29-41% for the samples extracted by LLME-SPME and 11-17% for the samples extracted by SPME. The enrichment factors with the addition of hexane are 2.6±0.2, 2.4±0.4, and 2.2±0.3 for the samples at low, medium, and high concentrations.

172 172 Table 6-3. Absolute recoveries and enrichment factor of TCE by LLME-SPME (n = 3) TCE TCE in SPME LLME-SPME Enrichment concentration the vial TCE on Recovery TCE on Recovery factor (ng ml -1 ) (ng) column (ng) (%) column (ng) (%) ± ± ± ± ± ±

173 A TIC Before RT Alignment Response After Normalization Retention Time (min) B TIC After RT Alignment Retention Time (min) Figure 6-7. Effect of RT alignment: TCE chromatograms before RT alignment (A) and TCE chromatograms after RT alignment CLS and ILS model and data analysis CLS and ILS models were constructed from the modeling data set. Compressed two-way data sets (c-two-way) were compared with the original two way data sets (two-way). The benefits of this data set are: first, it is smaller which reduces the calculation time; second, it is more specific because all possible molecular masses for TCE and TCE-d are in the range of 130 to 137 Th. RT drift among different runs is

174 174 inevitable when manual injection with a portable instrument was used. The width of the RT window for the TCE peak is only about 0.03 min, so it is very sensitive to RT drift (Figure 6-6 A). The RT alignment effect is demonstrated in Figure 6-6, and the chromatograms have been successfully aligned.

175 175 Table 6-4. Effect of normalization, RT alignment and two data representations on RMSPE by CLS RT two-way (ng ml -1 ) c-two-way (ng ml -1 ) Normalization alignment TCE TCE-d TCE TCE-d RMSPE * 0.05 *

176 176 The effects of RT alignment, normalization, and two different data representations on the root mean square of prediction error (RMSPE) of TCE and TCE-d for modeling data are reported in Table 6-3. The minimum errors for both TCE and TCE-d were achieved with the CLS model after normalization and RT alignment by using the c- two-way data representation. The calculation time was reduced from 52 min to 7 min when the c-two-way data was used instead of the full two-way data. It should be noted that the calculation time included the normalization and the RT alignment process for both modeling data set (21 objects) and calibration data set (15 objects). Therefore, all data sets were preprocessed by normalization and RT alignment and the c-two-way data representation was used further. TCE/TCE-d ratios of calibration data set were predicted by established CLS and ILS models. Calibrations were constructed by linear regression calibration with a weighting factor ( 1 ) or ordinary linear regression without using weighting factor between TCE concentrations and TCE/TCE-d ratios. Calibration curves from the CLS models were linear using both regression methods and gave relatively good fits (coefficients of determination (R 2 ) greater than 0.993). Calibration curves from the ILS models were not linear using regression with the weighting factor and gave lower coefficients of

177 177 determination (R 2 less than 0.707) but linear using ordinary linear regression (R 2 > 0.997). However, much greater relative errors (RE) were observed by using ordinary linear regression, especially for samples at lower concentrations (Table 6-4). For ordinary linear regression the larger deviations present at higher concentration have a stronger influence (weight) on the calibration, so the accuracy for lower concentration samples were lower [173]. Linear regression with a weighting factor of 1 was used and TCE/TCE-d ratios predicted by CLS model gave better performance with respect to lower errors and higher correlation coefficients. The construction of the model to resolve the overlapping isotopic peaks between the analyte and the IS is important for this method and the modeling data should be collected carefully. Once an accurate model was constructed, the method would be applicable for field analysis because the use of internal standard (TCE-d) could reduce the systematic error. For the reference method, the coefficient of determination (R 2 ) from PLS model was less than 0.93 which indicated that the direct application of PLS to TCE/TCE-d data sets did not perform satisfactorily. Larger errors were obtained when using the PLS model to predict TCE concentrations from the validation data sets (Table 6-

178 178 5). Therefore, the proposed method performed better than the reference PLS method.

179 179 Table 6-5. Predicted concentrations by weighted least squares regression and linear least squares regression for calibration set (n = 3) Cal data from CLS Cal data from ILS model Cal data from CLS Cal data from ILS model Added model by weighted LS by weighted LS model by linear LS by linear LS concentration (ng ml -1 ) Measured concentration (ng ml -1 ) RE (%) Measured concentration (ng ml -1 ) RE (%) Measured concentration (ng ml -1 ) RE (%) Measured concentration (ng ml -1 ) RE (%) 10 10±1 2±11 15±23 47±228-6±1-156±12 50±10 400± ±4-7±12-12±37-140±123 13±4-55±14 30±20 10± ±1-5±1 73±16-27±16 85±1-14±1 80±9-20± ±10 17±3 420±122 40±41 364±11 21±4 260±66-10± ±200-7± ±500 79± ±200-2± ±300 1±28

180 180 Table 6-6. Prediction results of synthetic water (n = 3) and TCE contaminated river water (n = 4) with PLS regression Added Measured concentration (ng ml -1 ) concentration (ng ml -1 ) RE (%) Synthetic sample ± ± ± ± ±70 70±20 River sample ± ± ±50 17±8

181 181 A B TCE/TCE-d Ratio 8 6 y = 0.06x r = Concentration (ng ml -1 ) Predicted concentration (ng ml -1 ) Spiked tap water samples Spiked river water samples Prepared concentration (ng ml -1 ) Figure 6-8. Example calibration curve for TCE/TCE-d ratios predicted by CLS model with respect to prepared TCE concentrations (A) and a plot of predicted concentration with respect to prepared concentration of TCE in water samples; the line is a reference line (B) Applications Spiked tap water samples at 3 concentrations were analyzed in triplicate with the proposed method. Real water samples were collected from Hocking River, Athens, Ohio; were extracted using LLME-SPME; and the extracts were analyzed with the portable GC/MS instrument. The results revealed that the river water was free of detectable TCE contamination. TCE contaminated river water samples

182 were prepared by spiking with the TCE standard solutions. TCE/TCE-d ratios were predicted with the established CLS model and the concentrations of TCE were calculated from the calibration constructed using linear regression with a weighting factor of 1. A representative calibration curve and plot of the predicted TCE concentrations are given in Figure 6-7. The prediction results of the river water samples are reported in Table 6-6. The relative errors ranged between -12%-10% Effectiveness of LLME-SPME on other volatile organic contaminants in water. The toluene and TCB were selected to test the effectiveness of LLME-SPME method because both are among the top 15 most 182 frequently detected VOCs in aquifers [177]. Water samples contain 20 ng ml -1 of toluene and TCB each were used. The extraction conditions for toluene and TCB are the same as the optimum conditions for TCE. The peak areas of extracted molecular ions (e.g., m/z 92 for toluene, m/z 182 for TCB) were used to compare the extraction efficiencies between regular SPME method and LLME-SPME methods with hexane, pentane, and chloroform. The results are graphed in Figure 6. For both toluene and TCB, the LLME-SPME with hexane method gives the better extraction efficiency than the other methods (p-values of 1.5

183 and 0.06 for toluene and TCB by paired t-test evaluation between the largest peaks and second largest peaks with 95% confidence intervals). The extraction efficiencies for toluene and TCB by LLME- SPME method would be further improved after optimization. Therefore the LLME-SPME method could be used for the analysis of various VOCs in aqueous matrices. Table 6-7. Prediction results of spiked tap water (n = 3) and TCE contaminated river water (n = 4) Added Measured Relative error concentration concentration (%) (ng ml -1 ) (ng ml -1 ) Spiked tap water sample 30 28± ± ±15 10 River sample 20 22± ±60-1

184 184 A Peak area Toluene SPME LLME-SPME with hexane LLME-SPME LLME-SPME with pentane with chloroform B Peak area SPME TCB LLME-SPME with hexane LLME-SPME LLME-SPME with pentane with chloroform Figure 6-9. Comparison of regular SPME and LLME-SPME for the extraction of toluene (A) and TCB (B). (n = 3) 6.3 Conclusion In this study, a novel sample preparation method LLME-SPME was developed. This method significantly improved the extraction efficiency compared with SPME and would be suitable for field analysis because of its simplicity. Different organic solvents were compared and hexane was selected because of the best extraction efficiency offered. The response surface for extraction temperature and time was modeled by fitting the full second-order polynomial model to the peak areas obtained from a central composite design. For a fast screening method, non-equilibrium extraction with 15 min as extraction time was used and 31 C was selected as the optimum temperature at this condition. Other parameters such as SPME fiber coatings, and effects of dispersive solvent, stirring, and salt were also optimized.

185 CLS was demonstrated for the first time as a method to resolve overlapping isotopic peak clusters between an analyte and its isotopic internal standard, thereby allowing the use of less expensive 185 deuterated standards. The CLS model resolved TCE and TCE-d spectra and then a univariate linear regression calibration with weighting factor of 1 was applied to quantify the TCE in water samples. The proposed method enables simple isotopic analogs of analytes (one H or C atom is isotopic labeled) to apply as internal standards. It has wide applications especially when analytes have isotopic distributions that would overlap with an internal standard or when sophisticated isotopic analogs of the analytes with three or more 2 H or/and 13 C- atoms are prohibitively expensive or even impossible to obtain.

186 186 CHAPTER 7: SUMMARY AND FUTURE WORK The determination of PCBs is difficult and costly because the PCB congeners are similar in structure and great effort is needed to separate them. PCBs were produced and disposed of in the environment as Aroclors, so the quantification of Aroclors is important. The application of chemometric methods (i.e., FuRES, PLS-DA, PLS) enables to identify, classify and quantify the Aroclors without the full separation of PCB congeners and complicated sample preparation. In Chapter 2, a headspace SPME method was developed for the extraction of PCBs from soil matrices. The SPME-GC.MS method was validated on the determination of single Aroclor (i.e., Aroclor 1260) and was applied to the classification of 7 different Aroclors in Chapter 3 and the quantification of 2 Aroclor mixtures (i.e., Aroclor 1254 and 1260) in Chapter 4. The developed headspace SPME method was also used in Chapter 5 in a field study for the analysis of PCBs with portable GC/MS instrument. During the study of analyzing PCBs by portable SPME-GC/MS system, we found that the quantitative analysis of PCBs was understated which might be caused by the complexity and low volatility of PCBs. Therefore a simpler and more volatile compound TCE was attempted to be analyzed using portable SPME-GC/MS system and promising result was achieved. The field method

187 187 development for the determination of TCE was described in Chapter 6. A novel extraction method, LLME-SPME, was introduced and fully discussed. This method can significantly improve the extraction efficiency comparing with the regular SPME method for TCE in water. Because the extraction efficiency of LLME-SPME for TCE varies at different concentrations, a suitable internal standard should be used for the quantitative study. To overcome the cross-contribution problem which is caused by intensity contribution in mass spectra between TCE and deuterated TCE, two chemometric methods, CLS and ILS, were evaluated for resolving the overlapping isotopic mass peak clusters between TCE and TCE-d. CLS results have demonstrated a success of this approach and enable the application of TCE-d as internal standard for the quantitation of TCE. In the study of classifying 7 different Aroclors in Chapter 3, the method cannot classify samples with Aroclor mixtures. In the future, the classification on the samples with Aroclor mixtures should be studied. Modulo compression was applied as data preprocessing method in this study and showed improvement on classification rates and sensitivity. Although modulo compression was developed in 1968, it has not been widely used. It may be an important data preprocessing method and have wide application on the classification

188 188 of a category of compounds. For the quantification study of Aroclors in Chapter 4, two Aroclors were quantified. The quantification of samples with more than two Aroclors should be investigated in the future study. The methods developed in Chapter 2, 3 and 4 are simple and fast which is possibly applied to the on-site analysis of PCBs. To develop and validate a field method, more efforts are need besides of the study in Chapter 5. The differences of PCB distributions between Aroclors in portable GC/MS TIC profiles has been observed. It would be more beneficial if the analysis of different Aroclors was performed in real matrices and the chemometric methods were applied to the study for the classification of Aroclors. Moreover, a suitable internal standard should be selected and could make quantitation of Aroclors using portable SPME-GC/MS system possible. In Chapter 6, LLME- SPME has been developed and applied to extracting TCE from water. The underlying physical chemistry governing the extraction enrichment is not clear. More studies are needed to clarify the mechanism of LLME-SPME process. It is also valuable to apply this method to other compounds to demonstrate the capability of LLME-SPME on the extraction if similar successes on extraction efficiency are achieved. The result of the application of chemometric methods such as CLS and ILS to discriminating analyte (i.e., TCE) from its isotopic internal

189 189 standard (i.e., TCE-d) is convincing. It should be applied to the quantitative studies of more complex compound such as PCBs in the future. Other chemometric methods should be evaluated to resolve overlapped isotopic mass peaks and may give better results.

190 190 REFERENCES [1] D. Consonni, R. Sindaco, P.A. Bertazzi, Blood levels of dioxins, furans, dioxin-like PCBs, and TEQs in general populations: A review, , Environ Int, (2012). [2] R.A. Hites, Dioxins: an overview and history, Environ Sci Technol, 45 (2011) [3] UNEP, Final act of the conference of plenipotentiaries on the Stockholm convention on persistent organic pollutants. United Nations Environment Programme, Geneva, Switzerland, (2001) 44. [4] U. EPA, Ambient water quality criteria for polychlorinated biphenyls., U.S. Environmental Protection Agency, Washington, DC, [5] M. Schmitz, G. Scheeder, S. Bernau, R. Dohrmann, K. Germann, Dioxins in primary kaolin and secondary kaolinitic clays, Environ Sci Technol, 45 (2011) [6] G.F. Fries, Transport of organic environmental contaminants to animal products, Rev Environ Contam Toxicol, 141 (1995) [7] G.F. Fries, D.J. Paustenbach, D.B. Mather, W.J. Luksemburg, A congener specific evaluation of transfer of chlorinated dibenzo-pdioxins and dibenzofurans to milk of cows following ingestion of

191 191 pentachlorophenol-treated wood, Environ Sci Technol, 33 (1999) [8] M. Van den Berg, L. Birnbaum, M. Denison, M. De Vito, W. Farland, M. Feeley, H. Fiedler, H. Hakansson, A. Hanberg, L. Haws, M. Rose, S. Safe, D. Schrenk, C. Tohyama, A. Tritscher, J. Tuomisto, M. Tysklind, N. Walker, R.E. Peterson, The 2005 world health organization reevaluation of human and mammalian toxic equivalency factors for dioxins and dioxin-like compounds, Toxicol Sci, 93 (2006) [9] A. Demond, A. Franzblau, D. Garabrant, X. Jiang, P. Adriaens, Q. Chen, B. Gillespie, W. Hao, B. Hong, O. Jolliet, J. Lepkowski, Human exposure from dioxins in soil, Environ Sci Technol, 46 (2012) [10] M.T. Ahmed, Y.Y. Mosleh, H. El Sharabasy, A. Diab, P. Tundo, Risk assessment of some PCDDs, PCDFs, dioxin-like PCBs, and PCBs in contaminated dairy products: An Egyptian pilot case study, Hum Ecol Risk Assess, 13 (2007) [11] S.H. Safe, Comparative toxicology and mechanism of action of polychlorinated dibenzo-p-dioxins and dibenzofurans, Annu Rev Pharmacol Toxicol, 26 (1986) [12] A. Poland, J.C. Knutson, 2,3,7,8-tetrachlorodibenzo-p-dioxin and related halogenated aromatic hydrocarbons: examination of the

192 192 mechanism of toxicity, Annu Rev Pharmacol Toxicol, 22 (1982) [13] B.A. Vorderstrasse, L.B. Steppan, A.E. Silverstone, N.I. Kerkvliet, Aryl hydrocarbon receptor-deficient mice generate normal immune responses to model antigens and are resistant to TCDD-induced immune suppression, Toxicol Appl Pharmacol, 171 (2001) [14] J. Mimura, K. Yamashita, K. Nakamura, M. Morita, T.N. Takagi, K. Nakao, M. Ema, K. Sogawa, M. Yasuda, M. Katsuki, Y. Fujii-Kuriyama, Loss of teratogenic response to 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) in mice lacking the Ah (dioxin) receptor, Genes Cells, 2 (1997) [15] J. Linden, S. Lensu, J. Tuomisto, R. Pohjanvirta, Dioxins, the aryl hydrocarbon receptor and the central regulation of energy balance, Front Neuroendocrinol, 31 (2010) [16] J. Tuomisto, T. Vartiainen, J.T. Tuomisto, Synopsis on Dioxins and PCBs, National Institute for Health and Welfare, Helsinki, Finland, [17] M.M. Ulaszewska, E. Zuccato, E. Davoli, PCDD/Fs and dioxin-like PCBs in human milk and estimation of infants' daily intake: a review, Chemosphere, 83 (2011)

193 193 [18] G.M. Frame, J.W. Cochran, S.S. Bowadt, Complete PCB congener distributions for 17 aroclor mixtures determined by 3 HRGC systems optimized for comprehensive, quantitative, congener-specific analysis, Hrc-J High Res Chrom, 19 (1996) [19] J.W. Newman, J.S. Becker, G. Blondina, R.S. Tjeerdema, Quantitation of Aroclors using congener-specific results, Environ Toxicol Chem, 17 (1998) [20] S. Safe, Polychlorinated biphenyls (PCBs) and polybrominated biphenyls (PBBs): biochemistry, toxicology, and mechanism of action, Crit Rev Toxicol, 13 (1984) [21] J. Castro-Jimenez, C. Gonzalez, Immunoassay-based screening of polychlorinated biphenyls (PCB) in sediments: requirements for a new generation of test kits, J Environ Monit, 13 (2011) [22] S. Jensen, The PCB story, Ambio, 1 (1972) [23] P.A. Bertazzi, L. Riboldi, A. Pesatori, L. Radice, C. Zocchetti, Cancer mortality of capacitor manufacturing workers, Am J Ind Med, 11 (1987) [24] P. Gustavsson, C. Hogstedt, C. Rappe, Short-term mortality and cancer incidence in capacitor manufacturing workers exposed to polychlorinated biphenyls (PCBs), Am J Ind Med, 10 (1986)

194 194 [25] M.M. Prince, A.M. Ruder, M.J. Hein, M.A. Waters, E.A. Whelan, N. Nilsen, E.M. Ward, T.M. Schnorr, P.A. Laber, K.E. Davis-King, Mortality and exposure response among 14,458 electrical capacitor manufacturing workers exposed to polychlorinated biphenyls (PCBs), Environ Health Perspect, 114 (2006) [26] D.P. Brown, Mortality of workers exposed to polychlorinated biphenyls--an update, Arch Environ Health, 42 (1987) [27] R.D. Kimbrough, M.L. Doemland, M.E. LeVois, Mortality in male and female capacitor workers exposed to polychlorinated biphenyls, J Occup Environ Med, 41 (1999) [28] R.D. Kimbrough, M.L. Doemland, J.S. Mandel, A mortality update of male and female capacitor workers exposed to polychlorinated biphenyls, J Occup Environ Med, 45 (2003) [29] G.M. Buck, J.E. Vena, E.F. Schisterman, J. Dmochowski, P. Mendola, L.E. Sever, E. Fitzgerald, P. Kostyniak, H. Greizerstein, J. Olson, Parental consumption of contaminated sport fish from Lake Ontario and predicted fecundability, Epidemiology, 11 (2000) [30] B. Bush, A.H. Bennett, J.T. Snow, Polychlorobiphenyl congeners, p,p'-dde, and sperm function in humans, Arch Environ Contam Toxicol, 15 (1986)

195 195 [31] A. Axmon, L. Rylander, U. Stromberg, L. Hagmar, Miscarriages and stillbirths in women with a high intake of fish contaminated with persistent organochlorine compounds, Int Arch Occup Environ Health, 73 (2000) [32] E.A. Emmett, M. Maroni, J. Jefferys, J. Schmith, B.K. Levin, A. Alvares, Studies of transformer repair workers exposed to PCBs: II. Results of clinical laboratory investigations, Am J Ind Med, 14 (1988) [33] E.A. Emmett, M. Maroni, J.M. Schmith, B.K. Levin, J. Jefferys, Studies of transformer repair workers exposed to PCBs: I. Study design, PCB concentrations, questionnaire, and clinical examination results, Am J Ind Med, 13 (1988) [34] J.L. Jacobson, S.W. Jacobson, Intellectual impairment in children exposed to polychlorinated biphenyls in utero, N Engl J Med, 335 (1996) [35] J.L. Jacobson, S.W. Jacobson, P.M. Schwartz, G.G. Fein, J.K. Dowler, Prenatal exposure to an environmental toxin - a test of the multiple effects model, Dev Psychol, 20 (1984) [36] J.L. Jacobson, S.W. Jacobson, H.E. Humphrey, Effects of exposure to PCBs and related compounds on growth and activity in children, Neurotoxicol Teratol, 12 (1990)

196 196 [37] J.L. Jacobson, S.W. Jacobson, H.E. Humphrey, Effects of in utero exposure to polychlorinated biphenyls and related contaminants on cognitive functioning in young children, J Pediatr, 116 (1990) [38] P. Stewart, J. Reihman, E. Lonky, T. Darvill, J. Pagano, Prenatal PCB exposure and neonatal behavioral assessment scale (NBAS) performance, Neurotoxicol Teratol, 22 (2000) [39] A.B. Smith, J. Schloemer, L.K. Lowry, A.W. Smallwood, R.N. Ligo, S. Tanaka, W. Stringer, M. Jones, R. Hervin, C.J. Glueck, Metabolic and health consequences of occupational exposure to polychlorinated biphenyls, Br J Ind Med, 39 (1982) [40] Y.L. Guo, M.L. Yu, C.C. Hsu, W.J. Rogan, Chloracne, goiter, arthritis, and anemia after polychlorinated biphenyl poisoning: 14-year follow-up of the Taiwan Yucheng cohort, Environ Health Perspect, 107 (1999) [41] A. Fischbein, J.N. Rizzo, S.J. Solomon, M.S. Wolff, Oculodermatological findings in workers with occupational exposure to polychlorinated biphenyls (PCBs), Br J Ind Med, 42 (1985) [42] U. EPA, Method 1668C: Chlorinated biphenyl congeners in water, soil, sediment, biosolids, and tissue by HRGC/HRMS, EPA Method 1668C, U.S. Environmental Protection Agency, Washington, DC, 2010.

197 197 [43] M.C. Bruzzoniti, R. Maina, V. Tumiatti, C. Sarzanini, L. Rivoira, R.M. De Carlo, Fast low-pressure microwave assisted extraction and gas chromatographic determination of polychlorinated biphenyls in soil samples, J Chromatogr A, 1265 (2012) [44] W.J. Crinnion, Polychlorinated biphenyls: persistent pollutants with immunological, neurological, and endocrinological consequences, Altern Med Rev, 16 (2011) [45] M. Van den Berg, L. Birnbaum, A.T. Bosveld, B. Brunstrom, P. Cook, M. Feeley, J.P. Giesy, A. Hanberg, R. Hasegawa, S.W. Kennedy, T. Kubiak, J.C. Larsen, F.X. van Leeuwen, A.K. Liem, C. Nolt, R.E. Peterson, L. Poellinger, S. Safe, D. Schrenk, D. Tillitt, M. Tysklind, M. Younes, F. Waern, T. Zacharewski, Toxic equivalency factors (TEFs) for PCBs, PCDDs, PCDFs for humans and wildlife, Environ Health Perspect, 106 (1998) [46] U. EPA, Polychlorinated Biphenyls (PCBs) by Gas Chromatography, EPA method 8082A, U.S. Environmental Protection Agency, Washington, DC, [47] D. Muir, E. Sverko, Analytical methods for PCBs and organochlorine pesticides in environmental monitoring and surveillance: a critical appraisal, Anal Bioanal Chem, 386 (2006)

198 198 [48] H. Lohninger, K. Varmuza, Selective detection of classes of chemical-compounds by gas-chromatography mass-spectrometry pattern-recognition - polycyclic aromatic-hydrocarbons and alkanes, Anal Chem, 59 (1987) [49] G.R. Magelssen, J.W. Elling, Chromatography pattern recognition of Aroclors using iterative probabilistic neural networks, J Chromatogr A, 775 (1997) [50] A.D. Hewitt, S.A. Shoop, Rapid assessment of trichloroethylene in ground water, Ground Water Monit R, 14 (1994) [51] Z. Ouyang, R.J. Noll, R.G. Cooks, Handheld miniature ion trap mass spectrometers, Anal Chem, 81 (2009) [52] Y. Peng, D.E. Austin, New approaches to miniaturizing ion trap mass analyzers, TrAC Trends in Analytical Chemistry, 30 (2011) [53] T.W. Burgoyne, G.M. Hieftje, R.A. Hites, Design and performance of a plasma-source mass spectrograph, J Am Soc Mass Spectr, 8 (1997) [54] D.L. Swingler, Small magnetic analyzer mass spectrometer with a velocity filter for residual gas analysis, vacuum, 21 (1971)

199 199 [55] R.R.A. Syms, T.J. Tate, M.M. Ahmad, S. Taylor, Fabrication of a microengineered quadrupole electrostatic lens, Electron Lett, 32 (1996) [56] O.J. Orient, A. Chutjian, V. Garkanian, Miniature, high-resolution, quadrupole mass-spectrometer array, Rev Sci Instrum, 68 (1997) [57] D.H. Holkeboer, T.L. Karandy, F.C. Currier, L.C. Frees, R.E. Ellefson, Miniature quadrupole residual gas analyzer for process monitoring at millitorr pressures, J Vac Sci Technol A-Vac Surf Films, 16 (1998) [58] W.B. Brinckerhoff, G.G. Managadze, R.W. McEntire, A.F. Cheng, W.J. Green, Laser time-of-flight mass spectrometry for space, Rev Sci Instrum, 71 (2000) [59] A.J. White, M.G. Blamire, C.A. Corlett, B.W. Griffiths, D.M. Martin, S.B. Spencer, S.J. Mullock, Development of a portable time-of-flight membrane inlet mass spectrometer for environmental analysis, Rev Sci Instrum, 69 (1998) [60] R.E. Kaiser, R.G. Cooks, G.C. Stafford, J.E.P. Syka, P.H. Hemberger, Operation of a quadrupole ion trap mass-spectrometer to achieve high mass charge ratios, Int J Mass Spectrom Ion Process, 106 (1991)

200 200 [61] S.A. Lammert, W.R. Plass, C.V. Thompson, M.B. Wise, Design, optimization and initial performance of a toroidal rf ion trap mass spectrometer, Int J Mass Spectrom, 212 (2001) [62] E.R. Badman, R.C. Johnson, W.R. Plass, R.G. Cooks, A miniature cylindrical quadrupole ion trap: Simulation and experiment, Anal Chem, 70 (1998) [63] O. Kornienko, P.T.A. Reilly, W.B. Whitten, J.M. Ramsey, Electron impact ionization in a microion trap mass spectrometer, Rev Sci Instrum, 70 (1999) [64] E.R. Badman, R.G. Cooks, Special feature: Perspective - Miniature mass analyzers, J Mass Spectrom, 35 (2000) [65] S.A. Lammert, A.A. Rockwood, M. Wang, M.L. Lee, E.D. Lee, S.E. Tolley, J.R. Oliphant, J.L. Jones, R.W. Waite, Miniature toroidal radio frequency ion trap mass analyzer, J Am Soc Mass Spectr, 17 (2006) [66] G.E. Patterson, A.J. Guymon, L.S. Riter, M. Everly, J. Griep- Raming, B.C. Laughlin, Z. Ouyang, R.G. Cooks, Miniature cylindrical ion trap mass spectrometer, Anal Chem, 74 (2002) [67] O. Kornienko, P.T.A. Reilly, W.B. Whitten, J.M. Ramsey, Micro ion trap mass spectrometry, Rapid Commun Mass Sp, 13 (1999)

201 201 [68] M.G. Blain, L.S. Riter, D. Cruz, D.E. Austin, G. Wu, W.R. Plass, R.G. Cooks, Towards the hand-held mass spectrometer: Design considerations, simulation, and fabrication of micrometer-scaled cylindrical ion traps, Int J Mass Spectrom, 236 (2004) [69] E.R. Badman, R.G. Cooks, Cylindrical ion trap array with mass selection by variation in trap dimensions, Anal Chem, 72 (2000) [70] M. Fico, J.D. Maas, S.A. Smith, A.B. Costa, Z. Ouyang, W.J. Chappell, R.G. Cooks, Circular arrays of polymer-based miniature rectilinear ion traps, Analyst, 134 (2009) [71] J.D. Maas, P.I. Hendricks, Z. Ouyang, R.G. Cooks, W.J. Chappell, Miniature monolithic rectilinear ion trap arrays by stereolithography on printed circuit board, J Microelectromech Syst, 19 (2010) [72] D.E. Austin, M. Wang, S.E. Tolley, J.D. Maas, A.R. Hawkins, A.L. Rockwood, H.D. Tolley, E.D. Lee, M.L. Lee, Halo ion trap mass spectrometer, Anal Chem, 79 (2007) [73] C.R. Bowerbank, T.V. Truong, M. Vincent, A.M. Hoffman, M.L. Lee, D.W. Later, Profiling of essential oil VOCs and SVOCs by GC-TMS, Lc Gc N Am, (2010)

202 202 [74] C.R. Bowerbank, T.C. Wirth, D.W. Later, Extraction and identification of volatile organic compounds from water in 70 s using SPME and a hand portable GC-TMS, Lc Gc N Am, (2009) [75] D.W. Later, E.D. Lee, J.L. Oliphant, C.R. Bowerbank, CUSTODION (TM) SPME syringe for rapid sample collection and sample preparation of drinking water for the GC-MS determination of trihalomethanes, Lc Gc N Am, (2008) [76] J.A. Contreras, J.A. Murray, S.E. Tolley, J.L. Oliphant, H.D. Tolley, S.A. Lammert, E.D. Lee, D.W. Later, M.L. Lee, Hand-portable gas chromatograph-toroidal ion trap mass spectrometer (GC-TMS) for detection of hazardous compounds, J Am Soc Mass Spectrom, 19 (2008) [77] U. EPA, Method 1613B: Tetra- through octa-chlorinated dioxins and furans by isotope dilution HRGC/HRMS, EPA method 1613B, US Environmental Protection Agency, Washington, DC, [78] U. EPA, Method 3540C: Soxhlet extraction, EPA method 3540C, U.S. Environmental Protection Agency, Washington, DC, [79] S. Ozcan, Analyses of polychlorinated biphenyls in waters and wastewaters using vortex-assisted liquid-liquid microextraction and gas chromatography-mass spectrometry, J Sep Sci, 34 (2011)

203 203 [80] A. Sjodin, R.S. Jones, C.R. Lapeza, J.F. Focant, E.E. McGahee, D.G. Patterson, Semiautomated high-throughput extraction and cleanup method for the measurement of polybrominated diphenyl ethers, polybrominated biphenyls, and polychlorinated biphenyls in human serum, Anal Chem, 76 (2004) [81] Z. Zhang, S.M. Rhind, Optimized determination of polybrominated diphenyl ethers and polychlorinated biphenyls in sheep serum by solidphase extraction-gas chromatography-mass spectrometry, Talanta, 84 (2011) [82] S.H. Patil, K. Banerjee, S. Dasgupta, D.P. Oulkar, S.B. Patil, M.R. Jadhav, R.H. Savant, P.G. Adsule, M.B. Deshmukh, Multiresidue analysis of 83 pesticides and 12 dioxin-like polychlorinated biphenyls in wine by gas chromatography-time-of-flight mass spectrometry, J Chromatogr A, 1216 (2009) [83] Z. Zhang, E. Ohiozebau, S.M. Rhind, Simultaneous extraction and clean-up of polybrominated diphenyl ethers and polychlorinated biphenyls from sheep liver tissue by selective pressurized liquid extraction and analysis by gas chromatography-mass spectrometry, J Chromatogr A, 1218 (2011) [84] J.L. Martinez Vidal, A. Garrido Frenich, L. Barco Bonilla Mde, R. Romero-Gonzalez, J.A. Padilla Sanchez, Simultaneous extraction of

204 204 polycyclic aromatic hydrocarbons and polychlorinated biphenyls in agricultural soils by pressurized liquid extraction and determination by gas chromatography coupled to tandem mass spectrometry, Anal Bioanal Chem, 395 (2009) [85] G. Rocco, C. Toledo, I. Ahumada, B. Sepulveda, A. Canete, P. Richter, Determination of polychlorinated biphenyls in biosolids using continuous ultrasound-assisted pressurized solvent extraction and gas chromatography-mass spectrometry, J Chromatogr A, 1193 (2008) [86] A. Muller, E. Bjorklund, C. von Holst, On-line clean-up of pressurized liquid extracts for the determination of polychlorinated biphenyls in feedingstuffs and food matrices using gas chromatography-mass spectrometry, J Chromatogr A, 925 (2001) [87] U. EPA, Method 3541: Automated soxhlet extraction, U.S. Environmental Protection Agency, Washington, DC, [88] R. Cariou, J.P. Antignac, P. Marchand, A. Berrebi, D. Zalko, F. Andre, B. Le Bizec, New multiresidue analytical method dedicated to trace level measurement of brominated flame retardants in human biological matrices, J Chromatogr A, 1100 (2005)

205 205 [89] F. Yang, S. Jin, D. Meng, Y. Xu, Solid phase extraction with pyrenebutyric acid-bonded silica for analysis of polychlorinated biphenyls in sewage water by gas chromatography-mass spectrometry, Chemosphere, 81 (2010) [90] C. Thomsen, V.H. Liane, G. Becher, Automated solid-phase extraction for the determination of polybrominated diphenyl ethers and polychlorinated biphenyls in serum--application on archived Norwegian samples from 1977 to 2003, J Chromatogr B Analyt Technol Biomed Life Sci, 846 (2007) [91] U. EPA, Method 3545A: Pressurized fluid extraction (PFE), U.S. Environmental Protection Agency, Washington, DC, [92] A. Sarafraz-Yazdi, A. Amiri, Liquid-phase microextraction, Trac- Trend Anal Chem, 29 (2010) [93] X. Liu, A. Zhao, A. Zhang, H. Liu, W. Xiao, C. Wang, X. Wang, Dispersive liquid-liquid microextraction and gas chromatography-mass spectrometry determination of polychlorinated biphenyls and polybrominated diphenyl ethers in milk, J Sep Sci, 34 (2011) 7. [94] G. Li, L. Zhang, Z. Zhang, Determination of polychlorinated biphenyls in water using dynamic hollow fiber liquid-phase microextraction and gas chromatography-mass spectrometry, J Chromatogr A, 1204 (2008)

206 206 [95] C. Basheer, H.K. Lee, J.P. Obbard, Application of liquid-phase microextraction and gas chromatography-mass spectrometry for the determination of polychlorinated biphenyls in blood plasma, J Chromatogr A, 1022 (2004) [96] S. Ozcan, A. Tor, M.E. Aydin, Determination of selected polychlorinated biphenyls in water samples by ultrasound-assisted emulsification-microextraction and gas chromatography-massselective detection, Anal Chim Acta, 647 (2009) [97] M. Xie, Z.Y. Yang, L.J. Bao, E.Y. Zeng, Equilibrium and kinetic solid-phase microextraction determination of the partition coefficients between polychlorinated biphenyl congeners and dissolved humic acid, J Chromatogr A, 1216 (2009) [98] Y.H. Wang, Y.Q. Li, J. Zhang, S.F. Xu, S.G. Yang, C. Sun, A novel fluorinated polyaniline-based solid-phase microextraction coupled with gas chromatography for quantitative determination of polychlorinated biphenyls in water samples, Anal Chim Acta, 646 (2009) [99] R. Lopez, F. Goni, A. Etxandia, E. Millan, Determination of organochlorine pesticides and polychlorinated biphenyls in human serum using headspace solid-phase microextraction and gas chromatography-electron capture detection, J Chromatogr B, 846 (2007)

207 207 [100] K.J. Chia, T.Y. Lee, S.D. Huang, Simple device for the solidphase microextraction screening of polychlorodibenzo-p-dioxins and polychlorodibenzofurans in heavily contaminated soil samples, Anal Chim Acta, 527 (2004) [101] F. Rezaei, A. Bidari, A.P. Birjandi, M.R. Milani Hosseini, Y. Assadi, Development of a dispersive liquid-liquid microextraction method for the determination of polychlorinated biphenyls in water, J Hazard Mater, 158 (2008) [102] X. Liu, J. Li, Z. Zhao, W. Zhang, K. Lin, C. Huang, X. Wang, Solid-phase extraction combined with dispersive liquid-liquid microextraction for the determination for polybrominated diphenyl ethers in different environmental matrices, J Chromatogr A, 1216 (2009) [103] N. Fattahi, S. Samadi, Y. Assadi, M.R. Hosseini, Solid-phase extraction combined with dispersive liquid-liquid microextraction-ultra preconcentration of chlorophenols in aqueous samples, J Chromatogr A, 1169 (2007) [104] S. King, J.S. Meyer, A.R.J. Andrews, Screening method for polycyclic aromatic hydrocarbons in soil using hollow fiber membrane solvent microextraction, J Chromatogr A, 982 (2002)

208 208 [105] M.D. Luque de Castro, F. Priego-Capote, Ultrasound-assisted preparation of liquid samples, Talanta, 72 (2007) [106] C.L. Arthur, J. Pawliszyn, Solid-phase microextraction with thermal-desorption using fused-silica optical fibers, Anal Chem, 62 (1990) [107] G. Vas, K. Vekey, Solid-phase microextraction: a powerful sample preparation tool prior to mass spectrometric analysis, J Mass Spectrom, 39 (2004) [108] S. Risticevic, H. Lord, T. Gorecki, C.L. Arthur, J. Pawliszyn, Protocol for solid-phase microextraction method development, Nat Protoc, 5 (2010) [109] H. Bagheri, E. Babanezhad, F. Khalilian, An interior needle electropolymerized pyrrole-based coating for headspace solid-phase dynamic extraction, Anal Chim Acta, 634 (2009) [110] W. Guan, Y.J. Wang, F. Xu, Y.F. Guan, Poly(phthalazine ether sulfone ketone) as novel stationary phase for stir bar sorptive extraction of organochlorine compounds and organophosphorus pesticides, J Chromatogr A, 1177 (2008) [111] W.N. Guan, F. Xu, W.M. Liu, J.H. Zhao, Y.F. Guan, A new poly (phthalazine ether sulfone ketone)-coated fiber for solid-phase

209 209 microextraction to determine nitroaromatic explosives in aqueous samples, J Chromatogr A, 1147 (2007) [112] F.C.M. Portugal, M.L. Pinto, J.M.F. Nogueira, Optimization of Polyurethane Foams for Enhanced Stir Bar Sorptive Extraction of Triazinic Herbicides in Water Matrices, Talanta, 77 (2008) [113] S. Risticevic, V.H. Niri, D. Vuckovic, J. Pawliszyn, Recent developments in solid phase microextraction, Anal Bioanal Chem, 393 (2009) [114] H. Lord, J. Pawliszyn, Evolution of solid-phase microextraction technology, J Chromatogr A, 885 (2000) [115] H. Kataoka, H.L. Lord, J. Pawliszyn, Applications of solid-phase microextraction in food analysis, J Chromatogr A, 880 (2000) [116] L. Kudlejova, S. Risticevic, D. Vuckovic, Solid phase microextraction method development, Handbook of Solid Phase Microextraction: SPME, (2007) [117] S.B. Hawthorne, C.B. Grabanski, D.J. Miller, Solid-phasemicroextraction measurement of 62 polychlorinated biphenyl congeners in milliliter sediment pore water samples and determination of K-DOC values, Anal Chem, 81 (2009) [118] R. Montes, M. Ramil, I. Rodriguez, E. Rubi, R. Cela, Rapid screening of polychlorinated biphenyls in sediments using non-

210 210 equilibrium solid-phase microextraction and fast gas chromatography with electron-capture detection, J Chromatogr A, 1124 (2006) [119] M.R. Criado, I.R. Pereiro, R.C. Torrijos, Determination of polychlorinated biphenyls in ash using dimethylsulfoxide microwave assisted extraction followed by solid-phase microextraction, Talanta, 63 (2004) [120] D. Poli, A. Caglieri, M. Goldoni, A.F. Castoldi, T. Coccini, E. Roda, A. Vitalone, S. Ceccatelli, A. Mutti, Single step determination of PCB 126 and 153 in rat tissues by using solid phase microextraction/gas chromatography-mass spectrometry: Comparison with solid phase extraction and liquid/liquid extraction, J Chromatogr B, 877 (2009) [121] C. Wu, J. Schaum, Exposure assessment of trichloroethylene, Environ Health Persp, 108 (2000) [122] A.D. Delinsky, J.V. Bruckner, M.G. Bartlett, A review of analytical methods for the determination of trichloroethylene and its major metabolites chloral hydrate, trichloroacetic acid and dichloroacetic acid, Biomed Chromatogr, 19 (2005) [123] B. Bakke, P.A. Stewart, M.A. Waters, Uses of and exposure to trichloroethylene in US industry: A systematic literature review, J Occup Environ Hyg, 4 (2007)

211 211 [124] H.A. Barton, C.D. Flemming, J.C. Lipscomb, Evaluating human variability in chemical risk assessment: Hazard identification and doseresponse assessment for noncancer oral toxicity of trichloroethylene, Toxicology, 111 (1996) [125] C.S. Scott, J. Jinot, Trichloroethylene and cancer: systematic and quantitative review of epidemiologic evidence for identifying hazards, Int J Env Res Pub He, 8 (2011) 4238-Nii_4163. [126] D. Wartenberg, D. Reyner, C.S. Scott, Trichloroethylene and cancer: Epidemiologic evidence, Environ Health Persp, 108 (2000) [127] ATSDR, Toxicological profile for trichloroethylene, 11 (1997) 124. [128] R.M. Ray, M.M. Mumtaz, Development of a priority list of chemical mixtures occurring at 1188 hazardous waste sites, using the HazDat database, Food Chem Toxicol, 34 (1996) [129] U. EPA, Tricholoethylene health risk assessment: Synthesis and characterization. External Review Draft; EPA/600/P-01/022A; U.S. Environmental Protection Agency: Washington, DC, [130] P.I. Beamer, C.E. Luik, L. Abrell, S. Campos, M.E. Martinez, A.E. Saez, Concentration of trichloroethylene in breast milk and household

212 212 water from Nogales, Arizona, Environ Sci Tech, 46 (2012) [131] P.G. Forkert, L. Lash, R. Tardif, N. Tanphaichitr, C. Vandevoort, M. Moussa, Identification of trichloroethylene and its metabolites in human seminal fluid of workers exposed to trichloroethylene, Drug Metab Dispos, 31 (2003) [132] D.J. Munch, D.P. Hautman, USEPA Method 551.1: Determination of chlorination disinfection byproducts, chlorinated solvents, and halogenated pesticides/herbicides in drinking water by liquid-liquid extraction and gas chromatography with electroncapture detection, US Environmental Protection Agency: Cincinatti, OH, [133] B. Dehon, L. Humbert, L. Devisme, M. Stievenart, D. Mathieu, N. Houdret, M. Lhermitte, Tetrachloroethylene and trichloroethylene fatality: Case report and simple headspace SPME-capillary gas chromatographic determination in tissues, J Anal Toxicol, 24 (2000) [134] M.I. Cervera, J. Beltran, F.J. Lopez, F. Hernandez, Determination of volatile organic compounds in water by head space-solid-phase microextraction gas chromatography coupled to tandem mass spectrometry with triple quadrupole analyzer, Anal Chim Acta, 704 (2011)

213 213 [135] N. Jakubowska, B. Henkelmann, K.W. Schramm, J. Namiesnik, Optimization of a novel procedure for determination of VOCs in water and human urine samples based on SBSE coupled with TD-GC-HRMS, J Chromatogr Sci, 47 (2009) [136] Y.Z. Liu, S. Muralidhara, J.V. Bruckner, M.G. Bartlett, Trace level determination of trichloroethylene in biological samples by headspace solid-phase microextraction gas chromatography/negative chemical ionization mass spectrometry, Rapid Commun Mass Sp, 22 (2008) [137] D. Poli, P. Manini, R. Andreoli, I. Franchini, A. Mutti, Determination of dichloromethane, trichloroethylene and perchloroethylene in urine samples by headspace solid phase microextraction gas chromatography-mass spectrometry, J Chromatogr B, 820 (2005) [138] N. Xu, S. Vandegrift, G.W. Sewell, Determination of chloroethenes in environmental biological samples using gas chromatography coupled with solid phase micro extraction, Chromatographia, 42 (1996) [139] J. Ai, Headspace solid phase microextraction. Dynamics and quantitative analysis before reaching a partition equilibrium, Anal Chem, 69 (1997)

214 214 [140] U.E.P. Agency, EPA Method 8260C: Volatile organic compounds by gas chromatography/mass spectrometry (GC/MS), [141] W.T. Chang, D.L. Lin, R.H. Liu, Isotopic analogs as internal standards for quantitative analyses by GUMS - evaluation of crosscontribution to ions designated for the analyte and the isotopic internal standard, Forensic Sci Int, 121 (2001) [142] D.L. Stalling, T.R. Schwartz, W.J. Dunn, S. Wold, Classification of polychlorinated biphenyl residues - isomers Vs homolog concentrations in modeling Aroclors and polychlorinated biphenyl residues, Anal Chem, 59 (1987) [143] W.J. Dunn, D.L. Stalling, T.R. Schwartz, J.W. Hogan, J.D. Petty, E. Johansson, S. Wold, Pattern-recognition for classification and determination of polychlorinated-biphenyls in environmental-samples, Anal Chem, 56 (1984) [144] C.Y. Ma, C.K. Bayne, Differentiation of Aroclors using linear discrimination for environmental-samples analyzed by electron-capture negative-ion chemical ionization mass-spectrometry, Anal Chem, 65 (1993) [145] P.D. Harrington, N.E. Vieira, P. Chen, J. Espinoza, J.K. Nien, R. Romero, A.L. Yergey, Proteomic analysis of amniotic fluids using analysis of variance-principal component analysis and fuzzy rule-

215 215 building expert systems applied to matrix-assisted laser desorption/ionization mass spectrometry, Chemometr Intell Lab, 82 (2006) [146] P.B. Harrington, Fuzzy multivariate rule-building expert systems - minimal neural networks, J Chemometr, 5 (1991) [147] X.B. Sun, P. Chen, S.L. Cook, G.P. Jackson, J.M. Harnly, P.B. Harrington, Classification of cultivation locations of panax quinquefolius L samples using high performance liquid chromatography - electrospray ionization mass spectrometry and chemometric analysis, Anal Chem, 84 (2012) [148] P.D. Harrington, J. Kister, J. Artaud, N. Dupuy, Automated principal component-based orthogonal signal correction applied to fused near infrared-mid-infrared spectra of french olive oils, Anal Chem, 81 (2009) [149] Z.F. Wang, P. Chen, L.L. Yu, P.D. Harrington, Authentication of organically and conventionally grown basils by gas chromatography/mass spectrometry chemical profiles, Anal Chem, 85 (2013) [150] B.W. Wabuyele, P.D. Harrington, Fuzzy optimal associative memory for background prediction of near-infrared spectra, Appl Spectrosc, 50 (1996)

216 216 [151] B.B. Li, J. Morris, E.B. Martin, Model selection for partial least squares regression, Chemometr Intell Lab, 64 (2002) [152] P. Geladi, B.R. Kowalski, Partial least-squares regression - a tutorial, Anal Chim Acta, 185 (1986) [153] F.A. Inon, J.M. Garrigues, S. Garrigues, A. Molina, M. de la Guardia, Selection of calibration set samples in determination of olive oil acidity by partial least squares-attenuated total reflectance-fourier transform infrared spectroscopy, Anal Chim Acta, 489 (2003) [154] S. Wold, A. Ruhe, H. Wold, W.J. Dunn, The collinearity problem in linear-regression - the partial least-squares (Pls) approach to generalized inverses, Siam J Sci Stat Comp, 5 (1984) [155] M. Sjostrom, S. Wold, W. Lindberg, J.A. Persson, H. Martens, A multivariate calibration-problem in analytical-chemistry solved by partial least-squares models in latent-variables, Anal Chim Acta, 150 (1983) [156] L. An, J. Deng, L. Zhou, H. Li, F. Chen, H. Wang, Y.T. Liu, Simultaneous spectrophotometric determination of trace amount of malachite green and crystal violet in water after cloud point extraction using partial least squares regression, J Hazard Mater, 175 (2010)

217 217 [157] D.M. Haaland, E.V. Thomas, Partial least-squares methods for spectral analyses.1. Relation to other quantitative calibration methods and the extraction of qualitative information, Anal Chem, 60 (1988) [158] O.E. Denoord, The influence of data preprocessing on the robustness and parsimony of multivariate calibration models, Chemometr Intell Lab, 23 (1994) [159] Z.F. Xu, X.B. Sun, P.D. Harrington, Baseline correction method using an orthogonal basis for gas chromatography/mass spectrometry data, Anal Chem, 83 (2011) [160] L.R. Crawford, J.D. Morrison, Computer methods in analytical mass spectrometry empirical identification of molecular class, Anal Chem, 40 (1968) [161] Y. Lu, P. Chen, P.B. Harrington, Comparison of differential mobility spectrometry and mass spectrometry for gas chromatographic detection of ignitable liquids from fire debris using projected difference resolution, Anal Bioanal Chem, 394 (2009) [162] P. Chen, Y. Lu, P.B. Harrington, Application of linear and nonlinear discrete wavelet transforms to MALDI-MS measurements of bacteria for classification, Anal Chem, 80 (2008)

218 218 [163] G.M. Frame, R.E. Wagner, J.C. Carnahan, J.F. Brown, R.J. May, L.A. Smullen, D.L. Bedard, Comprehensive, quantitative, congenerspecific analyses of eight aroclors and complete PCB congener assignments on DB-1 capillary GC columns, Chemosphere, 33 (1996) [164] M. Llompart, K. Li, M. Fingas, Headspace solid-phase microextraction for the determination of polychlorinated biphenyls in soils and sediments, J Microcolumn Sep, 11 (1999) [165] A. Derouiche, M.R. Driss, J.P. Morizur, M.H. Taphanel, Simultaneous analysis of polychlorinated biphenyls and organochlorine pesticides in water by headspace solid-phase microextraction with gas chromatography-tandem mass spectrometry, J Chromatogr A, 1138 (2007) [166] M. Zhang, P.d.B. Harrington, Automated pipeline for classifying Aroclors in soil by gas chromatography/mass spectrometry using modulo compressed two-way data objects, Talanta, 117 (2013) [167] G.L. Long, J.D. Winefordner, Limit of detection a closer look at the IUPAC definition, Anal Chem, 55 (1983) 712A-724A. [168] M. Zhang, P.d.B. Harrington, N.A. Kruse, J.R. Bowman, S.A. Lammert, E.D. Lee, G.P. Jackson, Development of an expedited field

219 219 study method for PCBs in sediments and soils using portable GC/MS, 61st ASMS Conference on Mass Spectrometry and Allied TopicsMinneapolis, MN, [169] M. Zhang, G.P. Jackson, N.A. Kruse, J.R. Bowman, P.d.B. Harrington, Determination of Aroclor 1260 in soil samples by gas chromatography/mass spectrometry with solid phase microextraction, J Sep Sci, 37 (2014) [170] R. Boque, F.X. Rius, Multivariate detection limits estimators, Chemometr Intell Lab, 32 (1996) [171] T. Kotiaho, On-site environmental and in situ process analysis by mass spectrometry, J Mass Spectrom, 31 (1996) [172] Mengliang Zhang, Glen P. Jackson, Natalie A. Kruse, J.R. Bowman, P.d.B. Harrington, Determination of Aroclor 1260 in soil samples by gas chromatography/mass spectrometry with solid phase microextraction, J Sep Sci, 37 (2014) [173] A.M. Almeida, M.M. Castel-Branco, A.C. Falcao, Linear regression for calibration lines revisited: weighting schemes for bioanalytical methods, J Chromatogr B, 774 (2002) [174] A. Garrido Frenich, R. Romero-Gonzalez, J.L. Martinez Vidal, R. Martinez Ocana, P. Baquero Feria, Comparison of solid phase microextraction and hollow fiber liquid phase microextraction for the

220 220 determination of pesticides in aqueous samples by gas chromatography triple quadrupole tandem mass spectrometry, Anal Bioanal Chem, 399 (2011) [175] D.L. Massart, B.G.M. Vandeginste, D. S.N.;, M. Y.;, L. Kaufman, Chemometrics: a textbook, Elsevier, Amsterdam, [176] M. Rezaee, Y. Assadi, M.R.M. Hosseinia, E. Aghaee, F. Ahmadi, S. Berijani, Determination of organic compounds in water using dispersive liquid-liquid microextraction, J Chromatogr A, 1116 (2006) 1-9. [177] John S. Zogorski, Janet M. Carter, Tamara Ivahnenko, Wayne W. Lapham, Michael J. Moran, Barbara L. Rowe, Paul J. Squillace, P.L. Toccalino, Volatile Organic Compounds in the Nation's Ground Water and Drinking-Water Supply Wells, in: U.S.D.o.t.I.a.U.S.G. Survey (Ed.)U.S. Geological Survey, Reston, Virginia, 2006, pp

221 221 APPENDIX A: PUBLICATIONS [1] Zhang, M.; Harrington, P. B., Determination of Trichloroethylene in Water using Liquid-liquid Microextraction Assisted Solid Phase Microextraction and Classical Least Squares Resolution of Overlapping Isotopic Peak Clusters. Analytical Chemistry 2014 (Under review). [2] Wang, Z.; Zhang, M.; Harrington, P. B., A Comparison of Three Algorithms for the Baseline Correction of Hyphenated Data Objects. Analytical Chemistry 2014, 86, DOI: /ac501658k [3] Zhang, M.; Harrington, P. B., Simultaneous Quantification of Aroclor Mixtures in Soil Samples by Gas Chromatography/Mass Spectrometry with Solid Phase Microextraction using Partial Least-Squares Regression. Chemosphere 2015, 118, DOI: /j.chemosphere [4] Zhang, M.; Jackson, G. P.; Kruse, N. A.; Bowman, J. R.; Harrington, P. B., Determination of Aroclor 1260 in Soil Samples by GC/MS with Solid Phase Microextraction. Journal of Separation Science 2014, 00, 1-6. DOI: /jssc

222 222 [5] Zhang, M.; Harrington, P. B., Automated pipeline for classifying Aroclors in soil by gas chromatography/mass spectrometry using modulo compressed two-way data objects. Talanta 2013, 117, Some contents in this dissertation are reprinted with permission from Zhang, M.; Harrington, P. B., Simultaneous Quantification of Aroclor Mixtures in Soil Samples by Gas Chromatography/Mass Spectrometry with Solid Phase Microextraction using Partial Least-Squares Regression. Chemosphere 2015, 118, , Copyright (2014), with permission from Elsevier. License number: Zhang, M.; Jackson, G. P.; Kruse, N. A.; Bowman, J. R.; Harrington, P. B., Determination of Aroclor 1260 in Soil Samples by GC/MS with Solid Phase Microextraction. Journal of Separation Science 2014, 37(19), , with permission from 2014 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. License number: Zhang, M.; Harrington, P. B., Automated pipeline for classifying Aroclors in soil by gas chromatography/mass spectrometry using modulo compressed two-way data objects. Talanta 2013, 117, , Copyright (2014), with permission from Elsevier. License number:

223 223 APPENDIX B: PRESENTATIONS [1] Zhang, M.; Harrington, P. B., Field Analysis of Trichloroethylene in Water using Liquid-liquid Microextraction Assisted Solid Phase Microextraction with Portable Gas Chromatography/Mass Spectrometry. FACSS SciX 2014 Conference. Reno, MD. Oct (Oral) [2] Zhang, M.; Harrington, P. B., Determination of Aroclor 1254 and 1260 in soil samples by headspace solid phase microextraction GC/MS using partial least-squares regression. 62nd ASMS Conference on Mass Spectrometry and Allied Topics. Baltimore, MD. June [3] Wang, Z.; Zhang, M.; Harrington, P. B., Reconstruction of Mass Spectra Using Fuzzy Optimal Associative Memories (FOAMs). 62nd ASMS Conference on Mass Spectrometry and Allied Topics. Baltimore, MD. June [4] Zhang, M.; Harrington, P. B.; Kruse, N. A.; Bowman, J. R.; Lammert, S. A.; Lee, E. D.; Jackson, G. P., Development of an expedited field study method for PCBs in sediments and soils using portable GC/MS. 61st ASMS Conference on Mass Spectrometry and Allied Topics. Minneapolis, MN. June 2013.

224 224 [5] Zhang, M.; Jackson, G. P., Expedited field survey and sampling techniques for polychlorinated biphenyl congeners in soil. 9th Annual Ohio Valley Mass Spectrometry Symposium. Columbus, OH. April (Oral)

225 !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! Thesis and Dissertation Services!