Biometeorological modelling and forecasting of monthly ambulance demand for Hong Kong

Transcription

1 Hong Kong Journal of Emergency Medicine Biometeorological modelling and forecasting of monthly ambulance demand for Hong Kong HT Wong, PC Lai, S Chen Introduction: Given the aging population in Hong Kong and the ever rising demand for emergency ambulance services, this study aimed to examine the effects of seasonality and weather on the demand for emergency ambulance services in Hong Kong. The feasibility of using time series models and selected weather factors to forecast average daily ambulance demand over a month was also assessed. Methods: Monthly statistics for ambulance demand from 1998 to 2007 were obtained for analysing the effects of seasonality and weather on the demand for emergency ambulance services in Hong Kong. The effectiveness of weather factors in forecasting ambulance demand was also examined by comparing the performance of the autoregressive integrated moving average (ARIMA) model against other commonly used models. Results: The lowest temperatures during cooler months were found to be negatively associated with average daily ambulance demand (adj-r 2 =0.38), while the average amount of cloud cover and highest temperatures were found to be positively associated with average daily ambulance demand during hotter months (adj-r 2 =0.34). When the analysis was stratified spatially by ambulance command units, Hong Kong Island had the highest adj-r 2 during cool and hot months, reported at 0.55 and 0.46 respectively. With the inclusion of average temperature, the ARIMA models outperformed other models for both short- and long-term predictions. Conclusions: Our findings suggest that weather factors, especially temperature, are significantly related to and useful for predicting ambulance demand. (Hong Kong j.emerg.med. 2017;24:3-11) Keywords: Emergency medical services, forecasting, statistical model, temperature, weather Background It is widely believed that weather conditions have a significant impact on human health. Researchers have reported that death rates increase during extreme hot Correspondence to: Chen Si, PhD The University of Hong Kong, Department of Geography, Hong Kong csissi@connect.hku.hk Lai Poh Chin, PhD State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Science and Natural Resource Research, Chinese Academy of Sciences, Beijing, People's Republic of China; and National Taiwan University, Department of Geography, Taiwan Wong Ho Ting, PhD and cold weather conditions. 1-6 More people die in the winter than the summer months and higher than normal mortality rates have been attributed primarily to cardiovascular and respiratory illnesses. 7 Rooney et al 8 argued that elevated temperatures could exacerbate chronic illnesses, as evidenced by the August 2003 heat wave in France. 9 It was reported that about 15,000 more deaths occurred between August 1 and 20, 2003, which represented a 60% increase in mortality from all causes, compared to the same period in Furthermore, the percentage increase in mortality for people aged and those 75 years and older were 30% and 70% respectively. From 1976 to 1995, above average mortality due to heat waves was observed throughout Europe 10 and the situation continued through to In fact, the heat-mortality relationship was evident among cities in different climates and their mathematical relationships were rather similar. 12 Gosling et al 13 presented a detailed

2 4 Hong Kong j. emerg. med. Vol. 24(1) Jan 2017 review of the relationships between temperature and mortality, as examined through epidemiological and synoptic climatological methods, with an outlook on the potential impact of heat-related mortality from climate change. It is not uncommon for temperatures in Hong Kong to rapidly plunge below 10 o C or rise above 32 o C in a relatively short time interval. Abrupt temperature changes are related to Hong Kong's geographic location in southern China (Figure 1). Hong Kong's hilly terrain and its many outlying islands are subject not only to monsoon-influenced humid subtropical climate with moderately warm temperatures for nearly half of the year but also to cold northerly winds from China. 14 Yan attempted to investigate the seasonality of mortality and weather-mortality relationships in Hong Kong from 1980 to The results showed a negative relationship between minimum temperature and mortality but a positive association between the amounts of cloud cover and mortality. Although weather-mortality association was not found among people aged 0-24 and 25-44, a strong association was discovered among elderly people aged 65 years and older. The vulnerability of the elderly to extreme temperatures, especially net effective temperature, was further confirmed by Leung et al. 2 The above observations of weather-related mortality suggest possible relationships between weather and ambulance demand. If the relationship between Figure 1. A location map of Hong Kong. weather and ambulance demand could be established, it would be possible to build a forecast system for ambulance demand to facilitate forward planning and improve deployment of emergency ambulances. Noticeable progress has been made by researchers in the development of forecast systems for ambulance demand over the past decades. The time series models for predicting ambulance demand were built on "moving average," "means with moving average smoothing," and "autoregressive integrated moving average". 15 The results showed that "means with moving average smoothing", explaining about 54.3% of variance, yielded the best model. In a study to predict the volume of ambulance runs using weather factors and a backpropagation neural network, Liao et al. 16 showed that their resultant models successfully predicted the volume of ambulance runs to within a 6% error rate. The above models are not without drawbacks. For example, the multiple regression analysis and the backpropagation neural network method are restricted by the effects of selected predictors. However, time series analysis which caters for both predictor and temporal effects seems to have greater modelling and predictive powers. The potential of time series modelling was demonstrated by the ARIMA model developed to forecast disease incidence of the Ross River Virus in Cairns, 17 which showed that relative humidity and rainfall were significantly associated with disease incidence. Time series modelling was also used to assess seasonal association of monthly suicide data and weather conditions 18 and to evaluate the accuracy of five different methods of forecasting the incidence of malaria. 19 It was also observed that time series modelling was used widely because it could be carried out by adequately trained non-specialists. 20 This research aimed to examine weather factors of Hong Kong and their seasonality effects on the demand for emergency ambulance services. The spatial variability of ambulance demand would also be assessed. Upon confirming weather factors that would be significantly related to ambulance demand in Hong Kong, the feasibility of time series modelling to forecast average daily ambulance demand based on selected

3 Wong et al./modelling of monthly Hong Kong ambulance demand 5 weather factors would be evaluated. The performance of the ARIMA model would be compared against other commonly used linear regression models. Given the aging population of Hong Kong and the ever rising demand for emergency ambulance services, this study would have significant public health implications in terms of early mobilisation and better preparedness for rapid responses and more effective deployment of ambulances. The forecasting methodology should have direct relevance to Asian cities with rapidly aging population and limited ambulance resources. Methods Ambulance demand statistics (A) The study made use of monthly ambulance demand statistics collected by the Fire Services Department (FSD) from January 1998 to December Figure 2 is a time series plot of the data, showing a linear trend of increasing ambulance demand within the study period. Four potential outliers were observed in 2003 and The first three reflected interference from the severe acute respiratory syndrome (SARS) outbreak, but the fourth dip in 2006 could not be accounted for. It was also reported that for many years, February had the lowest number of ambulance calls (personal communication with the FSD in February 2009). Meteorological statistics Monthly readings of the lowest daily temperature, highest daily temperature, average relative humidity, average amount of cloud cover, total rainfall, and average wind speed for the same time period were obtained from the website of the Hong Kong Observatory (HKO). Figure 3 is a time (B) series plot of monthly relative humidity, as well as the highest and lowest temperatures from 1998 to Yearly periodicity is evident from the data, with peaks in the summer and dips in the winter. Data pre-processing The presence of outliers and the linear trend of increasing ambulance demand as revealed in Figure 2 prompted the need for data pre-processing or (B) normalisation. The 2003 data were excluded from the analysis of weather-ambulance demand relationships due to the SARS epidemic. Data points in February 2006 were not excluded as there was no ground to regard them as outliers. These data were transformed to show average daily ambulance demand and normalisation was attempted to neutralise temporal variations in call volume. Normalisation essentially removes the effects of population growth to enable better visualisation and straightforward comparison of seasonal fluctuations and recurrent demands for ambulance services. However, the development of a forecast system must account for effects of population growth and this process could make use of the original data that have not been normalised. Figure 2. A time series plot of monthly ambulance demand, Figure 3. Time series plots of monthly relative humidity against highest and lowest temperatures,

4 6 Hong Kong j. emerg. med. Vol. 24(1) Jan 2017 Seasonality and weather relationships with ambulance demand Seasonal patterns of the data series were examined by considering three components: random, seasonal, and non-seasonal. A harmonic analysis was applied to assess the percentage of total variance attributable to each component a higher percentage would signify more significance. 21 Having confirmed the presence of seasonality effects, the relationships between selected weather factors (temperature, relative humidity, amount of cloud cover, rainfall, wind, etc.) and ambulance demand were then analysed through regression modelling. The dataset in the regression analysis was first split into two subgroups by time periods corresponding to cool and hot months. Months with the average daily lowest temperature of 24 o C or below (i.e., not above 24 o C) constituted the cool months. Conversely, months with the average daily highest temperature 26 o C or above (i.e., not below 26 o C) were the hot months. The two subgroups were analysed separately to determine seasonal differences. Moreover, the geographic variation of ambulance demand was also examined for three main regions: Hong Kong Island, Kowloon, and the New Territories (see Figure 1). Development of forecast models for ambulance demand Selected weather factors confirmed to have an impact on ambulance demand were employed as independent variables to model and predict ambulance demand. Three types of forecast models were compared in this study: (i) simple linear regression; (ii) seasonal adjustment based on the last three observations; and (iii) ARIMA model. The simple linear regression method requires no introduction, but the seasonal adjustment method contains a term to account for seasonal effects and another term to correct for trend effects in the three preceding months. 19 The equation for seasonal adjustment is written as follows: Eqn. (1) where the seasonal average A t+m is the historical average of observed values in each calendar month and the correctional term is the mean deviation from the seasonal average of observed values in the past three months. The ARIMA model involves an autoregressive process that relates an observation to past observations and a moving average process to relate an observation to past error terms. 22 Mathematically, the autoregressive process AR(p) is expressed as: X t =ε t +φ 1 X t-1 +φ 2 X t φ p X t-p Eqn. (2) where ε t is the error term at time t and φ p is the autoregressive coefficient of the observation X t-p ; and the moving average process MA(q) is written as: X t =ε t +θ 1 ε t-1 +θ 2 ε t θ q ε t-q Eqn. (3) where ε t is the error term at time t and θ q is the moving average coefficient of the error term ε t-q. When the two processes are combined into an autoregressive integrated moving average model, the ARIMA(p,d,q) expression becomes: Y t =φ 1 Y t-1 +φ 2 Y t φ p Y t-p +ε t +θ 1 ε t-1 +θ 2 ε t θ q ε t-q Eqn. (4) where Y t is the d-times differenced X t. The above models were developed and tested using the SPSS Forecasting module. 23 To assess the prediction accuracy of these models, the multi-year dataset was subdivided into two subsets: (i) historical ( ) and (ii) validation (2007). The historical data were used for model development, while the validation data were used to test the prediction accuracy of each forecast model. Each individual forecast model would yield ambulance demand forecasts for one to twelve months. The average absolute percentage error (AAPE) was computed as an objective measure of prediction accuracy, with a smaller value indicating better performance. 16,24 Results Seasonality of ambulance demand Our results showed the presence of a significant biseasonal pattern in the average daily ambulance demand (p<0.001), with 22% (p<0.001) of total variance attributable to seasonal harmonics with 2 cycles (Table 1). This result is consistent with our expectation that the cooler and hotter months would show greater ambulance demand. Moreover, the existence of seasonality in ambulance demand also suggests the suitability of applying the ARIMA model for making demand forecasts. 23

5 Wong et al./modelling of monthly Hong Kong ambulance demand 7 Weather effects (1) Overall situation for Hong Kong A multiple regression model must satisfy three major assumptions: (1) the residuals are normally distributed about a zero mean; (2) the residuals have constant variance; and (3) the sample size is 15 times larger than the number of predictors. The study found that the second assumption of constant variance was not met because the plot of residuals against the "lowest temperature" predictor exhibited a funnel-shaped distribution, indicating an inverse relationship. Thus, the weighted multiple linear regression method which makes no specific assumption about variance was selected for data analysis in place of the ordinary multiple linear regression. As the situation applied to all weather factors examined in this study, the weighted multiple linear regression analysis was adopted throughout the paper. The backward variable selection method was applied in the regression analysis and separately on the two datasets of defined time periods to identify potential predictors. Predictors deemed to be significantly associated with ambulance demand were identified and retained. The following weather variables were considered potential predictors of average daily ambulance demand for both cool and hot months: the lowest and highest temperatures, average relative humidity, average amount of cloud cover, total rainfall, and average wind speed. Results of the weighted multiple linear regressions are shown in Table 2. Weather factors, including lowest temperature, average wind speed, and average relative humidity were found negatively associated with average daily ambulance demand (adj-r 2 = 0.38, p<0.001) over the cool months. In contrast, weather factors including highest temperature and average amount of cloud cover were found positively associated with average daily ambulance demand in the hot months (adj-r 2 = 0.34, p<0.001). (2) Spatial differences in ambulance demand The 2006 Population By-census conducted by the Hong Kong Census and Statistics Department showed that the demographic constructs of people living on Hong Kong Island, in Kowloon and the New Territories were quite different. 25 The proportion of elderly residents aged 65 years and older was lowest in the New Territories (10%) compared to those of Hong Kong Island and Kowloon (13.7% and 15.9% respectively). The Population By-census also showed that people in the New Territories had lower income with a larger proportion living in public housing. Because of these demographic differences, the datasets were further subdivided into smaller subsets of ambulance command units to examine possible spatial effects of weather factors. Table 3 summarises regression Table 1. Results of harmonic analysis on ambulance demand Component of variances All seasonal harmonics % 1 cycle 17 1% 2 cycles % 3 cycles 0 0% 4 cycles 61 2% 5 cycles 9 0% 6 cycles 0 0% Non-seasonal harmonics % Random variations % Total variance % Table 2. Results of regression analysis on weather factors Cool months (N=58) Hot months (N=57) Regression model Coefficient adj-r 2 Coefficient adj-r 2 Lowest temperature (B L-temp ) Highest temperature (B H-temp ) Average wind speed (B wind ) Average relative humidity (B RH ) Average cloud coverage (B cloud ) 1.79 Constant

6 8 Hong Kong j. emerg. med. Vol. 24(1) Jan 2017 results by different ambulance command units. It shows that Hong Kong Island registered the highest adj-r 2 in both cool and hot months (0.55 and 0.46 respectively). In contrast, the New Territories had the lowest adj-r 2 (0.12 in cool months and 0.22 in hot months). The observed differences between Hong Kong Island and the New Territories might be attributable to the spatial distribution of people aged 65 and older. As the elderly population is more easily affected by extreme temperatures, the regression model can better explain average daily ambulance demand on Hong Kong Island, which has a higher proportion of the elderly population (13.7%) than the New Territories (10%). Table 4 shows regression results that considered temperature as the only weather factor affecting ambulance demand. The regression coefficients of temperature for the hot months were always greater than those of the cool months, which might infer that hot temperatures had greater absolute impacts on the average daily ambulance calls. The regression equations also suggest that about 20 extra cases per day were expected for every 1 o C increase in the highest daily temperature during the hot months. Similarly but to a lesser extent, 11 extra cases per day were expected for every 1 o C decrease in the lowest daily temperature during the cool months. Table 4 also shows that the adj-r 2 values for both cool and hot months (0.46 and 0.41 respectively) for Hong Kong Island were much higher than those of the New Territories (0.12 and 0.22 respectively). We could infer from these coefficients that temperature is a more significant factor impacting the average daily ambulance calls on Hong Kong Island than the New Territories because a higher adj-r 2 means a higher proportion of explained variance. This observation is also consistent with the 2006 Population By-census reporting that a higher proportion of the elderly population, who were more susceptible to temperature changes, lived on Hong Kong Island than in the New Territories. Forecasts of daily average ambulance demand Figure 4 shows the AAPEs of forecast results over a 12-month period by three different forecast models. The ARIMA model exhibited the lowest average percentage errors for shorter-term forecasts and performed the best for 1- to 7-month forecasts. However, the simple linear regression model, which maintained a stable error rate of about 3%, surpassed the ARIMA model in prediction accuracy for longerterm forecasts. Both the ARIMA and seasonal adjustment models showed increasing instability for longer-term forecasts beyond 5 months. As the linear regression and ARIMA models had better overall performance, we attempted to add another factor of average temperature into the models to assess possible improvements in the prediction accuracy. Figure 5 shows noticeable improvement in the prediction accuracy with the inclusion of average temperature. The ARIMA model displayed slightly better improvement with smaller forecast errors than the linear regression model but the plots exhibited no time trend in the size or direction of forecast errors. Figure 4. Forecast errors of the three selected models. ARIMA: autoregressive integrated moving average Figure 5. Forecast error of the models with and without average temperature as a predictor.

7 Wong et al./modelling of monthly Hong Kong ambulance demand 9 Discussion The study employed the harmonic analysis and a simple method of weighted linear regression to assess the effects of seasonality and weather factors on ambulance demand and possible geographic variation. There was clearly a bi-seasonal pattern on ambulance demand, especially in areas with a higher proportion of elderly and lower income population who were more sensitive to changing weather conditions. The performances of three different forecast models of average daily ambulance demand based on weather factors were compared. Our empirical findings showed that the ARIMA and seasonal adjustment models without consideration of weather factors appeared more suitable for shorter-term forecasts of 1 to 6 months. Additionally, the linear regression model performed better for longer-term forecasts of more than 6 months. The results revealed that both the ARIMA and seasonal adjustment models placed more emphasis on recent observations, whereas the simple linear regression model treated all data in the series equally. It is common to expect more recent observations to be a better indicator of what is to come. In this regard, the ARIMA and seasonal adjustment models are more suited for shorter-term forecasts. By contrast, the simple linear regression model, which gives equal weight to all observations in a time series, has a more stable error rate. With both the regression and ARIMA models also showing improved forecast accuracy with the inclusion of average temperature, it appears that the ARIMA model has a better overall performance. Limitations Data resolution is a major limitation of this study. Although the study reported some observable patterns, the results were subject to certain data artefacts. For example, daily variation between the lowest and the highest temperatures could be very large but it was not reflected adequately in the monthly statistics. Moreover, public holidays might also have an impact on ambulance demand. 26 The interaction effects between relative humidity and temperature were also not considered in the study. For example, it has been noted that people are more easily affected in hot days of high relative humidity and cold days of low relative humidity. The use of other indicator measures, such as the meteorological index of net effective temperature, 2 may be a way to account for this interaction effect. In reference to results of regression analysis, the finding about the proportion of elderly contributing more to the difference in goodness-of-fit (i.e., between Hong Kong Island and the New Territories) requires further evidence to support. Putting aside the problem of data resolution described above, the regression model did not include other socio-demographic or deterministic factors. The magnitude of the constants in the regression equations (see Tables 3 and 4) had large values but their corresponding adjusted R 2 values were not high enough to exclude the effects of unobserved factors. Moreover, and the spatial data resolution was not sufficient to enable an accurate modelling. It is expected that stronger and more accurate relationships between weather variables and average daily ambulance demand would result had actual values of daily ambulance Table 3. Results of the regression analysis by spatial units Ambulance command units Regression equation adj-r 2 Cool months (N=58) Whole Territory -11.5(B L-temp )-5.77 (B wind )-2.67(B RH ) Hong Kong Island -3.29(B L-temp )-1.39(B wind )-0.23(B cloud ) Kowloon -5.04(B L-temp )-1.83(B wind ) New Territories -3.04(B L-temp ) Hot months (N=57) Whole Territory 19.35(B H-temp )+1.79(B cloud ) Hong Kong Island 4.8(B H-temp )+0.33(B cloud ) Kowloon 6.8(B H-temp )+0.91(B cloud ) New Territories 7.81(B H-temp )

8 10 Hong Kong j. emerg. med. Vol. 24(1) Jan 2017 Table 4. Results of regression analysis with a single variable of temperature Ambulance command units Regression equation adj-r 2 Cool months (N=58) Whole Territory (B L-temp ) Hong Kong Island -3.1(B L-temp ) Kowloon -4.75(B L-temp ) New Territories -3.04(B L-temp ) Hot months (N=57) Whole Territory 20.15(B H-temp ) Hong Kong Island 4.77(B H-temp ) Kowloon 7.31(B H-temp ) New Territories 7.81(B H-temp ) demand by ambulance depots been available for analysis. Conclusions A better understanding of weather effects on emergency needs by locality can help inform the vulnerability of certain communities in different weather situations. Such knowledge is helpful to the FSD as the managing authority to better prepare for expected rise in the demand for emergency services. Logistic plans to redeploy ambulances based on prior knowledge of service hot spots can assure ambulance and paramedic services be dispatched to locations where they are most needed and in the most expedient manner. In the long run, more accurate forecasts can lead to improved capital planning to reduce or eliminate unbudgeted or unnecessary purchases of ambulances without compromising preparedness and service quality. Acknowledgements This research is jointly supported by the General Research Fund from the Research Grants Council of Hong Kong and the Hui Oi Chow Trust Fund Project We are grateful to the following government departments of the Hong Kong Special Administrative Region for access to data records used in the present study: Census and Statistics Department, Hong Kong Observatory, and Fire Services Department. Declaration Earlier versions of this paper were printed and presented at the 10th Asian Urbanization Conference, 27 and the 11th International Conference on Computers in Urban Planning and Urban Management. 28 References 1. Braga ALF, Zanobetti A, Schwartz J. The effect of weather on respiratory and cardiovascular deaths in 12 U.S. cities. Environ Health Perspect 2002;110(9): Leung YK, Yip KM, Yeung KH. Relationship between thermal index and mortality in Hong Kong. Meteorol Appl 2008;15(3): Vaneckova P, Hart MA, Beggs PJ, de Dear RJ. Synoptic analysis of heat-related mortality in Sydney, Australia, Int J Biometeorol 2008;52(6): WMO. Weather, climate, and health. World Meteorological Organization 1999; pp WMO/WHO/UNEP. Climate change and human health. World Health Organization 1996; pp Yan YY. The influence of weather on human mortality in Hong Kong. Soc Sci Med 2000;50(3): Keatinge WR. Winter mortality and its causes. Int J Circumpolar Health 2002;61(4): Rooney C, McMichael AJ, Kovats RS, Coleman MP. Excess mortality in England and Wales, and in Greater London, during the 1995 heat wave. J Epidemiol Community Health 1998;52(8): Stephenson PJ. French Heatwave Politics - Repenting the cost of pentecost: charity begins at work or government guilt trip? Proceedings, Annual meeting of the ISA's 49th Annual Convention - Bridging Multiple Divides, San Francisco, CA, USA, Kovats RS, Koppe C. Heatwaves: past and future impacts. In Integration of public health with adaptation to climate change: lessons learned and new directions, Ebi K, Burton I, Smith J, Eds. Taylor & Francis, 2003;

9 Wong et al./modelling of monthly Hong Kong ambulance demand 11 pp Gault G, Larrieu S, Flamand C, Filleul L. Health impact of the 2006 heat wave based on syndromic surveillance in Gironde, France. Advances in Disease Surveillance 2007;4: Gosling SN, Mcgregor GR, Paldy A. Climate change and heat-related mortality in six cities part 1: model construction and validation. Int J Biometeorol 2007; 51(6): Gosling SN, Lowe JA, Mcgregor GR, Pelling M, Malamud BD. Associations between elevated atmospheric temperature and human mortality: a critical review of the literature. Clim Change 2009;92 (3): Hong Kong Observatory. Climate of Hong Kong. [cited 2014 May 23]. Available from: climahk_e.htm. 15. Tandberg D, Tibbetts J, Sklar DP. Time series forecasts of ambulance run volume. Am J Emerg Med 1998;16 (3): Liao PS, Tzeng YS, Chen TS. Ambulance Run volume prediction by back-propagation neural network. Proceedings, 4th China-Japan-Korea Joint Symposium on Medical Informatics, CJKMI 2002, Beijing, 2002; pp Tong S, Hu WB. Climate Variation and incidence of Ross River in Cairns, Australia: a time series analysis. Environ Health Perspect 2001,109(12): Ajdacic-Gross V, Lauber C, Sansossio R, Bopp M, Eich D, Gostynski M, et al. Seasonal associations between weather conditions and suicide - evidence against a classic hypothesis. Am J Epidemiol 2007;165(5): Abeku TA, Vias SJ, Borsboom G, Teklehaimanot A, Kebede A, Olana D, et al. Forecasting malaria incidence from historical morbidity patterns in epidemic-prone areas of Ethiopia: a simple seasonal adjustment method perform best. Trop Med Int Health 2002;7(10): Allard R. Use of time series analysis in infectious diseases surveillance. Bull World Health Organ 1998;76(4): Yip PSF, Chao A, Ho TP. A re-examination of seasonal variation in suicides in Australia and New Zealand. J Affect Disord 1998;47(1-3): Box GEP, Jenkins GM, Reinsel GC. Time series analysis: forecasting and control. John Wiley, 2008, pp IBM: IBM SPSS Forecasting. [cited 2014 May 23]. Available from: products/en/spss-forecasting. 24. Jones SM, Thomas AP, Evans RP, Welch SM, Haug PM, Snow GP. Forecasting daily patient volumes in the emergency department. Acad Emerg Med 2008;15(2): Census and Statistics Department: 2006 Population Bycensus, Statistics on Map Dashboard. [cited 2014 May 23]. Available from: dashboard/index_en_2006.html. 26. Tai CC, Lee CC, Shih CL, Chen SC. Effects of ambient temperature on volume, specialty composition and triage levels of emergency department visits. Emerg Med J 2007;24(9): Wong HT, Lai PC. Time series forecasts of monthly ambulance calls in Hong Kong, the 10th Asian Urbanization Conference, held in Hong Kong on August Wong HT, Lai PC. The inference of weather and its spatial variability on the demand for emergency ambulance services in Hong Kong, 11th International Conference on Computers in Urban Planning and Urban Management, held in Hong Kong on June 2009.