A Bayesian Nonparametric Model for Predicting Disease Status Using Longitudinal Profiles

Transcription

1 A Bayesian Nonparametric Model for Predicting Disease Status Using Longitudinal Profiles Jeremy Gaskins Department of Bioinformatics & Biostatistics University of Louisville Joint work with Claudio Fuentes (OSU) and Rolando de la Cruz (Universidad Adolfo Ibáñez) July 29, 2018 JSM Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 1

2 The Motivating Question Women undergoing assisted reproductive therapy (ART) in order to become pregnant are at heightened risk of early pregnancy loss. The concentration of the hormone β-hcg is associated with the growth of the fetus in early pregnancy and may be predictive of abnormal pregnancy (loss of fetus or complications leading to nonterminal delivery). Normal Pregnancies Abnormal Pregnancies Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 2

3 The Motivating Question If we know (a part of) a patient s β-hcg trajectory, can we use this information to predict whether she will go on to have a successful pregnancy? Let D = 1 denote disease (abnormal pregnancy) and let Y denote the vector of HCG-values for patient i. Fit a model f (Y D = 0) for the normal pregnancy patients and a model f (Y D = 1) for the abnormal pregnancy patients. Use Bayes Theorem to get a probability of abnormal pregnancy, given the HCG trajectory. pr(d = 1 Y) = f (Y D = 1)pr(D = 1) f (Y D = 1)pr(D = 1) + f (Y D = 0)pr(D = 0). Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 3

4 The Motivating Question Normal Pregnancies Abnormal Pregnancies A couple of concerns: The abnormal pregnancy model doesn t appear to fit very well. I may believe that there are many ways for the pregnancy to fail and a model that tries to explain all of these in the same way won t work well. We need a model that allows multiple types of trajectories, but we don t necessarily know how many there should be. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 4

5 Contents 1 Bayesian Nonparametric Model for Longitudinal Trajectory 2 Computational Considerations 3 Simulation Study 4 Data Application: Assisted pregnancy in Chilean women Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 5

6 Some Notation N = number of patients In pregnancy data: N = 173 D i = disease status (1 for disease, 0 for healthy) 49 (28.3%) abnormal pregnancies t ij = the jth measurement occasion for patient i n i = number of measurement occasions for patient i Measurement times are not aligned. n i ranges from 1 to 6, with 30% having only 1 measurement. Y ij = longitudinal biomarker measurement for patient i at the jth observation (time t ij ) Y ij = log 10 (β-hcg) Y i = (Y i1, Y i2,..., Y i,ni ) = vector of biomarkers for patient i Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 6

7 2-Component Model First, let s formalize the 2-component model that we compare our approach to (Marshall and Baron, 2000). D i Bern(φ) Y ij = f (t ij ; θ Di ) + γ i + ɛ ij f (t; θ) = θ exp{ θ 2 t θ 3 } This model has two components: γ i N(0, γ 2 ) ɛ i MVN ni ( 0ni, σ 2 R i (ρ) ) Healthy: probability 1 φ and parameter θ 0 = (θ 01, θ 02, θ 03 ) Disease: probability φ and parameter θ 1 = (θ 11, θ 12, θ 13 ) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 7

8 Bayesian Nonparametric Model for Longitudinal Trajectory We extend the 2-component model to a Bayesian Nonparametric version. (θ i, φ i ) DP (α, MVN 3 (θ, Σ) Beta(a, b)) D i Bern(φ i ) Y ij = f (t ij ; θ i ) + γ i + ɛ ij γ i N(0, γ 2 ) ɛ i MVN ni ( 0ni, σ 2 R i (ρ) ), Due to the DP choice, there are many ties in the (θ i, φ i )s across patients. In essence, we have a small number of clusters with unique values of these parameters. We will have multiple types of clusters. Clusters with φ (c) near 0 will contain mainly healthy patients and clusters φ (c) near 1 mainly diseased patients. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 8

9 Bayesian Nonparametric Model for Longitudinal Trajectory Let c i {1, 2, 3,...} be the cluster indicator for patient i, we can equivalently express the model as: k 1 pr(c i = k) = ψ k = V k (1 V j ) D i c i = k Bern(φ (k) ) j=1 Y ij c i = k = f (t ij ; θ (k) ) + γ i + ɛ ij γ i N(0, γ 2 ) ɛ i MVN ni ( 0ni, σ 2 R i ) θ (k) MVN 3 (θ, Σ) φ (k) Beta(a, b) V k Beta(1, α) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 9

10 Bayesian Nonparametric Model for Longitudinal Trajectory If we observe a patient s longitudinal trajectory y, we can predict disease status as follows. pr(d = 1 y) = E{I(D = 1) y} = E [E {I(D = 1) C = k, y} y] H = φ (k) pr(c = k y) = k=1 H k=1 φ (k) ψ k MVN (y; f (t; θ (k) ), σ 2 R i (ρ) + γ 2 11 ) H h=1 ψ h MVN(y; f (t; θ (h) ), σ 2 R i (ρ) + γ 2 11 ) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 10

11 Bayesian Nonparametric Model for Longitudinal Trajectory We will also need to specify priors for the remaining distributions: α Gamma(1, 1) γ 2 InvGamma(0.1, 0.1) σ 2 InvGamma(0.1, 0.1) ρ Unif(0, 1) θ MVN 3 (1 3, 10 2 I 3 ) Σ InvWish(5, I 3 ) a = b = 0.5 Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 11

12 Contents 1 Bayesian Nonparametric Model for Longitudinal Trajectory 2 Computational Considerations 3 Simulation Study 4 Data Application: Assisted pregnancy in Chilean women Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 12

13 Markov chain Monte Carlo sampling In order to obtain any posterior inference, we will need a Markov chain Monte Carlo (MCMC) sampling algorithm. However, there are some challenges to drawing inference in our model. Label Switching: The likelihood function is invariant to permutations of the cluster labels. (Stephens, 2000). For instance, an estimate φ (1) from the MCMC sample is non-sensical because the types of patient who define the first cluster changes over the course of the sampler. Obviously, cluster-defined parameters are not identifiable, but these two parameters are estimable: Prob. two patients are in the same cluster: pr(c i = C j y i, y j ) Disease predictions: pr(d = 1 y) = H k=1 φ(k) pr(c = k y) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 13

14 Markov chain Monte Carlo sampling If all we need is to make predictions about disease status, then label switching is not an issue. But if we want to make conclusions about particular trajectories, we will need posterior samples with identifiable parameters. Based on the posterior pairwise probabilities pr(c i = C j y i, y j ), we estimate an optimal partition, the mostly likely clustering configuration. Given the optimal partition, we can run a Stage II MCMC algorithm without updating the cluster membership to get a usable posterior sample. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 14

15 Estimating the optimal partition Dahl s Method (Dahl, 2006) Find ĉ = (ĉ 1, ĉ 2,..., ĉ n ) from MCMC sample that minimizes N i=1 j=1 N [I (ĉ i = ĉ j ) pr(c i = C j y)] 2. Hierarchical Clustering Using the pairwise clustering probabilities pr(c i = C j y i, y j ), we can obtain a dendrogram describing the clustering relationships There are various choices of the linkage criteria that can be used, and we consider the average linkage and Ward s method. 2 Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data

16 Estimating the optimal partition Hierarchical Clusterings This gives us a partition for each number of clusters k, but we still have to choose the value k. Some automated methods designed to choose k based on minimizing some criteria: Gamma measure, Tau measure, Silhouette index (Charrad et al, 2014). But most approaches to selecting k require an actual (rectangular) data matrix. Under average linkage, dendrogram height h represents that for i and j assigned to different clusters pr(c i C j y) h, so we can specify a value of h, such as 0.75 or 0.9. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 16

17 Model Choices Different Models under consideration: 1 Two-component model 2 Bayesian Nonparametric model with label switching Disease prediction through Bayesian model averaging (BMA) 3 Bayesian Nonparametric model with 2-stage estimation Choose optimal clustering by Dahl s method 4 Bayesian Nonparametric model with 2-stage estimation Choose optimal clustering through hierarchical clustering under each method for choosing k Because the number of clusters and the membership are unknown, the BMA choice most accurately represents what we can learn from the data. Theoretically, predictions under BMA should be best (lowest variance) by Rao-Blackwell considerations. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 17

18 Contents 1 Bayesian Nonparametric Model for Longitudinal Trajectory 2 Computational Considerations 3 Simulation Study 4 Data Application: Assisted pregnancy in Chilean women Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 18

19 Simulation Study #1 All Groups ( n = 200 ) Group 1 ( n = 80 ) Group 2 ( n = 40 ) φ= 0 φ= 0.2 Group 3 ( n = 30 ) Group 4 ( n = 30 ) Group 5 ( n = 20 ) φ= 0.2 φ= 0.9 φ= 1 Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 19

20 Simulation Study #2 All Groups ( n = 200 ) Group 1 ( n = 150 ) Group 2 ( n = 50 ) φ= 0 φ= 1 In each case, we generate 200 datasets with 200 observations and apply each of the methods. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 20

21 Simulation Study Results Out of sample Within sample Model Clusters % error AUC % error AUC Simulation Study #1 (5 components) 2-component % (0.9) (.013) 25.0% (2.5) (.013) BMA 8.1 (1.1) 21.7% (0.7) (.007) 20.1% (3.0) (.029) Dahl 9.0 (2.9) 22.6% (0.9) (.009) 20.2% (2.9) (.030) Avg(h =.75) 5.0 (1.1) 22.8% (1.5) (.022) 20.7% (3.3) (.035) Avg(median) 7.7 (1.1) 22.9% (1.8) (.033) 20.6% (3.3) (.043) Avg(Silhouette) 4.2 (0.9) 23.0% (2.2) (.018) 21.2% (3.2) (.036). Simulation Study #2 (2 components) 2-component % (1.0) (.007) 16.0% (2.1) (.027) BMA 3.0 (0.6) 18.0% (1.0) (.008) 16.0% (2.0) (.028) Dahl 2.5 (0.8) 18.0% (1.0) (.008) 16.1% (2.2) (.027) Avg(h =.75) 2.1 (0.2) 18.0% (1.1) (.009) 16.1% (2.0) (.028) Avg(median) 2.7 (0.7) 18.2% (1.6) (.032) 16.3% (2.7) (.045) Avg(Silhouette) 2.1 (0.4) 18.0% (1.1) (.009) 16.1% (2.2) (.029). Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 21

22 Simulation Study Results Comments: If the two-component model is not correct, it produces substantially worse predictions; when two-component model is correct, BMA and two-stage BNP estimators do just as well. BMA beats the two-stage BNP predictions but not by too much. The minor loss in prediction accuracy may be justified by more clear interpretations. The Dahl method tends to have too many clusters, even though it is among the best two-stage methods in prediction. We recommend prediction using the BMA estimates and interpretation through two-stage procedure under the average linkage with fixed h = 0.75, followed by the Silhouette index (either average or Ward linkage) and Dahl s method. Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 22

23 Contents 1 Bayesian Nonparametric Model for Longitudinal Trajectory 2 Computational Considerations 3 Simulation Study 4 Data Application: Assisted pregnancy in Chilean women Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 23

24 Data Application: Assisted pregnancy in Chilean women We estimate the accuracy using within sample prediction from the full data (n = 173) and from a 25-fold cross validation (n=35, 20% test set). Full data 25-fold CV Model Clusters % error AUC % error AUC 2-component % % (1.2) (.004) BMA % % (1.1) (.003) Dahl % % (1.0) (.004) Avg(h = 0.75) % % (1.4) (.005) Avg(Silhouette) % % (1.0) (.005) Ward(Silhouette) % % (1.0) (.005) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 24

25 Data Application: Assisted pregnancy in Chilean women We consider the interpretable conclusions based on the Stage II model fits from the Silhouette index with average linkage. Cluster A Cluster B n=16 Disease prob=0.971 (0.86, 1) n=25 Disease prob=0.942 (0.83, 1) Index Cluster C Cluster D Cluster E n=2 Disease prob=0.494 (0.06, 0.93) n=13 Disease prob=0.181 (0.03, 0.42) n=117 Disease prob=0.056 (0.02, 0.1) Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 25

26 References Gaskins, J., C. Fuentes, R. de la Cruz. (2017) A Bayesian Nonparametric Model for Predicting Pregnancy Outcomes Using Longitudinal Profiles. arxiv: Charrad, M., H. Ghazzali, V. Boiteau, and A. Niknafs. (2014) NbClust: An R package for determining the relevant number of clusters in a data set. Journal of Statistical Software 61:1 36. Dahl, D.B. (2006), Model-Based Clustering for Expression Data via a Dirichlet Process Mixture Model, in Bayesian Inference for Gene Expression and Proteomics. K.-A. Do, P. Muller, and M. Vannucci (eds.) Cambridge University Press. Marshall, G., and A.E. Baron. (2000) Linear discriminant models for unbalanced longitudinal data. Stat in Med. 19: Stephens, M. (2000) Dealing with label switching in mixture models. JRSSB 62: Jeremy Gaskins University of Louisville BNP Disease Classification from Longitudinal Data 26