Using Electronic Health Records Data for Predictive and Causal Inference About the HIV Care Cascade

Transcription

1 Using Electronic Health Records Data for Predictive and Causal Inference About the HIV Care Cascade Yizhen Xu Joseph Hogan Tao Liu Rami Kantor Department of Biostatistics School of Public Health Brown University Joint Statistical Meetings July 30, 2018 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

2 AMPATH Program in western Kenya AMPATH: Academic Model Providing Access to Healthcare PEPFAR-funded HIV care program based in Eldoret, Kenya Over 150,000 individuals in care at over 100 clinical sites Electronic health record: AMPATH Medical Record System data from several million clinical encounters augmented with lab data (CD4, others where available) stored on a central server expanding to mobile data entry Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

3 AMPATH Consortium Collaboration between Moi University and 11 North American universities Led by Indiana University Brown, Duke, Toronto, Purdue, Mount Sinai, several others. Focus on care, training, research Medical student / resident / trainee exchanges Long-term faculty in residence in Eldoret Local infrastructure that makes visits easy Kenyan-North American partnership is key feature Research program has generated significant new activity Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

4 Focus on capacity building in biostatistics Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

5 Collaborators on this project Brown Yizhen Xu Rami Kantor, MD Tao Liu, PhD Allison DeLong, MS Indiana U Beverly Musick, MS Moi / AMPATH Ann Mwangi, PhD Edwin Sang, MS U Toronto Paula Braitstein, PhD Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

6 HIV care cascade Conceptual model describing progression through stages of HIV care Key stages Identify new cases Link to care Initiate treatment Positive treatment outcomes (e.g., viral suppression) Retain in care More recently: used to frame policy goals Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

7 HIV care cascade Source: aids.gov Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

8 Goals for understanding cascade Prediction Generate predictive models of transition between states E.g., flag those who are at risk for negative outcomes Regression, Machine learning Evaluation Causal inference about a policy, treatment, exposure E.g., what is the effect of immediate treatment initiation, compared to marker-based initiation? Causal structural models Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

9 Can be complex to model progression through care Mugavero MJ, Norton WE, Saag MS. Health care system and policy factors influencing engagement in HIV medical care: piecing together the fragments of a fractured health care delivery system. Clin Infect Dis. 2011;52:S238-S246 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

10 AMPATH: Model of engagement in care Engaged Disen. 1 Disen. 2 Disen. 3 + Deceased Transfer Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

11 Challenge: Translate patient-level data into states CD4 count Days since enrollment Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

12 Challenge: Translate patient-level data into states CD4 count Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

13 Challenge: Translate patient-level data into states CD4 count CD4 count Days since enrollment Days since enrollment CD4 count CD4 count State Engaged Disengaged 1 Disengaged 2 Disengaged 3+ Transferred out Deceased Days since enrollment Days since enrollment Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

14 Summary of available data Number of observations State Engaged Disengaged 1 Disengaged 2 Disengaged 3+ Transferred out Deceased Days since enrollment Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

15 Analytic approach 1 Organize data into states 2 Specify model for observed data Transition between states Dependence of transitions on covariates Longitudinal model for covariates 3 How to use fitted observed-data models Summary of transition rates (fill in numbers on graph) Individual-level predictions Causal policy comparisons Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

16 Statistical object: Aggregated transition rates State at t j State at t j 1 Engaged Diseng. 1 Diseng. 2 Diseng. 3 + Transfer Death Engaged Diseng Diseng <.01 Diseng <.01 Assumes constant rate over time Not individual-specific Death rates under-estimated (need tracing data) Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

17 Prediction model: Observed-data regressions Multinomial regression for longitudinal data log{p jkl (x j )/p jkl (x j )} = x T j β jkl l = 1,..., L 1 p jkl (x j ) = probability of transition from k to l at time j x j = covariates at time j Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

18 Prediction model: Observed-data regressions Covariates: x j = vector of covariates observed just prior to t j CD4 count (baseline and time-varying) baseline viral load height, weight HIV stage (graded 1-4) age, gender, marital status treatment status travel time to clinic enrollment year calendar year Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

19 Regression models Multinomial regression for transition from engaged at j = 3 (day 600) State at t j 1 Engaged State at t j Disengaged Transfer Death age 0.02* 0.01* 0.01* male 0.18* Enrollment Year * 0.06* TravelTime WHO stage 0.05* * Married 0.15* * Height log Weight 0.26* * undetectable VL Previous ARV 0.38* 0.21* 0.12 CD4 Update 2.20* 1.49* 0.52* latest log CD * 0.13* 0.31* Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

20 Regression models Fit at each time and for each transition Can be used for prediction and/or variable selection Current version has linear covariate effects Can generalize to use machine learners for more flexibility e.g., BART for multinomial outcomes Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

21 Causal modeling Question: How would treat immediately impact progression through the care cascade? Comparison regimes: Policy 1: Treat immediately upon enrollment Policy 2: Treat when CD4 falls below 350 Outcome: State membership probability at each time interval Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

22 Causal structural model to compare treatment policies Structural model S j = state membership at time t j a j = treatment assigned at time t j a j = (a 0,..., a j ) P aj (S j ) = distribution of S j under regime a j To compare two different regimes a and a, want to compare P a (S J ) and P a (S J ) Example: treat immediately is the regime a J = (1, 1, 1,..., 1) Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

23 What we need to estimate causal models Observed data S j = state at time j X j = time-varying confounders (CD4) V = baseline confounders (age, gender, site, CD4) A j = observed ART status at time j Collection of predictive models P(Sj S j 1, X j 1, A j, V ) P(Xj S j 1, X j 1, A j, V ) Assumptions No unmeasured confounders First-order Markov dependence for S and X Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

24 G computation for estimating causal quantities Target: P a0 (S 1 ) when a 0 = 1 (state membership distribution if everyone receives treatment at baseline) Confounders: X 0 = baseline CD4 count, V = (age, gender) G computation: Implementation P 1 (S 1 ) = P(S 1 A 0 = 1, X 0, V ) P(X 0, V ) d(x 0, V ) n P 1 (S 1 ) = (1/n) P(S 1 A 0 = 1, X 0i, V i ) i=1 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

25 G computation for estimating causal quantities Target: P a0,a 1 (S 2 ) Patient state probabilities at time t = 2 under treatment regime a 0, a 1 Confounders: X j = CD4 count (could be other stuff) V = (age, gender, other baseline covariates) Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

26 How to use observed-data models as plug-ins Target: P a0,a 1 (S 2 ) P a0,a 1 (S 2 ) = P(S 2 A 1 = a 1, X 1, S 1, V ) P(X 1 A 0 = a 0, X 0, V, S 1 ) P(S 1 A 0 = a 0, X 0, V ) P(X 0, V ) d(s 1, X 1, X 0, V ) Plug in fitted models for state transitions Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

27 How to use observed-data models as plug-ins Target: P a0,a 1 (S 2 ) P a0,a 1 (S 2 ) = P(S 2 A 1 = a 1, X 1, S 1, V ) P(X 1 A 0 = a 0, X 0, V, S 1 ) P(S 1 A 0 = a 0, X 0, V ) P(X 0, V ) d(s 1, X 1, X 0, V ) Plug in fitted models for time-varying covariates Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

28 How to use observed-data models as plug-ins Target: P a0,a 1 (S 2 ) P a0,a 1 (S 2 ) = P(S 2 A 1 = a 1, X 1, S 1, V ) P(X 1 A 0 = a 0, X 0, V, S 1 ) P(S 1 A 0 = a 0, X 0, V ) P(X 0, V ) d(s 1, X 1, X 0, V ) Fix treatment regime or policy a 0, a 1 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

29 How to use observed-data models as plug-ins Target: P a0,a 1 (S 2 ) P a0,a 1 (S 2 ) = P(S 2 A 1 = a 1, X 1, S 1, V ) P(X 1 A 0 = a 0, X 0, V, S 1 ) P(S 1 A 0 = a 0, X 0, V ) P(X 0, V ) d(s 1, X 1, X 0, V ) Average over the distribution of specific population of interest Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

30 Implementation 90,000 individuals with up to 10 years of follow up Fit predictive models (with priors) Use 1/3 of data to fit predictive models Outcome models P(S j A j 1, X j 1, V ) Use multinomial regression Time varying covariate models P(X j A j 1, X j 1, S j 1, V ) Use BART Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

31 Implementation Validate predictive models Use 1/3 validation dataset to compare posterior predictions to observed outcomes In this table: P(S pred = S obs ) Selected time points t = 1 t = 3 t = 5 t = 7 t = 9 Engaged Disengaged Transfer Death average mode Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

32 Implementation Implement G computation using validated models Use 1/3 of data to generate posterior predictive outcomes under different treatment policies using G computation algorithm Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

33 Treat if CD4<350 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

34 Treat upon enrollment ( test and treat ) Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

35 Inferences Next few slides: Compare proportions in each state over time Use rate difference, 95% posterior predictive interval Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

36 Engaged in care Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

37 Disengaged from care Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

38 Mortality Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

39 Substantive conclusions Inferences suggest strong benefit of treatment Higher engagement in care Lower loss to follow up Importance of disengaged finding Many of those disengaged are likely to be deceased Estimates available from tracing data Mortality can be as high as 20% (Yiannoutsos et al, 2016) Consequence: Preventing LTFU preventing mortality Quantifying this = data integration problem Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

40 Mathematical modeling approaches Example: Ying et al., Lancet HIV, 2016 Model of individual-level progression through 9 HIV disease states Simulates HIV incidence and prevalence over 45 year period Uses model to capture effect of specific interventions Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

41 Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

42 Source of selected inputs Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

43 Mathematical and statistical modeling for causal inference Mathematical Modeling: Focus on the model (more model, less data) Statistical modeling: Focus on the data (more data, less model) Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

44 Resolution (Data) Stat model Math model Complexity Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

45 Resolution (Data) Stat model Math model Hybrid model Complexity Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46

46 Capacity Building Xu, Hogan, Kantor, Liu Brown.edu) HIV Care Cascade July 30, / 46