University of Pittsburgh · VA Pittsburgh

Soumik Purkayastha

Assistant Professor, Department of Biostatistics & Health Data Science, University of Pittsburgh School of Public Health

Mathematical Statistician, Center for Healthcare Evaluation, Research, and Promotion (CHERP), VA Pittsburgh Healthcare System

I develop statistical and AI methods for discovering structure, quantifying association, and extracting safety-relevant signal from complex health data. My methodological work spans weighted Bayesian network learning, information-theoretic association estimation, and federated inference. With colleagues at CHERP, I co-develop clinical AI for medical device safety surveillance and serve as statistician/co-investigator on health services and clinical research projects, with a focus on Veterans.

i

Statistical methods for complex biomedical data

I develop graphical models for learning causal structure and quantifying directional association under realistic biomedical data conditions — survey weighting, mixed-typed variables, ordinal outcomes, and beyond rigid parametric assumptions. Recent work spans copula-based mutual information, generative exposure models and graphical models for learning network structure, and federated learning.

copula models · density estimation · federated learning · graphical models · robust inference · structure learning
ii

AI models for medical device safety

With colleagues at VA Pittsburgh, I co-develop AI models for adverse-event surveillance. Our current focus is insulin pumps and continuous glucose monitors — the second-largest source of medical-device-related injuries reported to the FDA. Active extensions integrate extracted risk signal into provider-facing registries and Veteran-facing safety education materials.

adverse-event surveillance · deep learning · foundation models · trustworthy AI
iii

Health services research (HSR)

I serve as the statistician on a portfolio of health services and clinical research projects, most of them at VA Pittsburgh's Center for Healthcare Evaluation, Research, and Promotion (CHERP). Active threads include alcohol use disorder treatment, guideline-directed therapy for heart failure, social risks among sexual and gender minority Veterans, HPV vaccination, and organ transplantation.

causal inference · micro- and macro-simulation models · pharmacoepidemiology · survey design

Working on federated generalized estimating equations for distributed health data; survey-weighted ordinal Bayesian network learning for Veteran mental-health surveys; and the impact of declining drug-overdose mortality on deceased-donor organ transplantation.

Fast and consistent copula-based nonparametric estimator of mutual information.
R · Github
Bivariate causal discovery via the generative exposure model.
Python · Github
Covariate-adjusted bivariate causal discovery via the generative exposure model.
R · Github
Survey-weighted, covariate-adjusted ordinal Bayesian network learning.
R · in development
Federated generalized estimating equations for multicenter longitudinal data.
R · in development
Extended SEIR model with false-negative correction and symptom-based testing.
R · CRAN