Statistical and AI methods for biomedical research.

Most biomedical data violate the standard assumptions of classical causal inference and structure learning. Variables are mixed-typed; samples come from complex surveys with informative weights; outcomes arrive as ordinal Likert ratings; observations are distributed across institutions for privacy reasons. My methodological work targets this gap — extending information-theoretic, generative, and graphical methods to operate under realistic data conditions without sacrificing interpretability or inferential validity.

Three threads run through the work. The first develops information-theoretic measures of asymmetric association — copula-based mutual information for high-dimensional screening, and generative exposure mapping models that recover causal direction from observational data. The second extends ordinal Bayesian network learning to handle survey weights and observed covariates jointly, enabling causal-graph estimation from large national health surveys. The third develops federated estimators for generalized estimating equations, allowing population-level inference from distributed multicenter cohorts without pooling individual-level data.

These methods find application across substantive domains: epigenetic regulation networks, directional structure of self-reported mental-health symptoms, multicenter pharmacoepidemiology cohorts, and Veteran social-determinants surveys.

The U.S. Food and Drug Administration's MAUDE database receives hundreds of thousands of medical-device adverse-event narratives each year, written in clinical and patient prose. The narratives carry safety signal — root causes, clinical consequences, contextual cues about who was using the device and how — but the sheer volume makes manual review infeasible. Surveillance at scale requires automated extraction; doing it well requires methods that are auditable, transparent, and grounded in domain expertise.

With colleagues at VA Pittsburgh, I co-develop AI models for adverse-event surveillance. Current systems combine rule-based multi-label classifiers for high-precision, auditable extraction of known risks with deep-learning and unsupervised methods for detecting emerging patterns that fall outside existing coding templates. All work is conducted within the VA Trustworthy AI Framework, with human-in-the-loop validation at every stage.

The current substantive 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, and broaden the pipeline to handle the unique reporting conventions of internal VA patient-safety systems.

At VA Pittsburgh's Center for Healthcare Evaluation, Research, and Promotion (CHERP), I serve as the lead statistician on a portfolio of health services research projects examining how Veterans receive — or fail to receive — evidence-based care. My role spans study design, causal inference, survey methodology, and pharmacoepidemiologic analysis. Several of these collaborations also surface methodological problems that motivate the work in Thrust I.

Active projects fall into five substantive areas. Medication initiation and persistence work, with Drs. Tim Anderson and Utibe Essien, examines hospital-discharge prescribing for alcohol use disorder, guideline-directed therapy for heart failure with reduced ejection fraction, and SGLT2 inhibitor adherence. Social determinants and equity work, with Dr. Leslie Hausmann, focuses on health-related social needs among LGB+ Veterans and racial, ethnic, and sex-based concordance in social-risk reporting. Cancer prevention work, with Dr. José Zevallos, models the projected burden of HPV-positive oropharyngeal cancer in the Veteran population under alternative vaccination strategies. Pharmacogenomic implementation work, with Dr. Christine Kistler, examines whether VA's national pharmacogenomic program reduces adverse drug events among older Veterans in nursing-home care. Sarcoidosis and military exposures work, with Dr. Mohamed Seedahmed, identifies disease phenotypes among Veterans with documented exposure histories and pairs prognostic-tool development with target-trial emulation of first-line therapies.