Research
Publications, preprints, and ongoing projects spanning machine learning, causal inference, and health informatics.
We introduce SciReason-1K, a benchmark of 1,000 expert-curated problems spanning physics, chemistry, and biology, and evaluate state-of-the-art LLMs against domain specialists. Our findings reveal systematic gaps in causal reasoning and quantitative inference that persist across model scales.
We propose a unified pipeline for causal effect estimation in large electronic health record datasets, combining propensity score weighting with doubly-robust estimators and sensitivity analyses for unmeasured confounding. Validated on three real-world cohorts.
We systematically study how label imbalance, feature shift, and temporal drift individually and jointly affect federated learning convergence and fairness. We propose FedAdapt, an adaptive aggregation strategy that reduces worst-silo accuracy gaps by up to 40%.