Himel Mallick, Ph.D.
Assistant Professor
Division of Biostatistics
Weill Cornell Medical College
Abstract. Modern biomedical studies routinely generate high-dimensional, multimodal data, yet integrating these heterogeneous sources into coherent, actionable insight remains a central challenge, especially when signals are weak and sample sizes are small. In this talk, I will present a suite of scalable Bayesian multimodal AI methods designed to learn generalizable and transferable structure across data types. First, I will discuss Bayesian foundation model approaches that pretrain on large reference cohorts to capture broad, context-independent structure and then treat that structure as an informative prior for data-limited settings, enabling zero-shot posterior inference where no paired data exist and few-shot fine-tuning where only small cohorts are available, carrying knowledge from data-rich source domains to sparsely sampled target diseases with calibrated uncertainty. Second, I will present Bayesian multiview integration tools that combine paired and unpaired measurements into unified predictive and interpretable models. Time permitting, I will introduce Bayesian multimodal mediation analysis capable of yielding coherent posterior summaries of direct and indirect effects from multimodal mediators. I will conclude by showing how Bayesian multimodal AI translates into actionable insight in cancer immunotherapy and inflammatory bowel disease.