By Tushar Mungle, Stanford University

Abstract 

In this talk, I will present ensemble machine learning approaches for robust and generalizable clinical phenotyping from real-world longitudinal EHR data. I will focus on a multimodal ensemble learning framework trained on data from one institution and externally evaluated on an independent site, demonstrating improved robustness and fairness compared to individual model baselines. These results highlight that while algorithmic strategies can address data distribution shifts, scaling evaluation to large multi-center settings is fundamentally constrained by label availability and consistency. Motivated by this bottleneck, I will discuss briefly, results from a proof-of-concept study in which large language models (LLMs) are explored as scalable tools for label generation and harmonization under weak supervision.  The broader goal of this research is to develop principled, reproducible methodologies for learning from longitudinal, heterogeneous data in real-world settings that can be replicated across domains and healthcare institutions.