YSPH Biostatistics Seminar- “Transfer Learning under Real-World Heterogeneity: Across-Site Contamination”

Note: BIS 526 students are required to attend in person. Others are invited to attend in person but may also attend via zoom.

Speaker Kun Chen, Ph.D.

Title- “Transfer Learning under Real-World Heterogeneity: Across-Site Contamination”

Abstract

Transfer learning and data fusion aim to improve prediction for a target population by borrowing information from related sources. There has been an increasing need for such methods in multi-site EHR networks and health information exchange (HIE) settings. In this talk, we discuss transfer learning under real-world heterogeneity, which can be either a friend or a foe. As a friend, heterogeneity is structured: sources and the target form clusters of similarity. In this regime, we present a three-stage, cluster-aware transfer-learning procedure that improves the bias–variance tradeoff and demonstrate it in facility-specific suicide risk modeling across 27 Connecticut hospitals. As a foe, heterogeneity is unstructured: external data may be large but contaminated by arbitrary outliers. In this regime, we study how to select/sample from the external data, deriving non-asymptotic theory that makes explicit how performance depends on target size, sampling rate, signal strength, outlier magnitude, and error tail behavior. The take-home message is simple: more is not necessarily better, and how much you can borrow depends on how much you have.