Unifying Domain Adaptation and Domain Generalization for Robust Prediction Across Minority Racial Groups

Abstract

In clinical deployment, the performance of a model trained from one or more medical systems often deteriorates on another system and such deterioration is especially evident among minority patients who often have limited data. In this work, we present a multi-source adversarial domain separation (MS-ADS) framework which unifies domain adaptation and domain generalization. MS-ADS is designed to address two types of discrepancies: covariate shift stemming from differences in patient populations, and systematic bias on account of differences in data collection procedures across medical systems. We evaluate MS-ADS for early prediction of septic shock on three tasks. On a task of domain adaptation across three medical systems, we show that by leveraging data from multiple systems while accounting for both types of discrepancies, MS-ADS improves the prediction performance across all three systems; on a task of domain generalization to an unseen medical system, we show that MS-ADS can perform better than or close to the gold standard supervised models built for the system; last but not least, on a task that involves both domain adaptation and domain generalization: generalization to unseen racial groups across medical systems, MS-ADS shows robust out-performance by addressing covariate shift across different racial groups and systematic bias across medical systems simultaneously.