Interpretable Machine Learning for Medical Applications

Abstract

Machine learning models are increasingly deployed in high-stakes medical settings, yet many remain opaque, limiting clinician trust and posing challenges to safe deployments. In clinical environments, traditional uninterpretable (“black box”) models are prone to silent failures due to poor generalization, reliance on medically irrelevant features, and sensitivity to changing data distributions. These limitations lead to significant risks when model predictions directly influence patient care.In this dissertation, I develop inherently interpretable machine learning models that align with domain knowledge, enabling clinicians to directly inspect the reasoning behind each prediction and correct it when necessary. For wearable cardiac monitoring (Chapter 2), I introduce methods that are interpretable at both the raw signal level and the learned representation level. In particular, I present SegMADe, which identifies localized segments corrupted by motion artifacts so unreliable regions can be excluded from downstream analysis; and SiamAF, which learns physiology-grounded representations for atrial fibrillation detection, yielding robustness across diverse patient populations and sensing hardware. In Chapter 3, I develop a case-based reasoning model for intensive care EEG analysis that matches expert-level performance while providing clinicians with faithful explanations of its predictions and enabling new clinical insights. To address challenges arising from dataset distribution shift in practice, Chapter 4 introduces a framework for explaining differences between datasets through influential examples and prototype-based concept discovery and summarization. This approach translates abstract distributional discrepancies into actionable insights, supporting safer model adaptation and deployment. Collectively, these contributions advance interpretable machine learning methods for clinical decision support.