Interpretable models for automatic sleep stage scoring

Abstract

This thesis aims to combine domain knowledge with deep learning to develop interpretable yet robust models for a particular clinical decision support system, sleep staging. The method is transferable to other areas where domain knowledge can be represented by a set of computational rules. Currently, sleep staging, a cardinal step for evaluating the quality of sleep, is a manual process, done by sleep staging experts who are trained over months. Moreover, it is tedious and complex as it can take the trained expert several hours to annotate just one patient's polysomnogram (PSG) from a single night. As a result, data-driven methods for automating this process have been explored extensively by the research community and deep learning models have demonstrated state-of-the-art performance in automating sleep staging. However, interpretability which defines other desiderata has largely remained unexplored. In this thesis, we propose SLEEPER: interpretable Sleep staging via Prototypes from Expert Rules, a method for automating sleep staging which combines deep learning models with expert-defined rules using a prototype learning framework to generate simple interpretable models. It derives a prototype, which is a representative latent embedding of PSG data fragments, for each sleep scoring rule and expert-defined feature. The inference models are simple and interpretable like a shallow decision tree whose nodes are based on a similarity index with those meaningful rules and features. We evaluate the method using two PSG datasets collected from sleep studies and demonstrate that it can provide accurate sleep stage classification comparable to human experts and deep neural networks with about 85% ROC-AUC and .7 K.