Interpretable machine learning methods for stroke prediction

Abstract

Machine learning has long been touted as the next big tool, revolutionizing scientific endeavors as well as impacting industries like retail and finance. Naturally, there is much interest in the potential of next improving healthcare. However, using traditional machine learning approaches in this domain has many difficulties, chief among which is the issue of interpretability. We focus on the medical condition of stroke, a particularly desirable problem to target because it is one of the most prevalent and yet preventable conditions affecting Americans today. In this thesis, we apply novel interpretable prediction techniques to the problem of predicting stroke presence, location, acuity, and mortality risk for patient populations at two different hospital systems. We show that using an interpretable, optimal tree-based approach is roughly as effective if not better than black-box approaches. Using the clinical learnings from these studies, we explore new interpretable methodologies designed with medical applications and their unique challenges in mind. We present a novel regression algorithm to predict outcomes when the population is comprised of notably different subpopulations, and demonstrate that this gives comparable performance with improved interpretability. Finally, we explore new natural language processing techniques for machine learning from text. We propose an alternate end-to- end framework for going from unprocessed textual data to predictions, with an interpretable linguistics-based approach to model words. Altogether, this work demonstrates the promise that new parsimonious, interpretable algorithms have in the domain of stroke and broader healthcare problems.