Predicting medical nonadherence using natural language processing

Abstract

Patient nonadherence is a multi-billion dollar problem in the United States healthcare system[1], accounting for over 93 million people and over 10% of American healthcare spending[2]. This study makes use of techniques in unsupervised natural language processing and human annotation to generate a set of predictive words that detect medical nonadherence. We defined nonadherence with more nuance than just patient medication practices by taking into account various psychosocial factors including adherence to dietary and therapeutic advice. Because of the multifaceted nature of the problem, our study analyzed the most multifaceted element of healthcare data: physician notes. We used natural language processing to extract meaningful keywords that predict nonadherence. We constructed three contextual categories of keywords that were statistically significant (p<0.05)predictors of nonadherence. Using our extracted key features, we made a nuanced contribution to the detection of nonadherence. These findings may be used to facilitate reduction of nonadherence in our healthcare system.