What Can AI Do in Precision Psychiatry? A Study in Electronic Health Records
CitationSheu, Yi-han. 2019. What Can AI Do in Precision Psychiatry? A Study in Electronic Health Records. Doctoral dissertation, Harvard T.H. Chan School of Public Health.
AbstractTreatment selection for depressive disorders is still largely a trial-and-error process. The work described in this thesis aims to combine epidemiological concepts with contemporary techniques in AI/machine learning/data science to improve selection of initial antidepressants treatment, utilizing data from a large electronic health record (EHR) database. We focused on adult patients first treated by non-psychiatrists; such patients are the majority of those receiving treatment for depressive disorder, and they are clinically distinct from those first treated by psychiatrists.
In the first chapter, we describe how we used multinomial logistic regression to predict the class of antidepressant chosen for initial treatment using a set of predictor variables derived from literature review and expert consultation. The variables were extracted from both structured and free-text EHR data through the application of natural language processing (NLP). The study provided supportive evidence that the basis of treatment decisions for first line depression treatment among non-psychiatrists was largely consistent with factors suggested by existing literature and expert opinion.
In the second chapter, we describe how we applied a deep-neural network (DNN)-based supervised NLP model on clinical notes to classify treatment response to antidepressants. While the DNN-based approach is perceived as a paradigm shift in NLP, application of deep learning-based NLP to medical texts is still scarce and warrants evaluation. We found that the estimated classification accuracy was limited, but acceptable for certain uses, such as imputing outcome labels in appropriate cases. However, with further improvements, it appears promising for a broader set of uses.
In the final chapter, we describe our work applying a machine learning model to predict treatment response utilizing predictors constructed in the first chapter, and outcome labels produced by combining expert-curated and imputed labels derived using the model developed in the second chapter. Our results showed that clinical characteristics can predict antidepressant treatment response to some degree, suggesting that with further optimization, such methods could lead to clinically useful decision support tools. In summary, the methods described in this thesis may be a first step towards a clinical support system for the treatment of depression and other conditions alike.
Citable link to this pagehttp://nrs.harvard.edu/urn-3:HUL.InstRepos:40976536