Towards Systematic Evaluation of Interventions in Major Depressive Disorder

Abstract

Learning health systems have the capability to conduct systematic evaluation of medical interventions. Systematic evaluation of medical interventions with electronic health record data requires an accurate and complete representation of the patient environment including patient health status, clinician interventions, and outcomes. To understand in place medical intervention practices for Major Depressive Disorder (MDD), we model longitudinal medication trajectories and characterize trajectory clusters using an LSTM autoencoder and K-means clustering. This is followed by a logistic regression analysis of clinician intervention effects on outcomes associated with severe MDD. We find that medication trajectory clusters correspond to MDD endotypes and that clinician action effects are likely confounded by unobserved variables. Our findings inform future work that will apply causal inference and longitudinal methods to identify intervention best practices in MDD from electronic health records.