Generating interpretable predictions about antidepressant treatment stability using supervised topic models

Importance: In the absence of readily-assessed and clinically-validated predictors of treatment response, pharmacologic management of major depressive disorder (MDD) often relies on trial and error. Objective: To utilize electronic health records to identify predictors of treatment response, while preserving interpretability of predictions despite large numbers of covariates. Design: Retrospective cohort study. Setting: Two academic medical centers in Boston, including outpatient primary and specialty care clinics. Participants: 81,630 adults with a coded diagnosis of MDD. Exposure: Treatment with 1 or more of 11 standard antidepressants. Main Outcomes and Methods: Stable treatment, intended as a proxy for treatment effectiveness, defined as continued prescription of an antidepressant for 90 days. We trained supervised topic models to extract 10 interpretable covariates from coded clinical data for stability prediction. Then, using data from one hospital system (Site A) we trained generalized linear models and ensembles of decision trees to predict stability outcomes from topic features that summarize patient history. We evaluated on held-out patients from Site A as well as all individuals from a second hospital system (B). Results: Among the 81,630 adults (31% male; age 18-80 with mean 48.46), we identified 55,303 who reached a stable treatment regimen during follow-up. For held-out patients from Site A, mean area-under-the-receiver-operating-characteristic-curve (AUC) discrimination for general stability outcome was 0.627 (95% confidence interval (CI) 0.615 - 0.639) for our supervised topic model with 10 covariates. In evaluation on site B, our approach achieved similar AUC of 0.619 (95% CI 0.610 - 0.627). Building models to predict stability specific to a particular drug did not improve upon predicting general stability, even when using a harder-to-interpret ensemble classifier and 9,256 coded covariates (specific AUC = 0.647, 95% CI 0.635- 0.658; general AUC = 0.661, 95% CI 0.648-0.672). Topics coherently captured clinical concepts associated with treatment response. Conclusions and Relevance: Coded clinical data available in electronic health records facilitated prediction of general treatment response, but not response to specific medications. While greater discrimination is likely required for clinical application, our results provide a simple and transparent baseline for such studies. Funding: We acknowledge funding from Oracle Labs, Harvard SEAS, and National Institute of Mental Health.