Assessment of a Prediction Model for Antidepressant Treatment Stability Using Supervised Topic Models

Area of Science:

• Computational psychiatry
• Clinical informatics
• Pharmacogenomics

Background:

• Major depressive disorder pharmacologic management often involves trial and error due to a lack of validated treatment response predictors.
• Predicting treatment response is crucial for optimizing patient outcomes and reducing healthcare costs.

Purpose of the Study:

• To evaluate a predictive model utilizing electronic health records (EHRs) to identify predictors of antidepressant treatment response in major depressive disorder (MDD) patients.
• To assess the model's ability to predict general treatment stability versus response to specific antidepressant medications.

Main Methods:

• A retrospective cohort study analyzed EHR data from 81,630 adults diagnosed with MDD across two academic medical centers.
• Supervised topic models extracted 10 interpretable covariates from coded clinical data to predict treatment stability (defined as 90-day continuous antidepressant prescription).
• Models were trained and validated using data from separate hospital systems, employing generalized linear models and decision tree ensembles.

Main Results:

• The supervised topic model achieved an area under the receiver operating characteristic curve (AUC) of 0.627 for predicting general treatment stability in one site and 0.619 in another.
• Predicting stability for specific drugs did not significantly improve general stability prediction, even with more complex models and numerous covariates.
• Extracted topics coherently represented clinical concepts linked to treatment response.

Conclusions:

• Coded clinical data within EHRs show potential for predicting general antidepressant treatment response in major depressive disorder.
• The current model's predictive accuracy requires improvement for direct clinical application but establishes a transparent baseline for future research.
• The study highlights the limitations in predicting response to specific medications using current EHR data and modeling approaches.