Supervised Machine-Learning Classification of Treatment-Resistant Depression in U.S. Claims Data

Accurate identification of individuals with treatment-resistant depression (TRD) is important to facilitate timely access to appropriate care. However, this process currently depends on subjective provider assessments and onerous medication calculations in real-world data. To reduce the burden of TRD identification, we developed TRD classification models using data extracted from a claims database. Using U.S. commercial claims data, we developed and tested two models using automated machine learning, as well as a rule-based model modified from a published TRD proxy. The highest performing model was designated the full-feature model. The parsimonious model was determined by implementing backward elimination and a clinically oriented consolidation strategy on the full-feature model. The rule-based model was adapted from a published proxy definition. The full feature model (ridge logistic regression) demonstrated the highest overall performance (AUC=0.96, F1=0.64) with 306 features. Backward elimination and implementation of the feature consolidation strategy resulted in a parsimonious model (logistic regression) with acceptable performance (AUC=0.92, F1=0.53) comprising 8 drug-class features. The rule-based model (decision tree) had the lowest AUC (0.82) and F1 score (0.40). To address the need for efficient TRD identification, we developed a parsimonious machine learning model capable of identifying individuals with TRD from claims data based on 8 drug class features. This model performs near the limit established by the full features model and has an interpretable architecture. Furthermore, it can be used to support population health and outcomes research and may reduce the subjectivity and variability in approaches to TRD identification in clinical practice.