Prediction of treatment-resistant depression using the 23andMe survey data

Patient-reported drug responses offer an alternative approach to design and develop a predictive model for treatment-resistant depression (TRD) in major depressive disorder (MDD). The variables described in phenotypic analysis (i.e., demographics, depression clinical characteristics, psychiatric comorbidities, and vulnerability characteristics) of the antidepressant efficacy survey from previously reported study among 23andMe, Inc. research participants were identified as predictors to classify TRD or non-TRD. A total of 101 features were engineered for 23,779 participants, with stratified split between training (80% of the total participants included in the model) and holdout (20% of the participants) datasets. Significant predictors included residual symptoms, years in life having ≥ 2 weeks depressive episode, stress level, duration of depressive episodes, age, suicidal ideation, symptoms of post-traumatic stress disorder, and latency. The final model achieved a model performance with receiver operating characteristics area under curve (AUC) at 78% and precision recall AUC at 44%. The proposed predictive model could improve the management of MDD by providing an assessment of risk factors and antidepressant treatment effects, thereby facilitating reliable identification of patients with TRD. The model could potentially address the gaps observed in real-world settings, where identification of TRD is often delayed.