Clinical predictors of treatment resistant depression

Abstract

Despite advances in the treatment of major depressive disorder (MDD) yet a substantial proportion of patients fail to achieve remission and instead develop treatment-resistant depression (TRD). Identifying robust clinical predictors of response is essential for early, personalized interventions. We analyzed a large, multicenter sample (N = 2953) from the Group for the Study of Resistant Depression (GSRD) project, which included previously studied cohorts (TRD I-III) and a newly recruited cohort (TRD IV, N = 294). Patients were categorized as responders, non-responders, or TRD. Sociodemographic and clinical variables, including current and retrospective MADRS items, were used to train an XGBoost classifier. Primary outcomes were the multi-class metrics area under the curve (AUC), accuracy, and F1-scores. Previously reported predictors were mainly confirmed in the new TRD IV sample. The XGBoost model showed a mean ROC AUC of 0.80 and an accuracy of 61 %, significantly above chance. Misclassification was more frequent among responders versus non-responders, while TRD was predicted most accurately (precision=0.73; recall=0.73). Measures of illness chronicity, such as duration of current episode, duration of disease lifetime, number of hospitalizations, and number of depressive episodes, as well as severity features, BMI and level of functioning were among the most important predictors. Secondary analyses using earlier cohorts to train and the new TRD IV sample to test confirmed stable performance metrics. Our findings highlight the central role of chronicity indicators, severity measures and functioning in predicting antidepressant response and TRD. Future work should include prospective validation and integration of biomarker data to further enhance predictive power.

Keywords: Antidepressants; Machine learning; Major depression; Predictive models; Treatment resistant depression.