A New Prediction Model for Evaluating Treatment-Resistant Depression

Objective: Despite a broad arsenal of antidepressants, about a third of patients suffering from major depressive disorder (MDD) do not respond sufficiently to adequate treatment. Using the data pool of the Group for the Study of Resistant Depression and machine learning, we intended to draw new insights featuring 48 clinical, sociodemographic, and psychosocial predictors for treatment outcome.

Method: Patients were enrolled starting from January 2000 and diagnosed according to DSM-IV. Treatment-resistant depression (TRD) was defined by a 17-item Hamilton Depression Rating Scale (HDRS) score ≥ 17 after at least 2 antidepressant trials of adequate dosage and length. Remission was defined by an HDRS score < 8. Stepwise predictor reduction using randomForest was performed to find the optimal number for classification of treatment outcome. After importance values were generated, prediction for remission and resistance was performed in a training sample of 400 patients. For prediction, we used a set of 80 patients not featured in the training sample and computed receiver operating characteristics.

Results: The most useful predictors for treatment outcome were the timespan between first and last depressive episode, age at first antidepressant treatment, response to first antidepressant treatment, severity, suicidality, melancholia, number of lifetime depressive episodes, patients' admittance type, education, occupation, and comorbid diabetes, panic, and thyroid disorder. While single predictors could not reach a prediction accuracy much different from random guessing, by combining all predictors, we could detect resistance with an accuracy of 0.737 and remission with an accuracy of 0.850. Consequently, 65.5% of predictions for TRD and 77.7% for remission can be expected to be accurate.

Conclusions: Using machine learning algorithms, we could demonstrate success rates of 0.737 for predicting TRD and 0.850 for predicting remission, surpassing predictive capabilities of clinicians. Our results strengthen data mining and suggest the benefit of focus on interaction-based statistics. Considering that all predictors can easily be obtained in a clinical setting, we hope that our model can be tested by other research groups.