Optimizing the predictive power of depression screenings using machine learning

Objective: Mental health self-report and clinician-rating scales with diagnoses defined by sum-score cut-offs are often used for depression screening. This study investigates whether machine learning (ML) can detect major depressive episodes (MDE) based on screening scales with higher accuracy than best-practice clinical sum-score approaches. Methods: Primary data was obtained from two RCTs on the treatment of depression. Ground truth were DSM 5 MDE diagnoses based on structured clinical interviews (SCID) and PHQ-9 self-report, clinician-rated QIDS-16, and HAM-D-17 were predictors. ML models were trained using 10-fold cross-validation. Performance was compared against best-practice sum-score cut-offs. Primary outcome was the Area Under the Curve (AUC) of the Receiver Operating Characteristic curve. DeLong's test with bootstrapping was used to test for differences in AUC. Secondary outcomes were balanced accuracy, precision, recall, F1-score, and number needed to diagnose (NND). Results: A total of k = 1030 diagnoses (no diagnosis: k = 775; MDE: k = 255) were included. ML models achieved an AUC(QIDS- 16) = 0.94, AUC(HAM- D-17) = 0.88, and AUC(PHQ- 9) = 0.83 in the testing set. ML AUC was significantly higher than sum-score cut-offs for QIDS-16 and PHQ-9 (ps <= 0.01; HAM_D-17: p = 0.847). Applying optimal prediction thresholds, QIDS-16 classifier achieved clinically relevant improvements (Delta balanced accuracy = 8%, Delta F1-score = 14%, Delta NND = 21%). Differences for PHQ_9 and HAM-D-17 were marginal. Conclusions: ML augmented depression screenings could potentially make a major contribution to improving MDE diagnosis depending on questionnaire (e.g., QIDS-16). Confirmatory studies are needed before ML enhanced screening can be implemented into routine care practice.