Optimizing the predictive power of depression screenings using machine learning.

Area of Science:

• Psychiatry and Mental Health
• Computational Medicine
• Clinical Diagnostics

Background:

• Traditional depression screening relies on self-report and clinician-rating scales with sum-score cut-offs.
• These methods are essential but may have limitations in diagnostic accuracy for major depressive episodes (MDE).

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) in enhancing the accuracy of depression screening.
• To compare ML-based detection of MDE against established sum-score cut-off approaches.

Main Methods:

• Data from two randomized controlled trials (RCTs) on depression treatment were utilized.
• Machine learning models were trained using 10-fold cross-validation with DSM-5 MDE diagnoses as ground truth.
• Predictors included self-report (PHQ-9) and clinician-rated scales (QIDS-16, HAM-D-17); performance was evaluated using Area Under the Curve (AUC) and other metrics.

Main Results:

• ML models achieved high AUCs: 0.94 for QIDS-16, 0.88 for HAM-D-17, and 0.83 for PHQ-9.
• ML significantly outperformed sum-score cut-offs for QIDS-16 and PHQ-9 (p ≤ 0.01).
• The QIDS-16 ML classifier yielded clinically relevant improvements in balanced accuracy (+8%), F1-score (+14%), and number needed to diagnose (-21%).

Conclusions:

• ML-augmented depression screening, particularly with the QIDS-16, shows promise for improving MDE diagnosis.
• Further confirmatory studies are necessary to validate ML-enhanced screening for routine clinical practice.