Cross-trial prediction of depression remission using problem-solving therapy: A machine learning approach

Background: Psychotherapy is a standard depression treatment; however, determining a patient's prognosis with therapy relies on clinical judgment that is subject to trial-and-error and provider variability. Purpose: To develop machine learning (ML) algorithms to predict depression remission for patients undergoing 6 months of problem-solving therapy (PST). Method: Using data from the treatment arm of 2 randomized trials, ML models were trained and validated on ENGAGE-2 (ClinicalTrials.gov, #NCT03841682) and tested on RAINBOW (ClinicalTrials.gov, #NCT02246413) for predictions at baseline and at 2-months. Primary outcome was depression remission using the Depression Symptom Checklist (SCL-20) score < 0.5 at 6 months. Predictor variables included baseline characteristics (sociodemographic, behavioral, clinical, psychosocial) and intervention engagement through 2-months. Results: Of the 26 candidate variables, 8 for baseline and 11 for 2-months were predictive of depression remission, and used to train the models. The best-performing model predicted remission with an accuracy significantly greater than chance in internal validation using the ENGAGE-2 cohort, at baseline [72.6% (SD = 3.6%), p < 0.0001] and at 2-months [72.3% (5.1%), p < 0.0001], and in external validation with the RAINBOW cohort at baseline [58.3% (0%), p < 0.0001] and at 2-months [62.3% (0%), p < 0.0001]. Model-agnostic explanations highlighted key predictors of depression remission at the cohort and patient levels, including female sex, lower self-reported sleep disturbance, lower sleep-related impairment, and lower negative problem orientation. Conclusions: ML models using clinical and patient-reported data can predict depression remission for patients undergoing PST, affording opportunities for prospective identification of likely responders, and for developing personalized early treatment optimization along the patient care trajectory.