Baseline machine learning prediction of 2-year remission from anxiety, depression, and eating disorders among college students after population-based guided self-help: A secondary analysis of a randomized controlled trial

Background: We examined whether machine learning could identify baseline variables that predicted 2-year follow-up prevention and remission from anxiety, depression, and eating disorders following population-based digital cognitive-behavioral therapy guided self-help (D-CBTgsh).Methods: Undergraduates at risk for or meeting criteria for anxiety, depression, and eating disorders who were randomized to screening+D-CBTgsh were examined. A five-step feature selection pipeline reduced 863 baseline questionnaire items to 37 predictors spanning demographics, mental health problems, treatment motivation, and treatment targets. Six machine learning algorithms were trained and tested using nested cross-validation among students from 13 U.S. colleges (n = 1472) to discriminate who would (versus would not) achieve two-year remission and prevention of panic, generalized anxiety, social anxiety, major depressive, and eating disorders. External validation was conducted in a separate set of 13 U.S. colleges (n = 823) to assess algorithm discrimination, calibration, clinical utility, and feature importance.Findings: When externally validated, the random forest algorithm outperformed all models, yielding an area under the receiver operating characteristic curve of .759 (95% CI = .727-.790) with excellent calibration (integrated calibration index = .086), corresponding to a large effect (d = 0.99) and greater net benefit than benchmark models. Predictors included lower generalized anxiety interference, behavioral avoidance, social fear, social anxiety interference, number of worry topics, fear of observation, depressed mood, perseverative thinking, and binge eating, along with higher feelings of calm/peacefulness, perceived need for treatment, and intention and willingness to seek help. Demographic predictors included being male at birth, Asian, older, and with lower current and past-year body mass index, fewer financial difficulties, and higher parental education.Interpretation: If implemented on college campuses, this predictive algorithm based on 37 single items created a scalable self-report measure that could be easily adopted to help allocate higher-intensity resources at a population-level.Funding: ClinicalTrials.gov: NCT04162847.