Machine Learning Model for Response to Internet-Delivered CBT vs Antidepressant Medication

Importance Many treatments exist for depression, yet none are universally effective. Multivariable predictive models support personalized treatment selection. Objective To develop a model predicting response to internet-delivered cognitive behavioral therapy (iCBT) and test its treatment specificity against antidepressant medications. Design, Setting, and Participants The Precision in Psychiatry prognostic study was a 4-week study collecting extensive baseline self-report and cognitive data online to predict early iCBT response, from February 2019 to May 2022. Patients in the iCBT group were recruited via an Irish mental health charity and a UK NHS Talking Therapies clinic. A separate antidepressant group was recruited globally online and via print advertisements. Participants were aged 18 to 70 years, fluent in English, had computer access, started treatment within 2 days of enrollment, and scored at least 10 on the Work and Social Adjustment Scale. Analysis was completed in December 2024. Exposures Low-intensity, clinician-guided iCBT with multimedia psychoeducation. Patients receiving antidepressants primarily received selective serotonin-reuptake inhibitors or serotonin-norepinephrine reuptake inhibitors. Main Outcomes and Measures Machine learning models were trained using the iCBT sample to predict change in depression severity (16-item Quick Inventory of Depressive Symptomatology–Self Report) at week 4. The best model was tested on holdout iCBT and antidepressant data. A separate model was trained on patients receiving iCBT only to assess treatment-specificity. Results Of 2674 patients screened, 883 completed baseline and final assessments, with 776 patients receiving iCBT (mean [SD] age, 31.8 [11.0] years; 600 [77.5%] female) and 107 patients receiving antidepressant medication (mean [SD] age, 30.1 [10.4] years; 78 [72.9%] female). Both samples had some treatment overlap (24% and 34%, respectively). Elastic net regression with 27 predictors best explained the variance in depression change ( R $^{2}$ = 14%; SD, 0.8%; 95% CI, 13.8%-14.2%). Key predictors included baseline depression, treatment expectation, transdiagnostic symptoms, and, less strongly, cognitive variables. The model performed well on holdout iCBT ( R $^{2}$ = 18.8%; root mean square error [RMSE], 0.88) and antidepressant ( R $^{2}$ = 17.9%; RMSE, 1.10) data. Retraining on 181 patients who received iCBT only increased treatment specificity in predictions ( R $^{2}$ = 19.3%; RMSE, 0.89) vs 71 patients who received antidepressants only ( R $^{2}$ = 10.8%; RMSE, 1.17). Conclusions and Relevance This prognostic study in a naturalistic setting found that self-reported data predicted iCBT response better than cognitive data. Model predictions generalized to patients receiving antidepressants, some of whom also received psychotherapy. Training models on single-treatment cohorts may yield more treatment-specific predictions.