- Baseline multimodal phenotypic models predicted clinically meaningful improvement after ICBT, AUCtest 0.732 to 0.749; random forest achieved best discrimination (AUC 0.749).
- Random forest and ensemble models outperformed a self-report screening benchmark (AUCtest 0.695), with significant paired DeLong tests (P=.04, P=.02, P=.03).
- Polygenic scores provided no independent predictive value; baseline phenotypic data enable feasible prognostic prediction and support prospective risk stratification.
J Med Internet Res. 2026 Aug 7;28:e100162. doi: 10.2196/100162.
ABSTRACT
BACKGROUND: Up to 50% of patients treated with internet-delivered cognitive behavioral therapy (ICBT) for depression and anxiety disorders do not experience clinically significant symptom reduction. Identifying these patients prior to the initiation of ICBT can support treatment planning.
OBJECTIVE: The aim of this study was to enhance baseline prediction of clinically meaningful improvement in patients treated with ICBT for common psychiatric disorders in routine care, which could ultimately inform treatment planning at intake.
METHODS: We developed multimodal predictive models integrating clinical, sociodemographic, and genetic data available pretreatment to predict clinically meaningful improvement in a sample of 1790 patients treated with ICBT for major depressive disorder, panic disorder, and social anxiety disorder. We applied machine learning algorithms of varying complexity (logistic regression, random forest [RF], extreme gradient boosting, support vector machines, soft voting, and stacking ensemble), with nested cross-validation, elastic net variable selection, multiple imputation, and temporal validation in a 20% holdout test set (n=356). The primary performance measure was the area under the receiver operating characteristic curve (AUC).
RESULTS: All full phenotypic models showed comparable performance (AUCtest 0.732-0.749), with RF achieving the best holdout discrimination (AUCtest 0.749, 95% CI 0.698-0.797). Compared with the benchmark model based on self-reported screening data (AUCtest 0.695, 95% CI 0.637-0.748), RF and both ensemble models incorporating register data showed higher discrimination in paired DeLong tests (P=.04, P=.02, and P=.03, respectively), whereas polygenic scores added no independent predictive value in this cohort and modeling setup (P=.97).
CONCLUSIONS: These promising results support the feasibility of baseline prognostic prediction of clinically meaningful improvement after ICBT and provide a basis for the prospective validation of model-informed risk stratification.
PMID:42567682 | DOI:10.2196/100162
Share Evidence Blueprint
Save to Google Notes
Search Google Scholar
Save as PDF
⭐ My Revision List
