JMIR Publications

ABSTRACT

Background:

Many adolescents with anxiety or depression do not receive needed mental health care, yet prediction research has largely focused on symptom or diagnostic outcomes rather than on whether symptomatic adolescents experience a treatment gap.

Objective:

This study developed and temporally validated interpretable prediction models for unmet need for mental health treatment among adolescents aged 12 to 17 years with current anxiety and/or depression.

Methods:

This analysis used the 2023 and 2024 National Survey of Children’s Health topical child-level files. The development cohort was restricted to adolescents aged 12 to 17 years with current anxiety and/or current depression in 2023; an independent 2024 cohort was used for temporal validation. The primary outcome contrasted adolescents with unmet need for mental health treatment or counseling with those who received treatment. Locked predictors covered symptom profile and severity, demographic and socioeconomic factors, caregiver characteristics, family communication and resilience, material hardship, and social support. Main models were survey-weighted logistic regression, elastic net, and explainable boosting machine (EBM). XGBoost was evaluated as a supplementary sensitivity model.

Results:

Development used 2,705 adolescents (323 cases) from 2023; validation used 3,158 adolescents (337 cases) from 2024. EBM and elastic net showed comparable performance and outperformed both XGBoost and survey-weighted logistic regression in the 2024 temporal validation cohort. Weighted area under the receiver operating characteristic curve (AUROC) was 0.685 (95% confidence interval 0.626-0.736) for EBM and 0.685 (95% confidence interval 0.627-0.734) for elastic net. Diagnostic analyses indicated that the underperformance of survey-weighted logistic regression reflected instability from extreme design weights rather than a limitation of logistic regression itself. Intercept-and-slope recalibration improved calibration for EBM and elastic net without affecting discrimination.

Conclusions:

Cross-year prediction of adolescent treatment gap was feasible, but discrimination was modest rather than high. In this national survey setting, interpretable models, especially EBM and elastic net, were competitive with or better than more flexible boosting. The study’s main contribution is a transparent, temporally validated prediction framework for a difficult service-gap outcome rather than a high-accuracy clinical decision tool. Clinical Trial: Not applicable. This study was a analysis of publicly available, deidentified survey data and did not constitute a clinical trial.