Machine learning for predicting unmet mental health treatment need among untreated adults with any mental illness: Evidence from a national survey.

BACKGROUND: Substantial treatment gaps persist in United States adults with mental illness, yet only a subset of untreated individuals perceive an unmet need for mental health care. We developed a machine learning model to predict perceived unmet mental health treatment need among untreated adults classified as having past-year any mental illness (AMI) . METHODS: Among 6004 untreated adults with past-year any mental illness in the 2024 National Survey on Drug Use and Health (weighted unmet need prevalence 21.0%), we trained an XGBoost classifier on sociodemographic, functional, substance use, and health variables. Discrimination was evaluated via test-set AUC and 100-fold repeated holdout validation; SHAP values assessed feature importance. RESULTS: The model achieved strong test-set discrimination, with an AUC of 0.789. Repeated holdout validation yielded a conservative expected AUC of 0.738 (SD = 0.030; 95% empirical interval, 0.670-0.785). Functional impairment remained a leading predictive domain, with WHODAS total score contributing 28.13% of cumulative gain, followed by age ≥65 years. Screening the highest-risk 10% identified 27.6% of all unmet-need cases, corresponding to a number needed to screen of 3.4. Among those reporting unmet need, believing they could handle the problem alone (70.8%), cost (66.2%), and not knowing where to go (48.5%) were the most common barriers. CONCLUSIONS: A robust XGBoost model reliably identifies untreated adults with perceived unmet mental health needs, highlighting functional impairment as the leading predictive domain and identifying cost, attitudinal barriers, and system-navigation difficulties as dominant obstacles.