Integrated Psychological–Behavioral Predictive Model Using Explainable Machine Learning

Background: Medication non-adherence remains a major challenge in psychiatric care, yet its prediction often relies on clinical and sociodemographic characteristics, while giving comparatively limited attention to modifiable psychological–cognitive factors. Integrating mental health literacy and medication-related beliefs with explainable machine learning approaches may provide a more informative framework for understanding and predicting adherence. Objective: This study examined the association of mental health literacy and medication-related beliefs with psychotropic medication adherence and evaluated their explanatory and predictive values using conventional statistical modeling, machine learning, explainable artificial intelligence, and structural path analysis. Methods: A cross-sectional study was conducted with 225 psychiatric patients. Medication adherence; mental health literacy; and beliefs about medication necessity, concerns, harm, and overuse were assessed using self-report measures, including a MARS-derived eight-item adherence measure. Associations were examined using Spearman correlations, hierarchical regression with robust inference, and a secondary observed-variable path representation of multivariable associations with 5000 bootstrap resamples. The Elastic Net, Random Forest, and Gradient Boosting models were evaluated using repeated five-fold cross-validation. Shapley Additive exPlanations (SHAP) were used to interpret the best-performing model. Results: Greater adherence was associated with higher mental health literacy (ρ = 0.361) and stronger necessity beliefs (ρ = 0.353), whereas concern (ρ = −0.474), perceived harm (ρ = −0.528), and overuse beliefs (ρ = −0.284) were negatively associated with adherence (all p < 0.001). The psychological–cognitive model explained 40.0% of the variance in the primary eight-item adherence composite. The integrated Random Forest achieved the best out-of-sample performance (R2 = 0.402, MAE = 0.958, RMSE = 1.299; n = 208), although the improvement over the psychological–cognitive Random Forest (R2 = 0.375) was modest. SHAP identified perceived harm and medication concerns as leading predictors. Conclusions: Medication adherence in psychiatric patients was more strongly characterized by psycho-cognitive factors than by sociodemographic and clinical characteristics alone. Mental health literacy, necessity beliefs, medication concerns, and perceived harm may represent particularly informative targets for individualized adherence assessments and future intervention development. Prospective external validation is required before clinical implementation.