Multi-Modal Ensemble Framework for Mental Health Disorder Prediction: A Novel Machine Learning Approach

Mental health disorders constitute a major global public health concern, affecting millions of individuals across diverse socioeconomic and cultural contexts. Accurate prediction of mental health outcomes at the population level remains challenging due to the complex and non-linear relationships among co-occurring disorders. Previous studies relying on traditional statistical approaches, particularly linear regression, have reported limited predictive performance, with an R² of approximately  0.7175. This limitation highlights the need for more advanced analytical frameworks capable of capturing comorbidity patterns and non-linear interactions among mental health conditions. This study proposes and evaluates a novel multi-modal ensemble machine learning framework to improve the prediction accuracy of eating disorder prevalence using global mental health data. The analysis utilizes country-level prevalence data for schizophrenia, depression, anxiety, bipolar disorder, and eating disorders across multiple countries and years. Eating disorder prevalence is modeled as the primary target variable, while other mental health disorders are incorporated as predictive features to represent clinically established comorbidity relationships. To enhance the representational capacity of the data, an extensive feature engineering strategy was applied, generating 19 additional features through polynomial transformations, interaction terms, ratio-based indicators, and aggregate burden measures. Unsupervised clustering techniques, including K-Means, DBSCAN, and hierarchical clustering, were employed to identify natural groupings of countries based on their mental health profiles. Furthermore, ten machine learning algorithms were systematically evaluated, including linear models, tree-based methods, neural networks, and support vector regression. The best-performing models were subsequently integrated into a stacking ensemble architecture. Experimental results demonstrate that the proposed stacking ensemble achieved a test R² score of 0.9955, corresponding to a 42.2% improvement over the baseline linear regression model. These results indicate that multi-modal ensemble approaches substantially enhance predictive accuracy and provide valuable insights to support evidence-based global mental health policy and targeted intervention planning.