A machine learning approach for early prediction of mental health crises

Abstract

The global mental health crisis, intensified by the COVID-19 pandemic, placed unprecedented strain on healthcare systems and highlighted the urgent need for proactive crisis prevention strategies. This study investigated the effectiveness of various machine learning (ML) models in predicting mental health crises within 28 days post-hospitalization, leveraging an eight-year longitudinal dataset. Multiple data preprocessing techniques, including feature selection (EFSA, RFECV), imputation, and class imbalance handling (SMOTE, Tomek links), were systematically applied to enhance model performance. Six traditional classifiers—logistic regression, support vector machine, k-nearest neighbors, naive Bayes, XGBoost, and AdaBoost—were evaluated alongside ensemble learning (EL) methods (bagging, boosting, stacking). Performance metrics such as accuracy, precision, recall, F1 score, and AUC-ROC were used for comprehensive assessment. Results demonstrated that ensemble methods, particularly boosting and bagging, consistently achieved high predictive accuracy (up to 93%), with XGBoost and AdaBoost emerging as top performers. Feature selection and class imbalance techniques further improved model robustness and generalizability. The findings underscored the potential of ML-driven approaches for early identification of at-risk patients, enabling more effective resource allocation and timely interventions in mental health care. Recommendations for integrating these predictive tools into clinical workflows were discussed to support data-driven decision-making.