Personalizing Suicide Risk Assessment: Machine Learning Extraction of Cross-Modal Interactions Between Psychosocial and Demographic Factors in Veterans

Veterans face elevated suicide risk compared to the general population. The research focused on developing predictive models using machine learning that combine structured data from electronic health records with textual information from clinical notes. Researchers analyzed data from 27,241 veterans and examined clinical documentation from 30, 90, or 270 days before death. They used XGBoost method to identify cross-modal interactions between textual and clinical variables. Results showed that incorporating these interactions into models improved predictive accuracy, particularly among patients with low and medium risk levels. The text-based representation outperformed the previous semantic approach. The study demonstrates the utility of interpretable natural language processing methods in uncovering clinically meaningful factors affecting suicide risk.