Work

Interpretable machine learning to identify problematic alcohol use in peacekeepers

Year: 2026

Type: article

Abstract: Introduction United Nations peacekeeping missions expose military personnel to significant psychological stress, impacting mental health and operational readiness. Despite predeployment screening efforts, problematic alcohol use remains a concern. Objectives This study aimed to evaluate whether supervised machine learning combined with interpretable artificial intelligence can identify and interpret factors associated with problematic alcohol use in Brazilian peacekeepers assessed prior to deployment. Materials and methods We performed a cross-sectional analysis of predeployment data from 771 Brazilian male military personnel assessed before deployment to the United Nations mission in Haiti. Problematic alcohol use was defined by a positive CAGE questionnaire result administered before deployment. Six supervised machine learning models were trained and evaluated using traditional metrics. Interpretability was addressed with Shapley values and feature importance, to clarify variable contributions. Results The best model, LightGBM, achieved an ROC-AUC of 0.80, with 75% sensitivity and 70% specificity on the test set. Interpretability analyses consistently identified lifetime trauma exposure and posttraumatic stress symptoms as the main predictors. Discussion These results highlight the value of interpretable machine learning in predeployment screening, enabling interpretable risk stratification from routine data. The importance of trauma exposure and posttraumatic stress aligns with evidence linking cumulative trauma to alcohol problems in military populations. Conclusions Interpretable machine learning may complement traditional methods in identifying multivariable patterns of problematic alcohol use in predeployment data. It also helps mitigate concerns about model opacity in clinical and operational contexts. External validation, calibration, and clinical utility assessment are required before operational use.

Source: Addictive Behaviors Reports

DOI: 10.1016/j.abrep.2026.100763