Machine learning prediction of gender-based violence among women in Somalia: a comparative analysis and SHAP interpretation

Gender-based violence (GBV) remains a major public health challenge in Somalia, yet evidence on predictive risk stratification using interpretable machine learning is limited. This study compared multiple machine learning classifiers to predict GBV among Somali women and used SHAP (SHapley Additive exPlanations) to improve the transparency of model interpretation. We conducted a cross-sectional predictive analysis using the nationally representative 2020 Somalia Health and Demographic Survey (N = 12,486). Six classifiers (gradient boosting, logistic regression, support vector machine, random forest, decision tree, and k-nearest neighbors) were evaluated using stratified train-test splitting, cross-validation, F1-scores, and the area under the receiver operating characteristic curve (AUROC; with 95% confidence intervals [CIs]). The weighted prevalence of GBV was 34.76%. Gradient boosting demonstrated the best balance of performance, achieving an F1-score of 61.37%, a recall of 55.01%, and an AUROC of 81.46% (95% CI: 79.74%–83.16%). Across all models, marital status and age ranked as the strongest predictors in our models. SHAP analysis revealed that divorced status, older age, and higher parity were strongly associated with higher predicted risk. Interpretable machine learning, particularly gradient boosting with SHAP, may serve as a hypothesis-generating tool for prioritizing screening resources in Somalia. However, the moderate recall indicates a “predictive ceiling” when relying solely on socio-demographic data, highlighting the need to integrate contextual and behavioral variables into future risk assessments.