Predicting suicide attempts among the elderly: A machine learning evaluation of a staged ideation-to-action approach...

Suicide among older adults remains a major global public health concern, particularly acute in South Korea. Effective prevention requires prediction methods grounded in robust theoretical frameworks. This study evaluates the utility of a Three-Step Theory (3ST) based approach-which posits a sequential progression from suicidal ideation to attempts-using advanced machine learning on nationally representative data from South Korea's Elderly Survey. We compare two predictive strategies: (a) a stepwise model that explicitly estimates suicidal ideation before attempts, and (b) a direct model that predicts attempts without this intermediate stage. Results strongly favor the stepwise approach (sensitivity = 0.531, specificity = 0.953, AUC = 0.828), suggesting both practical and theoretical advantages of modeling suicidal risk within a staged ideation-to-action framework. Error analysis further indicates that the two models capture complementary risk dimensions-progression-consistent versus early, non-progression markers-while a linear ensemble combining them achieves higher sensitivity with minimal loss of specificity. To address data imbalance, we evaluate oversampling (SMOTE) and undersampling techniques. Finally, SHapley Additive exPlanations (SHAP) analysis enhances interpretability and reveal distinct predictors across stages that are consistent with the sequential ideation-to-action logic proposed by the 3ST: depressive symptoms and life satisfaction in the first, and social connectedness and functional capacity in the second. Our findings provide new evidence consistent with 3ST and may inform the development of more precisely targeted interventions within large-scale suicide prevention policy.