Identifying high-risk elderly for suicide using machine learning

South Korea faces a significant public health issue with its high elderly suicide rate. Despite efforts by the government to mitigate this issue, the challenge remains in effectively targeting the highest risk groups, which hampers policy implementation. This study introduces a machine learning-based approach to accurately identify high-risk groups for elderly suicide, aiming to enhance intervention effectiveness. Given the practical necessity of utilizing readily accessible variables for prediction—specifically, those that can be acquired not through face-to-face interviews but from administrative records—we categorized the variables into administrative, passive, and active data according to accessibility. We developed the predictive model with various machine learning methods, sampling techniques, and data combinations. Using only administrative data, the best model achieved a predictive accuracy with an area under the curve (AUC) of 0.742. Adding passive data increased the AUC to 0.749, and using all data boosted it to 0.818. To enhance the identification of high-risk groups, we prioritized increasing the model's sensitivity over accuracy. This adjustment maintained an 80% accuracy rate, achieving sensitivities of 0.510 with just administrative data, 0.516 with the addition of passive data, and 0.662 with all data types. A SHAP analysis of the variables critical to the model’s predictions indicated that disease status and personal income were most influential when using administrative data alone. With the addition of passive data, financial status also emerged as significant, while the depression index and life satisfaction became crucial in models utilizing the full data.