Optimizing Suicide Risk Prediction in Korea: A Comparison of Model Performance Using Resampling Methods and Machine Learning Algorithms

Abstract

Objective Machine learning (ML) can assist in predicting suicide risk and identifying associated risk factors. Various resampling methods and algorithms must be applied to develop an ML prediction model with better performance. In this study, we developed an optimal Korean suicide prediction model by applying five ML algorithms, unsampled data, and two resampling methods. Methods In this study, data from the Korea National Health and Nutrition Examination Survey for 2017, 2019, and 2021 were integrated and analyzed to predict suicidal ideation in subjects aged ≥19 years. Logistic regression, random forest (RF), k-nearest neighbor, gradient boosting, and adaptive boosting were used as ML algorithms. Undersampling and oversampling are used as resampling methods to solve data imbalance problems. Results Among the study participants, 16,947 (95.14%) and 866 (4.86%) belonged to the control and suicidal ideation groups, respectively. Among the 15 ML models, the RF model exhibited excellent performance (sensitivity=0.781, area under the curve=0.870) in an algorithm trained with undersampled data. Conclusion Developing an optimized Korean suicide prediction model through additional validation based on the ML model developed in this study will help predict suicide risk factors caused by the interaction of individual, social, and environmental factors.