若年者の非自殺的自己傷害における将来の自殺企図の認識確率の予測：LightGBMおよびSHAPを使用した機械学習アプローチ
Yinan Duan1, Chang Lei1, Shicun Xu2
1Vanke School of Public Health, Tsinghua University, Beijing, 100084, China.
Objectives:
Non-suicidal self-injury (NSSI) is a strong predictor and a gateway to suicide attempts (SA) among youth. Therefore, understanding how individuals with NSSI develop thoughts of SA is crucial for early prevention. This study aims to analyze which factors can predict the perceived likelihood of future suicide attempts (PLFSA) among youth engaged in NSSI.
Methods:
Among the 96,218 respondents who completed the survey, 6955 participants (with NSSI history but without lifetime SA) were stratified into three groups based on their NSSI frequency over the past year: 0 days (N = 2951), 1-4 days (N = 2350), and 5 days or more (N = 1654). We employed the Light Gradient-Boosting Machine (LightGBM) algorithm with 5-fold cross-validation to build predictive models based on 56 features across 12 categories. To enhance the interpretability of the models, Shapley Additive Explanations (SHAP) were applied to assess feature importance.
Results:
Machine learning models effectively predicted the PLFSA in individuals with NSSI (AUC = 0.85-0.95), confirming known influencing factors (i.e., history of mental health disorders) and uncovering new ones (i.e., help-seeking behavior and self-relationship, defined as self-directed emotional and cognitive patterns). Notably, as the frequency of NSSI increases, the significant influencing factors shift progressively from external to internal domains.
Conclusion:
This study provides valuable insights for the early detection and intervention of NSSI, helping to prevent its progression to PLFSA. Healthcare professionals can utilize these findings to inform personalized screening and intervention strategies for youth engaging in NSSI, with a focus on tailored mental health care and suicide prevention.
