Development and validation of short-term, medium-term, and long-term suicide attempt prediction models based on a...

This study aimed to develop and validate prediction models for short-(3 months), medium-(1 year), and long-term suicide attempts among high-risk individuals in South Korea.

Data from the K-COMPASS cohort, a large prospective study conducted across five medical centers in South Korea between 2015 and 2023, were used. This cohort included 1246 high-risk individuals, with structured clinical assessments and follow-up data collected at multiple time points. Logistic regression and Cox proportional hazards models, along with machine learning methods (random forest, XGBoost, and random survival forest), were applied to predict suicide attempts, with internal and external validations conducted for each model.

In short-term and medium term prediction models, traditional logistic regression models showed moderate accuracy in the training cohort (AUC: 0.7461-0.8708) and lower but acceptable accuracy external validation (AUC: 0.5958-0.7051). Machine learning models showed higher accuracy in the training cohort (AUC: 0.8454-1.0000) but a decrease in the external validation cohort (AUC: 0.5948-0.7030). In long-term prediction, Cox models also demonstrated acceptable predictive accuracy, with a c-index of 0.780-0.786 in the training cohort, which decreased to 0.632-0.663 in the external validation cohort, whereas the random survival forest model showed 0.668-0.706 and 0.633-0.721 in both cohorts. The key predictors included younger age, prior suicide attempts, and psychiatric factors such as depression and anxiety.

Both traditional and machine learning models showed high accuracy in the internal validation and lower but acceptable accuracy in the external validation. Data reliance on self-reporting and missing medication specifics may affect prediction precision.