Multivariate machine learning regression approaches to predict adolescent suicide risk.

Area of Science:

• Psychiatry
• Machine Learning
• Adolescent Health

Background:

• Adolescent suicide is a critical public health issue.
• Existing prediction models often use binary classification and limited variables, failing to capture suicide risk complexity.
• There is a need for more comprehensive suicide risk prediction models.

Purpose of the Study:

• To develop an accurate and comprehensive suicide risk prediction model for adolescents.
• To utilize machine learning regression with a broad range of psychological and behavioral variables.
• To identify key predictors of suicide risk in a school-based adolescent population.

Main Methods:

• Recruited 2241 adolescents from school settings.
• Administered self-report questionnaires assessing 31 psychological and behavioral variables.
• Applied Lasso regression, support vector regression, and random forest regression, evaluated using 10-fold cross-validation and R².

Main Results:

• Lasso regression identified 13 significant features from the 31 variables.
• Random forest regression demonstrated the best predictive performance with an R² of 0.61.
• Top predictors included depression, emotion regulation, perceived burdensomeness, non-suicidal self-injury, and family function.

Conclusions:

• A multivariate, integrative approach significantly enhances suicide risk prediction accuracy in adolescents.
• Perceived burdensomeness, non-suicidal self-injury, and family function are crucial predictors alongside depression and emotion regulation.
• These findings suggest incorporating these variables into future screening frameworks for improved early detection and intervention.