Predicting suicide attempts in adolescents with longitudinal clinical data and machine learning

Area of Science:

• Computational psychiatry
• Adolescent health
• Machine learning in healthcare

Background:

• Adolescents exhibit high rates of nonfatal suicide attempts, posing a significant clinical challenge for risk prediction.
• Current screening methods are often time-consuming and difficult to implement at scale.
• Machine learning offers a potential solution for predicting suicide risk using routinely collected clinical data.

Purpose of the Study:

• To develop and validate a machine learning approach for predicting suicide risk in adolescents.
• To assess the performance of computational algorithms in identifying adolescents at risk for nonfatal suicide attempts.
• To evaluate the scalability and clinical utility of machine learning for suicide risk screening.

Main Methods:

• Retrospective, longitudinal cohort study using data from January 1998 to December 2015.
• Inclusion of 974 adolescents with nonfatal suicide attempts and multiple control groups.
• Utilized random forests algorithm with candidate predictors including diagnostic, demographic, medication, and socioeconomic factors.

Main Results:

• Computational models demonstrated strong performance in predicting suicide risk, with improved accuracy as attempts became more imminent.
• Discrimination metrics (AUC) were consistently high across comparisons with other self-injury, depressed, and general hospital controls.
• Random forests significantly outperformed logistic regression, and recalibration substantially improved prediction performance.

Conclusions:

• Machine learning applied to longitudinal clinical data presents a scalable method for enhancing suicide attempt risk screening in adolescents.
• This approach may overcome limitations of traditional screening methods, enabling broader identification of at-risk individuals.
• Validated computational models can aid clinicians in making more informed decisions regarding adolescent suicide prevention.