Prediction models for suicide reattempts by lasso regression through machine learning models: Single versus multiple suicide attempters

Abstract

The best predictor of a suicide attempt is a previous attempt, apart from psychiatric diagnoses also associated. Some studies found other indicators of great risk for suicide reattempts. Machine Learning algorithms offer the potential for systematic detection of features that carry greater risk for an event. This study sought to develop a classification algorithm distinguishing between Single Suicide Attempters (SSA) and Multiple Suicide Attempters (MSA) in a Spanish multicentre national cohort to explore prediction of subsequent attempts in suicidal patients. Two models including the same sociodemographic and clinical variables grouped in more specific (Model I) or broad (Model II) categories were developed to explore risk factors for suicide reattempts. A Least Absolute Shrinkage and Regression Operator logistic regression with a 10-fold cross-validation was adopted. 1443 adult patients from the SURVIVE cohort were included (582 SSA and 861 MSA). Both Model I (AUC = 0.696; BAC = 0.644) and Model II (AUC = 0.678; BAC = 0.621) outperformed naïve majority-class classification for SSA and MSA. Bipolar disorder type II, binge-eating disorder, and schizophrenia variables weighted heavier on Model I for suicide reattempt-related; while eating disorder diagnosis, Africa as birthplace, affective disorder diagnosis, being employed, schizophrenia-spectrum disorder and substance use disorder diagnoses were the most important suicide reattempt-related of Model II. Affective disorders, eating disorders and schizophrenia-spectrum disorders emerged as the most important variables in predicting reattempts. Both models showed similar sensitivity and specificity when discriminating between SSA and MSA. Identifying specific risk factors for reattempts could have a significant impact on tailoring prevention strategies and interventions.