Predicting Suicidal Ideation Among Youths With Autism Spectrum Disorder: An Advanced Machine Learning Study

Abstract

This study aimed to predict suicidal ideation among youth with autism spectrum disorder (ASD) by applying machine learning techniques. A cross-sectional sample of 368 ASD-diagnosed young people (aged 18-24 years) was recruited, and 34 candidate predictors-including sociodemographic characteristics, psychiatric symptoms (e.g., anxiety problems and depressive symptoms), behavioural measures (e.g., bullying victimization and insomnia severity) and adverse childhood experiences-were assessed using standardized instruments and parent-report checklists. After listwise deletion of missing data, recursive feature elimination (RFE) with a random forest wrapper was performed to identify the five most influential predictors. Four classification algorithms (logistic regression, random forest, eXtreme Gradient Boosting [XGBoost] and support vector machine [SVM]) were then trained on a 70/30 stratified split and evaluated on the hold-out test set using area under the curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value and accuracy. RFE identified anxiety problems, insomnia, bullying victimization, age and depression (PHQ-9) as the top predictors. Logistic regression achieved an AUC of 0.943 (sensitivity = 0.773, specificity = 0.957 and accuracy = 0.922), random forest an AUC of 0.948 (sensitivity = 0.727, specificity = 0.989 and accuracy = 0.939), XGBoost an AUC of 0.930 (sensitivity = 0.772, specificity = 0.989 and accuracy = 0.947) and SVM an AUC of 0.942 (sensitivity = 0.772, specificity = 0.978 and accuracy = 0.939). Across models, anxiety and insomnia emerged as the two most important risk factors, and XGBoost demonstrated the best overall balance of performance metrics, yielding the highest accuracy. Gradient-boosted tree models were thus shown to effectively integrate multidimensional data to predict suicidality in autistic youth, highlighting anxiety and sleep disturbances as critical targets for personalized risk assessment and prevention efforts.

Keywords: anxiety; depression; learning disability; machine learning; suicide.