Prediction and Prevalence of Self‐Harm and Nonsuicidal Self‐Injury in Children With Learning Disabilities: A Machine‐Learning Approach in Saudi Arabia

The current study aimed to estimate the prevalence of self-harm and nonsuicidal self-injury (NSSI) in children with learning disabilities (LD) in Saudi Arabia and to develop machine-learning (ML) models to identify individuals at elevated risk. In a cross-sectional study, 392 children with DSM-5 specific LD (aged 8-12 years) were recruited through clinical and community channels and assessed for lifetime NSSI and self-harm using structured interviews and self-report. A comprehensive set of sociodemographic, academic, clinical and psychosocial variables was screened using recursive feature elimination, and four supervised ML algorithms (penalized logistic regression, random forests, extreme gradient boosting, and support vector machines) and simple ensembles were trained and evaluated using tenfold cross-validation. Lifetime NSSI was reported by 16.1% of children and self-harm by 9.2%. All ML models showed excellent discrimination for NSSI (AUC up to 0.99), with extreme gradient boosting and majority-voting ensembles achieving the best overall performance. For self-harm, a weighted-average ensemble yielded the most favourable balance of sensitivity and precision (AUC = 0.93). Across outcomes and algorithms, peer victimization/bullying, emotion dysregulation and depressive symptoms emerged as the most robust predictors, whereas LD severity and anxiety symptoms contributed minimally. Self-harm and NSSI are common among Saudi children with LD, and ML models can accurately identify those at highest risk, highlighting bullying and emotion dysregulation as key intervention targets in educational and clinical settings.