Construction and validation of a machine learning-based prediction model for sudden cardiac death in chinese children with hypertrophic cardiomyopathy: a multicenter retrospective study

Background Hypertrophic cardiomyopathy (HCM) is a leading cause of sudden cardiac death (SCD) in children. While several risk prediction models exist, they were primarily developed in Western populations and may not fully capture risk factors relevant to Chinese pediatric HCM patients. We therefore aimed to develop and validate a machine learning-based model specifically for predicting SCD risk in Chinese children with HCM. Methods In this multicenter, retrospective cohort study, we enrolled 714 children with HCM from 10 tertiary centers in China between January 2001 and September 2025. The primary outcome was a composite endpoint of SCD events including SCD and aborted SCD (defined as successfully resuscitated cardiac arrest or appropriate implantable cardioverter-defibrillator therapy). Candidate predictors were selected using LASSO regression and the Boruta algorithm. Eight machine learning survival models were employed to construct both a clinical-only model and an integrated clinical-genetic model. Model performance was evaluated in two independent validation cohorts and assessed by the area under the receiver operating characteristic curve (AUC), concordance index (C-index), Brier score, calibration curves, and decision curve analysis (DCA). Results The cohort comprised 714 children with a median age of 5 years (IQR,0.6-10)(66.8% male) with 2816.26 patient-years of follow-up, 146 patients (20.4%) experienced a SCD event. Among 157 patients with detectable (likely) pathogenic variants, the MYH7 gene was most frequently implicated (45.2%; 71/157). Risk predictors included sex, unexplained syncope, specific electrocardiogram abnormalities, Z-scores of left ventricular posterior wall thickness and left atrial diameter, and left ventricular ejection fraction. Both the clinical model and the integrated clinical-genetic model demonstrated excellent discrimination. The clinical model achieved a C-index of 0.82 in the training cohort and 0.79 in validation cohort 1, and 0.78 in validation cohort 2. The integrated clinical-genetic model achieved a C-index of 0.83, 0.81, and 0.79 in the respective cohorts. Both models showed satisfactory calibration in internal and external validation, confirmed by calibration curves and Brier scores, and demonstrated potential clinical utility via DCA. To facilitate clinical application, an online risk prediction tool was developed based on the final models. Conclusion We developed and independently validated a clinically applicable machine learning model demonstrating excellent predictive performance for SCD risk in Chinese children with HCM. This tool, which incorporates clinical and genetic data, enables individualized SCD risk stratification and informs shared decision-making regarding primary prevention implantable cardioverter-defibrillator (ICD) implantation.