Artificial intelligence–driven temporal clinical and ECG modeling for personalized sudden cardiac death risk prediction in hypertrophic cardiomyopathy

Background Hypertrophic cardiomyopathy (HCM) is one of the leading causes of sudden cardiac death (SCD), particularly in young individuals. Current SCD risk stratification relies on established clinical and imaging markers. However, the identification of patients who most likely benefit from an implantable cardioverter defibrillator (ICD) is still challenging. Existing strategies are largely based on static variables, whereas arrhythmic risk in HCM is dynamic, and its temporal evolution remains insufficiently explored. Purpose To develop and validate a multimodal artificial intelligence (AI) model capable of capturing temporal clinical data and integrating them with electrocardiographic changes to enhance the prediction of SCD risk in patients with HCM. Methods We included 118 patients (mean age 51 ± 17 years, 25% women) with ≥3 longitudinal evaluations. Three standardized timepoints (T0, T1, T2) were defined to reflect disease evolution (first visit, intermediate visit, and last visit before SCD or last available evaluation for those without SCD). At each timepoint, synchronized clinical and echocardiographic variables corresponding to the ESC HCM Risk-SCD score were collected (Figure 1). ECG images were preprocessed and organized into temporal patient-specific sequences. Three models were developed:AI-Clinical model: feed-forward neural network using tabular variables.AI-ECG temporal model: convolutional neural network with recurrent layers.AI-Multimodal temporal model: fusion of clinical and ECG latent features processed through a Transformer encoder to capture temporal and cross-modal interactions. Data were split at the patient level (70% training, 30% testing). Performance was assessed using AUC-ROC, accuracy, precision, and recall. Interpretability was explored with Grad-CAM. Results During a follow-up of 8 ± 5 years, 16 (14%) patients experienced SCD. The static AI-Clinical model demonstrated limited discriminative performance (AUC: 0.63; Accuracy: 0.81; Recall: 0; Precision: 0). The temporal AI-ECG model improved prediction (AUC: 0.79; Accuracy: 0.83; Recall: 0.71; Precision: 0.56). The multimodal temporal Transformer model achieved the best overall performance (AUC 0.89; Accuracy 0.86; Recall 0.86; Precision: 0.60), outperforming unimodal approaches. In patients with SCD, predicted probabilities progressively increased from T0 to T2, suggesting dynamic ECG remodeling preceding events. Grad-CAM analysis demonstrated shifting model attention across ECG regions over time, supporting evolving electrophysiological substrates prior to SCD (Figures 1 and 2). Conclusions Longitudinal modeling of ECG evolution substantially improves SCD risk prediction in HCM compared with static models based solely on ESC clinical risk variables. Multimodal temporal AI may enable personalized, time-aware risk stratification and warrants prospective external validation.Figure 1Figure 2