A deep learning (DL) model based on 12-lead ECGs was developed and validated to assess the risk of sudden cardiac death (SCD). The model outperformed a conventional ECG risk model in both internal and external validation datasets. The DL model achieved an area under the receiver operating characteristic (AUROC) of 0.889 in the internal test dataset and 0.820 in the external validation dataset, compared to 0.712 and 0.743 for the conventional model. The DL model demonstrated higher sensitivity and specificity, indicating improved accuracy in distinguishing SCD cases from controls. The model was trained on 1,827 pre-cardiac arrest ECGs from 1,796 SCD cases and validated on 714 ECGs from 714 SCD cases. The model was also compared to a conventional 6-variable ECG risk score, which showed lower performance. The DL model was developed using a convolutional neural network and was optimized for lightweight architecture while maintaining high performance. The model was tested on both internal and external datasets, with results showing good performance in both. The study highlights the potential of DL-based ECG analysis for improving SCD risk stratification. The model's performance was evaluated using logistic regression and other statistical methods, and the results suggest that DL-based ECG analysis could be a valuable tool for early screening and risk assessment. The study also discusses the limitations of the model, including the need for further research to validate its performance in diverse populations and settings. The findings suggest that DL-based ECG analysis has the potential to improve SCD risk assessment and contribute to more effective prevention strategies.A deep learning (DL) model based on 12-lead ECGs was developed and validated to assess the risk of sudden cardiac death (SCD). The model outperformed a conventional ECG risk model in both internal and external validation datasets. The DL model achieved an area under the receiver operating characteristic (AUROC) of 0.889 in the internal test dataset and 0.820 in the external validation dataset, compared to 0.712 and 0.743 for the conventional model. The DL model demonstrated higher sensitivity and specificity, indicating improved accuracy in distinguishing SCD cases from controls. The model was trained on 1,827 pre-cardiac arrest ECGs from 1,796 SCD cases and validated on 714 ECGs from 714 SCD cases. The model was also compared to a conventional 6-variable ECG risk score, which showed lower performance. The DL model was developed using a convolutional neural network and was optimized for lightweight architecture while maintaining high performance. The model was tested on both internal and external datasets, with results showing good performance in both. The study highlights the potential of DL-based ECG analysis for improving SCD risk stratification. The model's performance was evaluated using logistic regression and other statistical methods, and the results suggest that DL-based ECG analysis could be a valuable tool for early screening and risk assessment. The study also discusses the limitations of the model, including the need for further research to validate its performance in diverse populations and settings. The findings suggest that DL-based ECG analysis has the potential to improve SCD risk assessment and contribute to more effective prevention strategies.