This study aimed to develop and validate a 12-lead ECG-based deep learning (DL) model for assessing the risk of sudden cardiac death (SCD). The DL model was trained on 1,827 pre-arrest 12-lead ECGs from 1,796 SCD cases in the Portland, Oregon, metro area and validated on 714 ECGs from 714 SCD cases in Ventura County, CA. The model achieved an AUROC of 0.889 (95% CI 0.861–0.917) in the internal held-out test dataset and 0.820 (0.794–0.847) in the external validation dataset. The DL model outperformed a previously validated conventional 6-variable ECG risk model, which had an AUROC of 0.712 (0.668–0.756) in the internal and 0.743 (0.711–0.775) in the external cohort. The DL model was also compared with a logistic regression model, which showed improved performance when combined with clinical variables. The study concluded that the ECG-based DL model provides better accuracy in distinguishing SCD cases from controls compared to conventional ECG-based models, suggesting its potential for improved SCD risk stratification. Further research is needed to validate the model in diverse clinical settings and to explore its broader applications.This study aimed to develop and validate a 12-lead ECG-based deep learning (DL) model for assessing the risk of sudden cardiac death (SCD). The DL model was trained on 1,827 pre-arrest 12-lead ECGs from 1,796 SCD cases in the Portland, Oregon, metro area and validated on 714 ECGs from 714 SCD cases in Ventura County, CA. The model achieved an AUROC of 0.889 (95% CI 0.861–0.917) in the internal held-out test dataset and 0.820 (0.794–0.847) in the external validation dataset. The DL model outperformed a previously validated conventional 6-variable ECG risk model, which had an AUROC of 0.712 (0.668–0.756) in the internal and 0.743 (0.711–0.775) in the external cohort. The DL model was also compared with a logistic regression model, which showed improved performance when combined with clinical variables. The study concluded that the ECG-based DL model provides better accuracy in distinguishing SCD cases from controls compared to conventional ECG-based models, suggesting its potential for improved SCD risk stratification. Further research is needed to validate the model in diverse clinical settings and to explore its broader applications.