Development and external validation of an AI-enabled 12-lead electrocardiogram model integrating age and sex for acute aortic dissection

Abstract Background Acute aortic dissection (AD) is a life-threatening cause of chest pain that can be difficult to recognise. Artificial intelligence (AI) analysis of routinely acquired electrocardiogram (ECG) may provide additional diagnostic information. We developed an AI-ECG model incorporating age and sex and evaluated its discrimination and operating characteristics in internal and external chest pain cohorts. Methods This retrospective two-centre study included 27,160 patients from Zhongshan Hospital, Fudan University, divided at the patient level into development, tuning, and internal test sets. External evaluation included 844 patients from the Qingpu Branch. The full model was compared with ECG-only and age-and-sex models. Classification thresholds were selected in the internal tuning set and applied unchanged to both evaluation cohorts. Calibration, alert burden, and performance in an aortic computed tomography angiography (CTA)-verified subset were also assessed. Results The internal test set included 5,433 patients with 30 AD events, and the external cohort included 844 patients with 31 AD events. The full model achieved areas under the receiver operating characteristic curve (AUCs) of 0.803 (95% confidence interval [CI] 0.716–0.880) internally and 0.857 (0.778–0.925) externally. Corresponding AUCs were 0.798 and 0.842 for the ECG-only model and 0.670 and 0.553 for the age-and-sex model. At the tuning-set Youden threshold, sensitivity was 63.3% internally and 77.4% externally, specificity was 81.0% and 76.9%, and alert rates were 19.3% and 25.1%, respectively. In 94 internal test patients with aortic CTA verification, the AUC was 0.781 (0.671–0.877). Conclusions ECG waveforms provided information for AD discrimination beyond age and sex in two retrospective chest pain cohorts. The observed sensitivity and alert burden do not support stand-alone diagnostic use. Prospective evaluation is needed to determine whether model-guided clinical reassessment improves diagnostic decisions and patient safety.