Deep learning electrocardiogram analysis for rapid rule-out of acute coronary syndrome in the emergency department: a multicenter diagnostic accuracy study.

BACKGROUND: Rapid and accurate exclusion of acute coronary syndrome (ACS) in patients presenting with chest pain remains a major clinical challenge. Deep learning models applied to the 12‑lead electrocardiogram (ECG) may improve diagnostic efficiency, yet multicenter validation data remain limited. METHODS: We conducted a prospective, multicenter diagnostic accuracy study enrolling 6743 consecutive adults presenting to five emergency departments with suspected ACS between January 2020 and December 2023. Patients with ST-elevation myocardial infarction were excluded. A deep learning ECG (DL-ECG) model based on a residual convolutional neural network was developed using nested five-fold stratified cross-validation. Model performance was compared against emergency physician clinical assessment (Standard Care) and the HEART score using the 30-day adjudicated ACS diagnosis as the reference standard. RESULTS: Among 6743 patients (mean age 61.9 years; 59.8% male), 1099 (16.3%) received an ACS diagnosis within 30 days. The DL-ECG model achieved an area under the receiver operating characteristic curve (AUC) of 0.954 (95% CI, 0.948-0.960), significantly exceeding Standard Care (0.921; P < 0.001) and the HEART score (0.876; P < 0.001). At matched sensitivity of 97.5%, the DL-ECG model demonstrated superior specificity (0.746 vs. 0.568 and 0.354) and negative predictive value (0.994 vs. 0.992 and 0.987). Performance was consistent across demographic subgroups and study centers. CONCLUSIONS: A deep learning model analyzing the standard 12‑lead ECG significantly outperformed both clinical assessment and the HEART score for ruling out ACS. If validated prospectively, this approach could facilitate earlier discharge of low-risk patients.