Augmenting an AI-ECG platform with automated STEMI detection: multicenter and prospective external validation

Introduction Rapid and reliable detection of ST-elevation myocardial infarction (STEMI) on 12-lead electrocardiograms (ECGs) is essential for timely reperfusion and improvement of clinical outcomes. Artificial intelligence (AI)–enabled ECG interpretation may support early triage by automating STEMI recognition across heterogeneous clinical settings. Purpose To develop and validate an AI-enabled ECG model for automated STEMI detection and to assess its generalizability across multicenter external cohorts. Methods A retrospective single-center cohort comprising 1,226 patients and their corresponding 12-lead ECGs (34.91% STEMI cases) was used for model development. An AI model was trained to identify STEMI presence (cases) or absence (controls) using standard 12-lead ECG recordings. External validation was conducted on 604 ECGs derived from four independent public datasets (Chapman, Georgia, PTB_XL, and SPH) to assess generalizability across heterogeneous data sources. Additional preliminary evaluation was performed in an interim prospective multicentric cohort comprising 339 patients. Model performance was evaluated using Monte Carlo Cross-Validation (MCCV; 100 iterations) with repeated patient-level random subsampling for robust performance estimation. Performance metrics included the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and F1-score, and were reported as mean values across MCCV iterations with 95% confidence intervals (CI). Results Dataset characteristics are summarized in Table 1, and performance metrics reported in Table 2. In internal validation, the model achieved an AUC of 0.963 (95% CI 0.962–0.965), with sensitivity of 91.96% (95% CI 91.46–92.47) and specificity of 90.72% (95% CI 90.27–91.17). External evaluation demonstrated preserved generalization across heterogeneous public datasets, with AUC values ranging from 0.819 (95% CI 0.815–0.823) to 0.926 (95% CI 0.923–0.929), despite differences in dataset composition and class distribution. Negative predictive value remained consistently high across external cohorts. In the interim prospective multicentric study cohort, the model achieved an AUC of 0.906 (95% CI 0.904–0.907), with specificity of 95.07% (95% CI 94.76–95.37) and sensitivity of 65.12% (95% CI 63.78–66.46). Conclusions This AI-enabled ECG model demonstrated strong discrimination in internal validation and consistent performance across heterogeneous external datasets, including a prospective multicentric cohort. These findings support the integration of automated STEMI detection within AI-ECG platforms for scalable clinical decision support, while highlighting the need for continued optimization in prospective real-world settings. Future research includes model optimization upon study completion to identify occlusive and non-occlusive myocardial infarction (OMI, NOMI) in addition to STEMI.