Diagnostic performance of an artificial intelligence-based electrocardiogram interpretation system for acute coronary occlusion in adults following out-of-hospital cardiac arrest

Background Rapid identification of acute coronary occlusion (ACO) following out-of-hospital cardiac arrest (OHCA) is essential to guide early coronary angiography and revascularisation. In OHCA patients without clear ST-elevation, early angiography has not consistently demonstrated a survival benefit compared with a delayed strategy. Interpretation of the post-arrest electrocardiogram (ECG) is difficult, and conventional STEMI criteria can fail to identify true ACO in patients who may still benefit from timely intervention. Purpose To compare the diagnostic accuracy of an artificial intelligence (AI)-based ECG interpretation system with clinician assessment using standard STEMI criteria for detecting angiographically confirmed ACO in adults following OHCA. Methods A retrospective analysis was performed of consecutive adults (≥18 years) presenting after OHCA of presumed cardiac origin to a regional cardiac arrest centre between January 2019 and July 2020. Patients were excluded if there was a clearly non-cardiac cause of arrest, severe pre-existing disability, or estimated life expectancy <6 months. Post-arrest 12-lead ECGs, coronary angiography findings, clinical characteristics and laboratory data were collected from electronic health records. ECGs were independently assessed by experienced clinicians using standard STEMI criteria and by an AI-based ECG interpretation system that generated a continuous STEMI probability score. Each method was evaluated against the reference standard of angiographically confirmed acute culprit vessel occlusion, defined as Thrombolysis in Myocardial Infarction (TIMI) 0–1 coronary flow. Results Of 219 screened patients, 154 had an analysable post-arrest ECG and underwent coronary angiography suitable for comparison. Among these, 66/154 (42%) had an acute culprit vessel occlusion. The AI-based ECG system achieved an area under the receiver operating characteristic curve (AUC) of 0.843 for detection of ACO. At an optimal decision threshold of 0.48, sensitivity was 82% and specificity was 75%. Clinician ECG interpretation using STEMI criteria showed comparable performance, with sensitivity 76% and specificity 81%. 66 ACOs confirmed at angiography, 16 (24%) did not have ST elevation on ECG, of these the AI successfully identified 11 cases of ACO, demonstrating potential incremental diagnostic yield in non-classical or borderline presentations following OHCA. Conclusion(s) An artificial intelligence-based ECG interpretation system demonstrated diagnostic performance comparable to experienced clinicians for identifying acute coronary occlusion in adults following out-of-hospital cardiac arrest. The system also detected cases of acute coronary occlusion that were not recognised as STEMI by clinicians, suggesting a possible role in triage and early invasive decision-making in this high-risk population when conventional post-arrest ECG criteria are inconclusive. Prospective validation is warranted.