External validation of an artificial intelligence-based electrocardiogram algorithm for early rule-out of acute myocardial infarction in two large European cohorts

Introduction Early identification of patients with acute myocardial infarction (AMI) remains challenging in the emergency department (ED), as current diagnostic pathways rely mainly on serial high-sensitivity cardiac troponin (hs-cTn) measurements and may delay decision-making. Artificial intelligence–based electrocardiogram (AI-ECG) models may enable rapid, non-invasive risk stratification at presentation. A novel AI-ECG model, AiTiA MI v2.0.0, was developed in a South Korean population to perform the early triage of patients with chest pain. Its performance in large European ED cohorts remains unknown. Methods In this secondary analysis of two prospective cohorts, adult patients presenting with symptoms suggestive of acute myocardial infarction (AMI) were enrolled in cohort A, an international multicenter study, and in cohort B a single-center study. Cohort A required written informed consent for participation, whereas Cohort B consecutively enrolled all eligible patients. All cases underwent central adjudication of the final diagnosis. Patients were excluded if a 12-lead ECG at presentation was unavailable or deemed invalid by the AI-ECG. As part of routine care, hs-cTnT was measured at presentation in all patients. Prespecified cut-offs from the original model were applied for initial rule-out and rule-in classification, followed by the baseline thresholds of the European Society of Cardiology (ESC) 0/1-hour algorithm following a theoretical ED workflow integration. AMI was defined as a composite of type 1 and type 2 MI. Findings Among 5,031 and 7,565 patients with an available AiTiA MI v2.0.0 probability and hs-cTnT measurement in cohort A and cohort B, respectively, the median age was 60 years (p25–p75: 47–73) and 58 years (p25–p75: 44–73), and 1584 (32%) and 2933 (39%) were female. AMI occurred in 964 patients (19%; 769 type 1 MI and 195 type 2 MI) and in 1384 patients (18%; 1,125 type 1 MI and 257 type 2 MI, 2 missing). When predefined thresholds were applied in cohort A (Figure 1) AiTiA MI v2.0.0 ruled out 706 patients (14%), achieving a sensitivity of 97.6% (95% CI: 96.6–98.5%) and a negative predictive value (NPV) of 96.7% (95% CI: 95.3–97.9%). In cohort B (Figure 2), the model ruled out 1,437 patients (17.5%), with a sensitivity of 97.5% (95% CI: 96.7–98.3%) and an NPV of 97.6% (95% CI: 96.8–98.3%). Among the 23 false-negative patients in cohort A, 9 had type 1 MI. Among the 34 false-negative patients in cohort B, 14 had type 1 MI. All had hs-cTnT superior to 5ng/l. Interpretation When validated in two prospective European cohorts recruiting consecutive adult patients presenting to the ED with symptoms suggestive of AMI, the raw thresholds performance of AiTiA MI v2.0.0 had good performance, comparable to its development. The number of false negatives could be explained from a distribution shift in AMI adjudication, with a higher ratio of type1 MI/type 2 MI in the European cohorts.