Development and validation of a novel prediction model integrating AI-ECG and 0h high-sensitivity cardiac troponin T for immediate rule-out of acute myocardial infarction

Background/Introduction Early triage of patients presenting to the emergency department (ED) with symptoms suggestive of acute myocardial infarction (AMI) is critical to reduce overcrowding and enable timely treatment. While the current European Society of Cardiology (ESC) 0/1-hour algorithm represents the standard of care, it requires serial high-sensitivity cardiac troponin (hs-cTn) measurements and delays clinical decision-making. Furthermore, these protocols do not formally integrate electrocardiogram (ECG) data beyond identifying ST-segment elevation. A novel deep-learning algorithm was developed to detect AMI solely using a 12-lead ECG. However, whether the formal integration of AI-ECG probabilities with baseline hs-cTn can provide superior diagnostic synergy for immediate (0h) rule-out remains unknown. This study aims to derive and validate a novel strategy for rapid and safe triage upon ED arrival. Methods Data from a prospective, international multicentre study were used to develop a logistic regression model combining raw AI-ECG probabilities from the first ECG at ED presentation with baseline hs-cTnT. Triage thresholds were calibrated to achieve a predefined rule-out sensitivity of 99.7%. External validation was performed in a retrospective, single-centre study including consecutive patients with suspected AMI. The primary outcome was centrally adjudicated AMI by independent cardiologists according to the fourth universal definition. The primary objective was to compare this novel combination versus the ESC-guideline endorsed standard 0h hs-cTnT benchmark (<5 ng/L) approach. Results Among 5,031 and 7,565 patients with an available AI-ECG probability and hs-cTnT measurement in the derivation and validation cohorts, respectively, the median age was 60 years (p25–p75: 47–73) and 58 years (p25–p75: 44–73), and 1,584 (32%) and 2,933 (39%) were female. AMI occurred in 19% (n=964) of patients in the derivation vs 18% (n=1,384) in the validation cohort. In the validation cohort, the novel clinical prediction model combining AI-ECG probabilities and 0h hs-cTnT identified 2,711 patients (35.8%) for immediate rule-out, compared to 2,079 patients (27.5%) using the standard 0h hs-cTnT benchmark (<5 ng/L), representing a 30.2% relative increase in rule-out proportion. The combined model achieved a sensitivity of 99.8% (95% CI: 99.4–100.0%) vs 99.7% (95% CI: 99.3–99.9%) and an NPV of 99.9% (95% CI: 99.7–100.0%) vs 99.8% (95% CI: 99.5–99.9%), missing only 3 vs 4 patients, respectively. Conclusion To our knowledge, this is the first study to derive and validate a strategy combining AI-ECG with 0h hs-cTnT. This approach enables a low-cost, rapid rule-out by identifying significantly more low-risk patients than the standard 0h benchmark, representing a 30.2% relative increase in efficiency while providing numerically superior safety.