Multi-centre study of an artificial intelligence ECG model for the detection of occlusive myocardial infarction within the UK

Background Pre-hospital triage strategies for ST-segment elevation myocardial infarction (STEMI) vary widely. Standard practices, such as autonomous paramedic activation or remote cardiology consultation, face evolving competition as artificial intelligence electrocardiogram interpretation develops. Evaluating this technology across distinctly different clinical pathways is essential to determine its real-world utility. Purpose This retrospective validation study assessed whether an artificial intelligence model [1] can accurately detect occlusive myocardial infarction (OMI) compared to both standard paramedic and cardiology triage. The primary aim was to determine if the technology could reduce false-positive pathway and emergency catheterisation laboratory activations without loss of diagnostic sensitivity. Methods The study retrospectively analysed 295 suspected STEMI cases collected consecutively in cohorts between 2024 – 2025 from two large regional heart attack centres in the UK. Centre A (177 patients) operated without pre-hospital cardiology review, relying on paramedic activation. Centre B (118 patients) used a nurse-led cardiology remote electronic electrocardiogram and clinical history review model. We compared the final clinical ground-truth diagnosis against the standard care decisions and the artificial intelligence predictions. True OMI was defined as Thrombolysis in Myocardial Infarction flow grade 0 to II, or grade III flow with a confirmed culprit vessel and troponin elevation. Blinded independent adjudication was performed of all clinical notes and angiography. Results Triage relying solely on paramedic activation resulted in high proportion of false-positive alerts (58% of total activations). In Centre A, applying the artificial intelligence electrocardiogram model yielded 94.7% sensitivity, 89.2% specificity, and 91.5% overall accuracy. This reduced false-positive activations (6% of total activations) without compromising patient safety. In Centre B, the existing cardiology remote review gatekeeper achieved 91.4% sensitivity and 91.6% specificity. The artificial intelligence model demonstrated comparable performance in this setting, achieving 94.3% sensitivity, 89.2% specificity, and 90.7% accuracy. Across both triage pathways, the artificial intelligence maintained a consistently low false-negative rate (5.5%). Conclusion The OMI artificial intelligence electrocardiogram model demonstrates broad clinical utility across diverse pre-hospital triage workflows. The model significantly improves specificity and overall diagnostic accuracy in paramedic-led pathways, safely reducing unnecessary laboratory activations. This matches the high diagnostic performance of a cardiology remote review pathway, but with opportunities for more effective resource utilisation. These findings suggest that artificial intelligence can effectively optimise STEMI pathway triage in a variety of underlying workflow structures.