A highlight of today’s full programme at the Digital Health Stage is a session on AI interpretation of ECGs. In one of presentations, Bjørn-Jostein Singstad (Akershus University Hospital - AHUS - Nordbyhagen, Norway) describes fine-tuning of an AI model to detect occlusion MI (OMI) from 12-lead ECGs.

This study included 16,346 digital ECGs from 10,880 patients recorded at a tertiary PCI centre and affiliated ambulances. ECGs were linked to cathlab data and troponin (TnT) measurements. The primary outcome was detection of OMI, defined as an acute culprit lesion with TIMI flow 0–2 and treated with PCI. An ECG foundation model was fine-tuned on the dataset using asymmetric focal loss and Muon optimisation. Patient-level data were stratified by outcome into 80% development and 20% test sets, with performance evaluated on the test set.

In total, 12% of cases in the test set were classified as OMIs, of which 37% were NSTEMIs. The fine-tuned model achieved an area under the receiver operating characteristic curve (AUROC) of 0.91 and an area under the precision–recall curve (AUPRC) of 0.73 for detecting all OMI cases. AUROC and AUPRC were 0.98 and 0.87 for ST-elevation OMI and 0.85 and 0.34, respectively, for OMI without visible ST-elevations.

Of the cases with door-to-balloon time of more than 120 minutes, the model flagged 47% of cases as OMI, of which 88% had NSTEMI and the median TnT was 1,213 ng/L (IQR 104–12,373). The remaining cases (53%) not flagged by the model were almost all NSTEMIs (98%), with significantly lower TnT levels (400 ng/L; IQR 42–4,123; p=0.05). The lower myocardial injury biomarker levels in the unflagged group may partly be explained by a higher prevalence of collateral circulation (p<0.05).

A 2-year prospective follow-up study is ongoing to further evaluate the diagnostic performance of the algorithm. If the algorithm again demonstrates promising results, an optimal threshold for direct referral to acute invasive angiography will be established. A randomised controlled trial will be conducted to assess the ability of the algorithm’s output from ambulance or emergency department ECGs to guide management of patients presenting with acute chest pain.