Artificial intelligence for detecting acute occluded myocardial infarction: fine-tuning and evaluation of an ECG foundation model
B J Singstad, B Turan, K R Hov, J Ravn, T Sovershaev, T A Aksnes, C H Eek, S Halvorsen, J Munkhaugen, I G Lunde, H Schirmer, A RanjbarAbstract
Background
ST-elevation myocardial infarction (STEMI) patients are usually revascularized within 120 minutes, whereas fewer than one-third of non-ST-elevation myocardial infarction (NSTEMI) occlusions are treated this quickly [1], potentially contributing to worse long-term outcomes [2].
Purpose
We aimed to fine-tune and validate a foundation model to detect occluded myocardial infarction (OMI) from 12-lead ECGs.
Methods
This study included digital ECGs recorded at a tertiary PCI center and affiliated ambulances between 2016 and 2022. ECGs were linked to data from the catheterization laboratory as well as troponin (TnT) measurements. The primary outcome was OMI, defined as an acute culprit lesion, qualitatively determined to be <3 months old, with TIMI flow 0–2 and treated with PCI. To be eligible, ECGs had to be recorded <72 hours prior to the acute invasive procedure. A foundation model was fine-tuned on our dataset using asymmetric focal loss and Muon optimizer. The data were split at the patient level and stratified by outcome into 80% development and 20% test sets, with performance evaluated on the held-out test set.
Results
We included 16 346 ECGs from 10 880 patients. In the test set, 254 (12%) cases were classified as OMI, of which 93 (37%) were NSTEMIs. As shown in the Figure, the fine-tuned OMI model achieved an area under the Receiver Operating Characteristic curve (AUROC) of 0.91 (95% CI, 0.88-0.93) and an Area Under the Precision–Recall Curve (AUPRC) of 0.73 (95% CI, 0.66-0.79) for detecting all OMI cases. For ST-elevation OMI, performance improved to an AUROC 0.98 (95% CI, 0.96-0.99) and an AUPRC 0.87 (95% CI, 0.77-0.94), whereas for OMI without visible ST-elevations the AUROC was 0.85 (95% CI, 0.81-0.89) and the AUPRC 0.34 (95% CI, 0.25-0.47). A prediction threshold of 0.181 was selected to obtain a positive predictive value (PPV) of =0.5 for NSTEMIs. The model's sensitivity, specificity, PPV, and negative predictive value (NPV) for all OMI cases, STEMIs and NSTEMIs are presented in the Table. Of the OMI cases with door-to-balloon time >120 minutes (n=120), the model flagged 56 cases (47%), of whom 88% had NSTEMI, and a median TnT of 1213 (IQR, 104-12373) ng/L. In contrast, the remaining 64 OMI cases (53%) with door-to-balloon time >120 minutes that were not flagged by the model (98 % NSTEMIs), TnT levels were significantly lower (400 ng/L; IQR, 42–4 123; p = 0.05). This difference may partly be explained by a higher prevalence of collateral circulation in the unflagged group (p < 0.05).
Conclusion
The fine-tuned foundation model shows performance within the range of state-of-the-art in detecting OMI from ECGs [3], including NSTEMIs with occlusions. Notably, the model identified almost half of OMIs currently receiving PCI later than 120 minutes upon hospital arrival, underscoring the model’s potential to enhance early diagnosis and timely management of under-recognized, but high-risk OMIs.Figure.ROC & PRC for OMI detectionTable.Model performance at threshold=0.181