An ECG AI more than doubled detection of occlusion MI, and still missed half
In short
Among 24,511 chest-pain patients in Swedish emergency departments (mean age 59, 52% male), 467 (1.9%) had occlusion myocardial infarction. An ECG AI model reached 52% sensitivity against 23% for conventional STEMI criteria, with specificity essentially tied at 99% versus 98% and positive predictive value 51% versus 17%. But 52% sensitivity means close to half of true occlusions were still missed. A clean ECG is not the same as ruling out a heart attack.
Start with the terms. The STEMI criteria long used in emergency departments ask whether the ST segment has risen by a set amount. The problem is that a coronary artery can be genuinely occluded without meeting that threshold. Hence occlusion myocardial infarction (OMI), a definition aimed at the blockage itself rather than at a waveform threshold. This analysis took Swedish emergency department data from 2017–2018, confirmed occlusion by angiography and expert adjudication, then ran the same ECGs through several reading methods.
How did the numbers split?
- The AI model — 52% sensitivity (47–57), 99% specificity, 51% positive predictive value, 99% negative predictive value.
- Conventional STEMI criteria — 23% sensitivity (19–27), 98% specificity, 17% positive predictive value, 98% negative predictive value.
- The Glasgow automated algorithm — 32% sensitivity, 98% specificity, 26% positive predictive value.
- Extended criteria (extra leads, modified Sgarbossa for left bundle branch block, a rule for left ventricular hypertrophy) — sensitivity rose to 41% while specificity fell to 95% and positive predictive value to 14%.
The most practical gap is 51% versus 17% positive predictive value. Under the old criteria, more than four of every five positives were false alarms; with the AI, half of the positives were real occlusions. The extended criteria teach the opposite lesson: loosening a threshold to gain sensitivity buys false positives at the same time. Three percentage points of specificity, across an emergency department that size, translate into hundreds of unnecessary activations.
What does this mean for a lifter?
Two things. First, a normal ECG is not an exclusion. Even the best reading method here missed close to half of the occlusions, which is why chest pain during or right after training should never end at "the ECG looked fine." Second, ECGs are especially unsteady in people with thick heart muscle. That the extended criteria needed a separate rule for left ventricular hypertrophy — ST elevation in V1–V3 of at least 0.25 of the R and S waves — is itself the evidence. We have covered how often an ECG misses left ventricular hypertrophy and how weakly that finding predicts; this result says the same limitation carries into acute care.
Chest pain during exercise, fainting, or breathlessness unlike your usual can be an emergency. Do not wait it out, and take from this study that a normal ECG does not rule out an infarction.
Frequently asked questions
How does occlusion MI differ from STEMI?
STEMI is defined by ST-segment elevation crossing a set threshold on the ECG; occlusion MI is defined by an artery actually being blocked. The definitions diverge because some patients have a genuine occlusion without meeting the elevation threshold.
How well did the AI ECG model perform?
Among 24,511 chest-pain patients, 467 had occlusion MI. The AI model reached 52% sensitivity, 99% specificity, and 51% positive predictive value; conventional STEMI criteria on the same data reached 23% sensitivity and 17% positive predictive value.
Is 52% sensitivity good enough?
No. It means close to half of true occlusions are missed, so an ECG result alone cannot rule out a heart attack. Symptoms and blood tests still carry the decision.
Would looser criteria catch more?
They catch more and misfire more. In this study the extended criteria raised sensitivity to 41% while specificity fell to 95% and positive predictive value to 14%.
Is an ECG harder to read in someone with a large heart?
Left ventricular hypertrophy changes the waveform and makes interpretation less reliable. The extended criteria in this study included a dedicated rule for exactly that case.
Source: PubMed