The AI ECG Is a Second Reader, Not a Rule-Out Test

At 02:11, a patient with chest pressure has an ECG completed at triage. The automated interpretation is quiet. The AI-ECG score is low. The patient still looks unwell, and the pain has not stopped. The next decision is not whether the algorithm is impressive. It is whether the team lets a reassuring number end the evaluation.

That is the safety problem with AI-assisted ECG interpretation. A model can find patterns that deserve attention, including patterns that do not meet classic ST-elevation criteria. Emergency physicians should use AI-ECG as a second reader that changes what we look at next, not as a rule-out test that makes bedside assessment optional.

The signal is useful, but the threshold is not the diagnosis

The best recent evidence is encouraging without supporting full automation. In a 2023 international evaluation of an AI model for occlusion myocardial infarction (OMI), the model was tested on geographically distinct European and United States cohorts. At its selected threshold, it reported 80.6% sensitivity and 93.7% specificity for OMI, with an area under the curve of 0.938. Sensitivity was lower in patients without ST elevation on the index ECG and in patients with broad QRS complexes (International evaluation of an artificial intelligence-powered electrocardiogram model detecting acute coronary occlusion myocardial infarction).

A prospective multicentre study gives a complementary picture. In 8,493 adults with suspected acute myocardial infarction across 18 Korean emergency departments, the ROMIAE study found an area under the curve of 0.878 for acute myocardial infarction. At its prespecified low-risk threshold, sensitivity was 99.6% and negative predictive value was 99.1%, but only 8.2% of the cohort entered that low-risk group. Emergency physicians were blinded to the AI output, so the study measured model performance, not the effect of placing the model inside a live clinician workflow (Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study).

Those numbers matter because a threshold is a decision instrument, not a diagnosis. A low score may support a low-risk pathway when the rest of the evaluation agrees. It cannot erase ongoing symptoms, a concerning examination, a changed tracing, an abnormal troponin pattern, or a clinician's concern.

Keep the first ten minutes human-led

The 2021 AHA/ACC chest-pain guideline recommends acquiring and reviewing an ECG for STEMI within 10 minutes of arrival. It also states that an initial normal ECG does not exclude acute coronary syndrome and recommends repeat ECGs when symptoms continue until other testing rules out the diagnosis (2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain). AI must fit inside that process. It must not delay the first ECG, replace clinician review, or become the reason a repeat tracing is skipped.

A safe sequence is simple. Obtain a usable ECG. Read the raw tracing against the symptoms, examination, and prior ECGs. Review the AI output as a prompt for a second look. Then decide on escalation, serial testing, imaging, consultation, or another pathway. The word “second” matters: the AI should not be the only reader, and its score should not automatically win a disagreement.

When the patient and the score disagree, that disagreement is the clinical signal. Document what the model said, what you saw, and why the plan followed the patient rather than the number.

Design the workflow around false reassurance

The most dangerous output may be a low-risk label that arrives while the team is looking for permission to stop. A 2025 study of AI detection in ECGs called normal by conventional computer algorithms found that, in a highly selected case series of 42 OMI patients, AI identified OMI in 29 of 37 initial ECGs that the conventional algorithm had called normal. The study could not estimate the AI false-positive rate and was not a representative ED sample. It is a warning about a blind spot, not proof that every normal ECG should trigger an invasive pathway (Artificial Intelligence Detection of Occlusive Myocardial Infarction from Electrocardiograms Interpreted as “Normal” by Conventional Algorithms).

Build a hard stop for discordance. If symptoms persist, the clinical story is high risk, the tracing has a new or subtle change, the AI output is high risk, or ECG quality is poor, the low-risk label cannot close the case. Route the patient to physician review and the local chest-pain pathway. Repeat the ECG when the story is still active. Use serial biomarkers and the rest of the clinical decision pathway rather than treating an isolated AI output as a disposition order (2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain).

Show the time of the ECG, model version, confidence band, signal-quality status, and intended use. Do not hide limits behind a single color. If performance is weaker in broad QRS complexes or tracings without ST elevation, make those conditions easy to recognize (International evaluation of an artificial intelligence-powered electrocardiogram model detecting acute coronary occlusion myocardial infarction).

Make deployment a clinical program, not an install

Before go-live, define the question the AI is allowed to answer. “Find OMI patterns that merit a second look” is bounded. “Clear chest pain” is a disposition claim and needs a higher bar. The 2025 Canadian consensus work on AI-based clinical decision support in emergency medicine recommends choosing a relevant problem and expert team, setting data standards, using AI-specific reporting, and addressing ethics and privacy (Establishing methodological standards for the development of artificial intelligence-based Clinical Decision Support in emergency medicine).

Test the full workflow. Measure time to human review and notification, repeat-ECG rates, false-positive burden, missed events, and performance across the patients your ED sees. Assign an owner, escalation route, fallback, and review date. For a case-based rehearsal of subtle ECG findings and human override, EM-Sim offers branching emergency-medicine scenarios.

Monitor after launch. Track poor-quality inputs, alert rates, score distributions, model updates, and cases where clinicians report that the output did not fit the patient. Chet Shermer's discussion of silent AI failure makes the operational point clearly: a system can keep returning plausible numbers after its data feed, patient mix, or workflow changes, so a green dashboard is not proof of safe performance (When an AI Model Fails Silently: Detection and Response Protocols for Emergency Physicians). Rehearse a case in which the AI is unavailable, conflicts with the ECG, or arrives too late to help.

Dr. Chet's Take

I have read thousands of ECGs in emergency departments, on HEMS transports, and in the Guard. The hard cases are not always dramatic. They are the patient whose tracing looks almost normal, the patient whose symptoms do not fit the first label, and the patient whose physiology is changing while the team waits for the next data point. This article gets the main point right: AI-ECG can widen our attention, but it does not own the patient. A score is a prompt. The patient is the problem. I want the emergency physician to see the tracing, hear the story, and decide what the next safe step is.

That being said, I do not want us to swing from enthusiasm to reflexive rejection. The international OMI work shows why a second reader may find patterns that classic STEMI rules miss. The ROMIAE results show that a carefully selected low-risk threshold can perform well in a defined study population. The honest answer is that those results do not tell me what will happen in my department until we measure the local workflow. I need to know when the output arrives, who sees it, what happens when it is wrong, and whether the team still repeats the ECG when the patient says the pain is getting worse. If we cannot answer those questions, we are adding another unowned voice to a crowded room.

Key Takeaways

• Use AI-ECG to find patterns that deserve a second look, not to declare a patient safe for discharge.
• Keep the AHA/ACC chest-pain process intact: rapid ECG review, clinical assessment, repeat ECGs when symptoms continue, and serial testing when indicated (2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain).
• Treat disagreement between the patient, tracing, and model as a reason to escalate human review.
• Before go-live, define intended use, local measures, owner, fallback, and review date (Establishing methodological standards for the development of artificial intelligence-based Clinical Decision Support in emergency medicine).

FAQ

Can an AI ECG rule out a heart attack?

Not by itself. The ROMIAE study reported strong low-risk performance at a prespecified threshold in its study population, but the model was one part of an evaluation and was not tested as a replacement for emergency physician assessment (Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study).

What should I do when the AI ECG disagrees with my read?

Review the raw tracing, symptoms, examination, prior ECGs, and the chest-pain pathway. Persistent symptoms or a concerning picture should drive repeat ECGs and further evaluation rather than automatic acceptance of either interpretation (2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain).

Can AI detect an occlusion when the ECG does not meet STEMI criteria?

Some models have detected OMI patterns in research cohorts that classic STEMI criteria did not identify. Performance varies by population and ECG features, and the international evaluation reported lower sensitivity in patients without ST elevation and in those with broad QRS complexes (International evaluation of an artificial intelligence-powered electrocardiogram model detecting acute coronary occlusion myocardial infarction).

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download it here: AI in EM Survival Guide.

Sources

• Gulati M, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain
• Kim KH, et al. Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study
• van der Sluijs et al. International evaluation of an artificial intelligence-powered electrocardiogram model detecting acute coronary occlusion myocardial infarction
• Smith SW, et al. Artificial Intelligence Detection of Occlusive Myocardial Infarction from Electrocardiograms Interpreted as “Normal” by Conventional Algorithms
• Kareemi H, et al. Establishing methodological standards for the development of artificial intelligence-based Clinical Decision Support in emergency medicine
• Chester Shermer, MD, FACEP. When an AI Model Fails Silently: Detection and Response Protocols for Emergency Physicians
• Related simulation training: EM-Sim
• Books by Dr. Shermer