Teaching ECG interpretation when AI can already read the strip

Your students can already photograph a rhythm strip and ask an AI what it shows. That changes what an ECG course has to teach, though not in the way it first seems: the reading still has to happen in the student’s head. The AI doing this work comes in three kinds, and a course that treats them as one ends up trusting too much or too little. A deep-learning model trained on the ECG signal has matched cardiologists on rhythm classification since 2019. A general chatbot handed a photo of a strip still loses to cardiologists on routine cases, and its answers change from one session to the next. A third kind of model reads things no human eye can see. Here’s what each one does, with the numbers, and what that leaves for the student to learn.

The first kind is trained on the signal itself. In the Stanford study published in Nature Medicine, a network trained on 91,232 single-lead recordings from 53,549 patients classified 12 rhythm classes with an average F1 score of 0.837, against 0.780 for the cardiologists it was compared with. A year later a Brazilian group trained on more than 2 million 12-lead exams from a telehealth network reported F1 scores above 80% and specificity over 99% for six abnormalities, ahead of the cardiology residents in the comparison. This is the kind of model behind the AI features that ECG equipment vendors advertise.

The second kind reads the printed image instead of the signal. A Yale team trained a network on 385,601 ECGs plotted in different layouts to detect a left ventricular ejection fraction under 40%, and it held an area under the curve of 0.90 to 0.94 across hospitals in four states. That matters for teaching because a photo of a strip is what a student, or a nurse on a unit without digital export, can actually put in front of a model.

The third kind gets the headlines: models that find what no reader can. Trained on 44,959 Mayo Clinic patients, a network detected an ejection fraction of 35% or less from the 12-lead ECG alone, with an area under the curve of 0.93. Another identified patients with atrial fibrillation from ECGs recorded while they were in normal sinus rhythm. Nobody can teach a student to do that by eye, so this branch replaces nothing in a course. It’s screening, and it belongs in the lecture on what the ECG is turning into.

The question students actually ask is about the chatbot in their pocket. A Turkish emergency medicine group has answered it three times with the same 40 cases from the book 150 ECG Cases, 20 routine and 20 harder, against 12 cardiologists and 12 emergency physicians.

The three papers are the first, second and third rounds. Two details in them matter more than the headline. The models’ answers to the same image varied between runs (GPT-4o only reached moderate agreement with itself, Fleiss kappa 0.473), and the whole series rests on 40 textbook cases rather than tracings from a real unit. For a course, that reads as: a chatbot is a second opinion a student can argue with, and a poor authority.

One study has measured AI inside an actual ECG course, and it was in nursing. In 2026, a team ran four models (a nursing-specific tutor, ChatGPT-5, Claude Sonnet 4 and Microsoft Copilot) through an 88-item ECG exam: 85.3%, 83.1%, 80.9% and 79.1% accuracy, with the nursing-specific model making the fewest waveform errors. They then put that tutor into a 4-credit ECG course. 43% of the students used it, and their grades averaged 95.1% against 88.8% for those who didn’t (P = .0048).

Two cautions come with it. Students who choose to use a tutor may be the students who study more, and even the best model in the exam got roughly one item in seven wrong. The authors’ conclusion was that domain-specific AI supports ECG learning. The one that follows for an instructor is that someone still has to hold the answer key.

The practical pattern follows from that. The student reads the tracing first and commits to an interpretation in writing. Then the model reads it, the two answers are compared, and the rubric decides. Disagreement is the teaching moment, and it only exists if the student answered before seeing the AI’s output.

That order is how the clinical interpretation exercise on our AI Platform is built. The educator picks the ECG modality, writes the clinical context and uploads the tracing. The platform reads the image once, for measurements only: rhythm, rate, axis, and which leads carry which change. It’s built not to name a diagnosis at that step; that stays with the educator. From the context and those measurements it drafts one question per dimension (systematic description, relevant findings, diagnostic judgment and the immediate plan) and a rubric with the critical findings, the accepted diagnosis and its synonyms, the plan and the red herrings, all of which the educator edits before publishing. The student gets the strip with a magnifier and the context, writes a free-text answer to each question and submits once. Grading is text against that rubric, never the image again: a weighted score, a penalty for each critical finding missed and each red herring taken, and a report that lists what was missed, why it matters, and an expert reading of the same strip.

Interpretation is the last step of a routine that starts at the bedside, and no model covers the first ones: the four limb electrodes and six precordials placed correctly, a clean acquisition, a strip checked before anyone reads it. The method for reading a strip in four questions is in teaching ECG rhythm interpretation.

Before an AI ECG tool goes into a course, four questions sort the useful ones from the rest.

1. What was it validated on? Signal or image, how many patients, which hospitals, and whether that population looks like the one your students will meet.
2. Does it give the same answer twice? Run the same tracing several times. The chatbot studies above did, and that’s where the variance showed.
3. Where were its errors? The 2026 nursing study counted waveform errors per model. A tool that fails on blocks and a tool that fails on ST changes need different supervision.
4. Does the student answer first? If the tool shows its reading before the student commits, it’s a reference, and references don’t build interpretation.

The Mayo Clinic review of AI-enhanced electrocardiography ends by appraising the field’s limitations and unknowns: models trained on one population, performance outside it, and what a clinician does with a positive screen nobody can see on the strip. The same list works as the last lecture of the course.