CMUH's exercise-ECG AI catches a 99% LAD stenosis rated moderate

運動心電圖AI輔助判讀系統 (Exercise ECG AI-Assisted Interpretation System), China Medical University Hospital

A deep-learning model that processes exercise ECG recordings and clinical features.

Exact release not established.

China Medical University Hospital (中國醫藥大學附設醫院人工智慧中心), director 許凱程

The model estimated the probability of coronary narrowing from exercise ECGs and clinical features.

The model combined exercise ECG recordings from three time points with 12 measured or derived clinical features.

The model analyzed exercise ECG signals and clinical features to assess coronary artery disease.

Not assessed

A patient with atypical symptoms has an exercise ECG at a hospital that runs the AI model alongside the manual reading. The model's probability of severe (over 70%) stenosis exceeds what the manual reading and risk score suggest. The cardiologist acts on it by ordering angiography. Angiography confirms a severe stenosis that can be treated at once by stenting.

Not assessed

The model read the exercise-ECG signals and report features. It returned a 72% probability of severe coronary narrowing, above the moderate-risk verdict of the manual reading and risk score. On that basis the cardiologist ordered angiography, which showed the 99% stenosis, and he stented it. The counterfactual rests on the treating cardiologist's statement, carried by three outlets and the hospital's release. He said the manual reading would have led to drug treatment and follow-up only. The hospital that says so also built the model. The dates of the reading and procedure are not reported. The published model's specificity of 0.60 means such flags are often false positives.

The model read the exercise-ECG signals and report features. It returned a 72% probability of severe coronary narrowing, above the moderate-risk verdict of the manual reading and risk score. On that basis the cardiologist ordered angiography, which showed the 99% stenosis, and he stented it. The counterfactual rests on the treating cardiologist's statement, carried by three outlets and the hospital's release. He said the manual reading would have led to drug treatment and follow-up only. The hospital that says so also built the model. The dates of the reading and procedure are not reported. The published model's specificity of 0.60 means such flags are often false positives.

Not assessed

angiography found a 99% LAD stenosis, stented

Improved Physical Health

Not assessed

People.

Not assessed

Not assessed

Not assessed

Not assessed

Yes

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

angiography found a 99% LAD stenosis, stented

Not assessed

angiography found a 99% LAD stenosis, stented

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Developer: China Medical University Hospital (中國醫藥大學附設醫院人工智慧中心), director 許凱程. Operator: China Medical University Hospital, Division of Cardiology (心臟血管系), Taichung.

Benefit: Taichung, Taiwan. Developer: Taiwan. Operator: Taiwan.

Not assessed

Not assessed

Not assessed

Not assessed

The model is run in addition to, not instead of, the clinician's reading. The hospital calls it assisted interpretation and a clinician made every decision. The clinician is willing to send a patient to invasive angiography on an AI probability score. With the published model's specificity of 0.60, some such patients will have no severe lesion. The patient is seen at a hospital that has the system. The sources describe it only at CMUH, in use from May 2022.

Not assessed

17 March 2025. The hospital's cardiology group published the algorithm in Frontiers in Artificial Intelligence in March 2025 with external validation on its own patients after February 2022 and at Asia University Hospital.

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

AIBD-0115-source-0001–0005

Numbered references below.

Not assessed

Independent, Developer

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

Not assessed

People.

People: 3.

People: A man with a 99% stenosis of the left anterior descending artery had it stented after the AI flagged a 72% probability of severe narrowing. The manual exercise-ECG reading had rated it moderate and left it for drug treatment and follow-up. The cardiologist put his chance of an acute infarction at 8 in 10. He said the stent removed the sudden-death risk. He left hospital, symptoms resolved. The hospital's headlines call it life-saving, but the harm was prospective and the account is the developer-hospital's own, so life-saving is not established.

Not assessed

Not assessed

Not assessed

Not assessed

Independently reported

Not assessed

Not assessed

Healthcare: cardiology diagnostics (exercise stress ECG). Hospital cardiology; in-house clinical AI.

The hospital's AI Center built the system to judge from an exercise ECG whether a patient's coronary arteries are severely narrowed (over 70%). Manual reading of a dozen tracings is slow and, the hospital says, only 40–50% accurate. Flagging this patient's severe stenosis was exactly that purpose.

Date basis unassessed: 2022 (precision unassessed). Updated: 22 October 2026.

Not assessed

SHA-256: ada0f6e4bf20…

oecd_aim: 2022-12-28-624f.

Not assessed

Not assessed

6 revisions. Latest: 22 October 2026.