AI ECG analysis is an AI model that reads an ordinary 12-lead ECG and finds signs of heart disease that a doctor reading the same tracing would not see. It is also called ECG-AI. Some of these models have been cleared by regulators as medical devices in their own right, which is called software as a medical device (SaMD). The first ones were cleared in the United States, where they are mainly used.
A doctor reads an ECG by eye, looking for known patterns, such as the shape that a heart attack leaves. The model reads it as numbers. A 10-second, 12-lead ECG recorded 500 times a second is 60,000 numbers. The model was trained on hundreds of thousands of ECGs. Each ECG was paired with what later tests found about that patient's heart. From these pairs, the model learnt small patterns in the tracing that no person had described.
One use is to find a weak heart pump. The ejection fraction is the share of the blood in the heart's main pumping chamber that the heart pushes out with each beat. A normal ejection fraction is above about 50%. Below 40%, the heart is weak. It is usually measured by an echocardiogram, an ultrasound scan of the heart, which takes a specialist's time and costs more. Many people with a weak heart have no symptoms and a normal-looking ECG.
Here is a typical case, at a family clinic in Minnesota, in the United States. A 64-year-old woman has a routine ECG before a minor operation. Her doctor reads it as normal. The model reads the same 60,000 numbers and gives a score of 0.83 for a low ejection fraction. This is above the alert level of 0.5 that the clinic has set. So the clinic's software shows a flag beside her ECG. The doctor orders an echocardiogram, which measures her ejection fraction at 35%. She starts treatment before any symptoms appear.
The model does not diagnose by itself. It flags who should have the next test, and a doctor decides. Similar models are being built for longer recordings from Holter and ambulatory monitors, which a patient wears for a day or more. In India, remote ECG interpretation services use AI in a related way, to read each ECG first and sort cases by urgency through triage. Each new version of a model can change what it flags, so each version must be tested and cleared again before it is used.