A new artificial intelligence model recognized several types of arrhythmias on electrocardiograms with more than 98% accuracy.
An arrhythmia is a heartbeat that is too fast, too slow, or irregular. It may be harmless or indicate a disorder requiring medical care. An electrocardiogram, often abbreviated as ECG, records the heart’s electrical activity using electrodes placed on the skin.
Pixabay illustration
The researchers compared several families of algorithms under the same conditions. Their model combines a graph network, which connects different points in the signal, with a transformer, capable of analyzing relationships across a long sequence. This combination makes it possible to track the shape and temporal evolution of heartbeats simultaneously.
The main tests used the MIT-BIH Arrhythmia database, a collection widely used to test automated systems. The model achieved 98.35% precision, 98.36% recall, and an F1 score of 98.34%. These indicators measure different aspects of classification quality.
Accuracy indicates the overall share of correct predictions. Precision measures the reliability of cases classified in a given category. Recall evaluates the proportion of cases that are actually present and found by the algorithm. The F1 score combines the latter two indicators to limit misleading results caused by an imbalance between categories.
The authors also performed five-fold cross-validation. This method trains the model several times on one portion of the data, then tests it on the remaining portion. An external evaluation was also conducted using the PTB Diagnostic ECG database to verify that the performance did not depend on a single dataset.
The system also incorporates an interpretation method called Grad-CAM. This highlights the areas of the signal that most influenced the decision. A physician can therefore assess whether the algorithm relies on coherent cardiac characteristics rather than on an irrelevant detail unrelated to the arrhythmia.
These results come from computational research and do not include a prospective trial involving patients or use in a clinical setting. Reference databases do not always reproduce the diversity of devices, populations, and recording conditions. Independent tests on hospital data will be necessary before considering diagnostic assistance.