High-performance AI expands electrocardiogram analysis across medical tasks

The university clinic in Innsbruck, Austria has developed an advanced artificial intelligence model, named xECG, that can analyze electrocardiograms (ECGs) with high performance across various medical tasks. Unlike conventional AI models, which are trained for a single, narrowly defined task, xECG is a foundation model that learns to understand the heart signal in all its diversity, much like a language model such as ChatGPT.

This comprehensive understanding allows xECG to apply its knowledge to many medical questions, even those for which only limited specific training data is available. The Digital Medicine in Cardiology research group at the University Clinic, led by Clemens Dlaska, developed xECG in collaboration with cardiologists Axel Bauer and Sebastian Reinstadler.

The xECG model combines a recently proposed architecture (xLSTM) with a training method from computer vision adapted for time-series data like ECGs. It is designed to efficiently process very long signals, such as nighttime recordings for sleep apnea diagnosis, and handles a broad spectrum of conceptually diverse tasks, including classification, regression, and survival prediction.

Trained on approximately 8 million ECGs from about 1.7 million patients, xECG outperforms other models in these diverse tasks. To establish clear scientific criteria for ECG foundation models, the Innsbruck team created the BenchECG evaluation framework, which defines three key requirements: an efficient and comprehensive foundation model must handle a broad spectrum of conceptually diverse tasks, different types of ECG recordings, and function reliably across diverse patient groups.

Using publicly available, highly diverse ECG datasets, the researchers systematically measured how well xECG's representations transfer across tasks under controlled benchmark conditions, finding that it outperforms other models across diverse ECG datasets.