New ECG Foundation Model Enables Broad Cardiac Diagnosis and Risk Prediction

This limits their usefulness across varied ECG formats and clinical questions, even as cardiovascular care increasingly depends on extracting value from large volumes of routine digital patient data. Now, new findings present an ECG foundation model designed to support multiple medical tasks using heart-signal data.

Medical University of Innsbruck researchers developed the xECG foundation model for electrocardiogram analysis. Unlike single-task models, xECG is designed to learn a broad representation of ECG signals and apply it to different clinical questions, including settings with limited task-specific training data. The model combines the xLSTM architecture, recently proposed by a team at JKU Linz, with a computer-vision training method adapted for the first time to time-series data such as ECGs.

The Innsbruck foundation model was trained on approximately 8 million ECGs from about 1.7 million patients. The team also introduced BenchECG, a freely accessible evaluation framework intended to define clear scientific criteria for ECG foundation models. BenchECG assesses whether models can manage diverse task types, different ECG recording formats ranging from 10-second 12-lead ECGs to long-term ECGs and smartwatch data, and varied patient groups from healthy individuals to high-risk patients.

Using publicly available ECG datasets totaling approximately 1.7 million ECGs from around 400,000 patients, the researchers evaluated how well model representations transferred across tasks under controlled benchmark conditions. The team found that most ECG foundation models evaluated only partly met these criteria, with strong performance in one area but weaknesses in other task domains. The research was published in npj Digital Medicine on September 14, 2026 and is part of a broader program at the Department of Cardiology in Innsbruck, where initial clinical applications of the model are in preparation.

“Our xECG model can also efficiently process very long signals, such as nighttime recordings for sleep apnea diagnosis, an area where many other models reach their limits. The computational cost of our system scales linearly with the signal length, which is what makes it so efficient, especially with very long ECG signals. By handling a broad spectrum of conceptually diverse tasks—such as classification, regression and survival prediction—it currently outperforms other models,” said Clemens Dlaska, leader of the Digital Medicine in Cardiology research group at the University Clinic for Internal Medicine III (Cardiology and Angiology).

Related Links
Medical University of Innsbruck