BenchECG and xECG: a benchmark and baseline for ECG foundation models

Abstract Electrocardiograms (ECGs) are inexpensive, widely used, and well-suited to deep learning. Interest has grown in ECG foundation models that generalise across diverse downstream tasks. However, current evaluation rarely reflects the full capabilities expected of ECG foundation models, with prior work often using narrow task selections and inconsistent datasets, hindering fair comparison. We introduce BenchECG, the first standardised, open benchmark for rigorous ECG foundation model evaluation, comprising eight publicly available datasets (421,171 patients; 1,674,704 recordings). BenchECG assesses model performance across diverse signals, patient populations, and ten conceptually distinct tasks, spanning classification, detection, regression, survival analysis, and segmentation, including explicit out-of-distribution testing. It enables systematic comparison by measuring how well model representations transfer across tasks under controlled benchmark conditions. Alongside BenchECG, we propose xECG, an ECG foundation model combining the efficiency of xLSTM recurrent architecture (5x faster training than transformer baselines) with SimDINOv2 self-supervised learning. xECG achieves the highest BenchECG score (0.838) and best average rank (1.5 out of 5 models), outperforming current publicly available state-of-the-art methods and providing a strong baseline for future ECG foundation models.