xECG

A sleep apnea study records a single ECG lead for eight hours; an ambulatory Holter tape runs thirty minutes; a resting clinical ECG lasts ten seconds. The ECG encoders that perform best on the ten-second case are vision transformers, and attention costs grow quadratically with sequence length, so in practice they are pinned to a ten-second window and longer recordings have to be chopped into fragments whose relationship the model never sees. The phenomena clinicians care about on those longer traces — apnoeic pauses, rhythm changes recurring across a night — live in the dependencies that the chopping discards.

xECG answers that with a recurrent encoder. It is built on xLSTM, a gated recurrent architecture that restores the linear-in-length cost of an RNN while keeping the parallel training behaviour that made transformers scale, and it is pretrained with SimDINOv2, a non-contrastive teacher-student objective adapted here from computer vision to time series. The Medical University of Innsbruck group behind it released the model in September 2025 alongside BenchECG, a standardised evaluation suite, with the peer-reviewed version appearing in npj Digital Medicine in September 2026.

The release is deliberately a baseline rather than a leaderboard entry. BenchECG assembles eight public datasets — 421,171 patients and 1,674,704 recordings — into ten tasks spanning classification, beat detection, regression, survival analysis and segmentation, and the authors re-evaluate ECG-JEPA, ECGFounder and ST-MEM under the same protocol rather than quoting their published numbers. xECG is the model they submit to their own benchmark.

#Key Features

• Bidirectional xLSTM encoder: Nine alternating sLSTM and mLSTM blocks read the patch sequence in opposing directions layer by layer, giving full-signal context without attention's quadratic cost.
• Two representation granularities: The encoder emits per-patch features that feed beat-level detection and segmentation heads directly, and an attention-pooled vector for signal-level classification, regression and survival modelling.
• Non-contrastive pretraining: SimDINOv2 trains a student to match a slowly updated teacher across augmented views, with a coding-rate regulariser holding features apart — no negative sampling and no large-batch requirement.
• Cross-patient views: Because the pretraining corpora hold multiple ECGs per patient, global and local views are drawn from different recordings of the same person, not only different crops of one trace.
• Raw-signal input: Signals are resampled to 100 Hz and patched, with no normalisation or filtering step, so the encoder learns from whatever the recording actually contains.

#Technical Details

xECG has 57.0M parameters, against 85.2M for ST-MEM and 85.4M for ECG-JEPA. An input ECG is resampled to 100 Hz and split along time only into non-overlapping 250 ms patches, each flattened and linearly projected into the embedding space before entering the encoder stack — a purely temporal patching that produces fewer tokens than the spatio-temporal schemes its transformer baselines use. Pretraining aggregates the CODE corpus (2,322,513 12-lead recordings from 1,676,384 patients), Chapman and Ningbo (45,152 twelve-lead ECGs) and INCART (75 thirty-minute Holter recordings), after filtering out signals containing NaNs, all-zero traces, or variance above 10 combined with amplitude beyond 15 mV. On BenchECG the published version reports the highest score, 0.838, and the best average rank, 1.5 across the five models compared, together with roughly five-fold faster training than the transformer baselines. The margin is widest on the long-context tasks — overnight apnea classification and arrhythmia detection in thirty-minute ambulatory recordings — and widest under linear probing, indicating the advantage is in the frozen representation rather than in what fine-tuning can recover.

#Applications

The pooled representation suits diagnostic classification on twelve-lead ECGs, age and laboratory-value regression, and mortality survival modelling; the patch-level representation suits R-peak detection, per-beat arrhythmia labelling and waveform segmentation. Because the same checkpoint handles reduced-lead and non-ECG cardiac waveforms — the benchmark includes single-lead overnight traces and wearable photoplethysmography — it is a plausible starting point for wearable and remote-monitoring work as well as for hospital twelve-lead pipelines. The repository's disclaimer states that the code and model are released for reproducible research, not for clinical care or commercial use.

#Impact

xECG is the first ECG foundation model to make a recurrent backbone the central claim, and its results are the clearest published evidence that long-context physiological signals reward that choice over a fixed-window transformer. The paired contribution matters as much: by publishing BenchECG with the model and re-running competing checkpoints inside it, the authors made the comparison reproducible in a subfield where each new model had arrived with its own task selection. Both the benchmark code and the pretrained weights are released under the MIT licence. The pretraining corpus is the limiting factor for reproduction — CODE is not openly downloadable and requires contacting its original authors — and every evaluation to date is retrospective and in silico, with no prospective clinical validation.