ER-JEPA: learning the 12-lead ECG without labels — state of the art on a benchmark, nothing clinical yet

A researcher proposes ER-JEPA, a self-supervised learning model that learns to represent the 12-lead electrocardiogram with no labels at all, by predicting part of the signal from the rest, through a two-stage hierarchical architecture — channels then time — built on a Vision Transformer. Pretrained on about 174,000 10-second traces from Chinese and Brazilian databases, it matches the state of the art on the ST-MEM benchmark and surpasses it on PTB-XL fine-tuning (AUC 0.936 for multi-label, 0.943 for multi-class), all on a single consumer graphics card. But the author acknowledges unstable training, the code is not yet released, evaluation is limited to two public test sets, and no figure is translated into clinical performance.

The context

The electrocardiogram (ECG) is one of the most common tests in medicine: fast, non-invasive, produced by the million each year. The paradox, for anyone wanting to train a model to read it, is that raw traces are abundant while labels — the diagnosis a cardiologist assigns to each trace — are scarce and costly. This is exactly where self-supervised learning (SSL) helps: a model is pretrained on a large mass of unlabelled data by solving a task "invented" from the structure of the signal (for instance, predicting a masked portion), then fine-tuned on a small labelled set for the clinical task of interest.

Two families compete here. Masked autoencoders reconstruct the missing raw signal sample by sample; the reference work for ECG, ST-MEM, belongs to this line. JEPAs (Joint-Embedding Predictive Architectures), originally proposed for images (I-JEPA), make the opposite bet: predict not the missing signal but its abstract representation, which spares the model from wasting capacity reproducing noise. The paper here ports that idea to the ECG.

The method

The model is called ER-JEPA (Event Reconstruction Joint-Embedding Predictive Architecture). Its distinctive feature is stacking two JEPAs in cascade, hence the name hierarchical JEPA (H-JEPA): a first module learns a representation across the twelve leads (the twelve viewpoints of the heart a standard ECG provides) within each short time window; a second then treats the sequence of these representations as a time series. The backbone is a Vision Transformer (ViT), the attention-based architecture that splits the input into fragments and weighs their relations. As in I-JEPA, a "target encoder" is updated by an exponential moving average of the weights, and prediction happens in representation space.

Pretraining draws on two large public 12-lead databases: Chapman-Shaoxing (45,152 recordings, collected at Shaoxing People's Hospital and Ningbo First Hospital, China) and a subsample of CODE-15 (from the Brazilian CODE database, Telehealth Network of Minas Gerais). After restricting to 10-second traces and removing incomplete recordings, 174,461 traces remain (43,561 Chinese and 130,900 Brazilian), resampled to 250 Hz. Training takes about 30 hours on a single NVIDIA RTX 3090 (24 GB), over 300 epochs (full passes over the data).

Evaluation follows the ST-MEM benchmark methodology, on two public test sets: PTB-XL (a German database from the Physikalisch-Technische Bundesanstalt) and CPSC2018 (a Chinese database). Two protocols are used: linear probing (freezing the pretrained encoder and training only a linear classifier on top — a test of the raw quality of the representations) and fine-tuning (retraining the whole model on the task). The tasks are multi-label classification (several possible abnormalities per trace) and multi-class (one category among several).

The results

On the ST-MEM benchmark, ER-JEPA matches the state of the art of transformer-based self-supervised models on PTB-XL and CPSC2018. It surpasses it in one specific configuration: fine-tuning on PTB-XL, with an AUC of 0.936 for multi-label and 0.943 for multi-class. The AUC (area under the ROC curve) measures the ability to correctly rank a positive case ahead of a negative one: 0.936 means that, drawing at random one abnormal and one normal trace from the test set, the model assigns the higher score to the right one about 94 times out of 100. The other claimed contribution is efficiency: at comparable performance, the model markedly reduces memory and inference time relative to competing multichannel ViTs.

Clinical translation: there is none, and this needs stating plainly. The paper reports neither sensitivity nor specificity at a decision threshold, no predictive value, no count of false positives and false negatives in a given population, no calibration curve. An AUC of 0.936 on a classification benchmark does not tell you how many infarctions or arrhythmias would be missed or over-detected in practice: it places the model in a methodological ranking, not in a cardiology ward. It is an evaluation of the quality of the learned representations, not a measure of clinical utility.

What is good

A real problem, a coherent approach. Exploiting the mass of unlabelled ECGs through self-supervision answers a genuine constraint of the field. And the two-stage structure — first across the twelve leads, then across time — mirrors the very structure of the ECG, where information lives both in the comparison between leads and in the evolution of the trace. Pretraining moreover crosses two very different populations (China and Brazil) and evaluation two others (Germany and China), placing the model up against some geographic diversity from the outset.

Light, frugal, accessible. Thirty hours on a single consumer RTX 3090, a marked reduction in memory and inference time: the opposite of foundation models that demand GPU clusters. This frugality makes the method reproducible by a modest lab and deployable on ordinary hardware. The licence is CC BY 4.0, permissive, allowing reuse including commercial — the opposite of non-commercial licences that block any real-world use.

A rare honesty about the method's fragility. The author does not hide that stacking two JEPAs makes the model prone to representation collapse (the model cheats by squashing all outputs toward a constant) and that performance varies strongly with batch size, dataset size and embedding dimension. He documents that about 1% of 500 pretraining runs give anomalous results. Reporting one's own instabilities, rather than showing only the best run, is a healthy scientific reflex.

What is less good

A beaten benchmark is not clinical use. This is a misleading metric by omission: the whole evaluation rests on two public test sets (PTB-XL, CPSC2018) and on classification AUCs, often on imbalanced multi-label tasks where a high AUC can coexist with many false positives in absolute terms. No calibration, no operating threshold, no prospective or even retrospective validation under real clinical conditions. The paper mostly claims to "match" the state of the art; the surpassing is confined to a single configuration (PTB-XL fine-tuning).

An acknowledged instability, and code still missing. The sensitivity to representation collapse is not a mere curiosity: it means the result depends on delicate settings and that a practitioner rerunning the training could land on one of the failed runs. Yet at publication the code is announced as "coming soon" but not provided, and this is a non-peer-reviewed preprint by a single author. Without published code or weights, reproducibility — the very argument for frugality — rests on the author's word.

A generalization asserted but not dissected. This is a population bias blind spot: despite the geographic spread, no analysis by sex, age, acquisition device or pathology subtype is presented. The test sets are public benchmarks that share preprocessing conventions, which eases transfer without proving robustness against a genuinely unseen hospital, ECG machine or population. So we do not know where the model breaks down, or for whom.

What it changes

For the research community, the contribution is clear and circumscribed: a lightweight self-supervision recipe for multivariate time series, a reusable idea (explicitly separating the channel axis from the time axis), and an honest flagging of an instability others will meet. The only credible judge of what comes next will be the release of the code and an evaluation that steps outside the benchmark: sensitivity, specificity and calibration on an external, ideally unseen, clinical cohort.

For clinicians, there is nothing to deploy today: a research tool, benchmark-evaluated, unvalidated, unapproved, still fragile in behaviour. For patients and the general public, the nuance is the same as for all ECG "foundation models": in time, learning to read the ECG without depending on a stock of expert-annotated labels could make automated analysis cheaper and more available — but this is upstream research, not a product, and the headline AUC describes a benchmark ranking, not diagnostic performance at the bedside.

To go further

The preprint is available on arXiv (2607.01145). On automated ECG reading in clinical conditions, see our decryptage of a deep-learning model applied to the ECG image in the emergency department; on arrhythmia prediction from a wearable ECG, that of a personalised atrial-fibrillation prediction model; and on what the representations of self-supervised models are really worth once probed, that of a probing of foundation models in oncology.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.