Foundation Models for ECG: Promise, Gaps, and What Comes Next

Large pretrained models have reshaped NLP and computer vision. The natural question is whether the same paradigm — pretrain on huge unlabeled corpora, fine-tune on small labeled datasets — translates to clinical time-series. For ECG specifically, the stakes are high: cardiac events kill more people than any other cause, and most of the world has no cardiologist nearby.

The short answer is: partially. And the gaps are instructive.

Why ECG is a good testbed

ECG is well-suited for this experiment for a few reasons:

1. Scale exists. Datasets like MIMIC-IV, PhysioNet, and hospital systems have millions of recordings. This is enough to pretrain on.
2. The signal is structured. Unlike free-form clinical notes, an ECG is a fixed-length, fixed-channel time-series. Tokenization choices are more constrained.
3. Labels are expensive but finite. Cardiology has a well-defined label set (rhythm disorders, conduction abnormalities, ischemia markers). This makes downstream evaluation tractable.

What works

Self-supervised pretraining on large ECG corpora using masked reconstruction (analogous to BERT’s masked token prediction) does generalize. Models pretrained on one hospital’s data and fine-tuned on another’s outperform task-specific models trained from scratch — even with 50–100x fewer labeled examples in the target domain.

The representation quality is particularly good for rhythm-based tasks: atrial fibrillation, supraventricular tachycardia, and heart block. These have clear temporal patterns that contrastive or masked objectives capture well.

Where it breaks down

Morphology is harder. Subtle ST-elevation or T-wave changes that indicate ischemia are sensitive to electrode placement, patient anatomy, and recording quality. A model trained on hospital A’s 12-lead acquisitions can fail badly on hospital B’s if their lead placement protocol differs by even a few centimeters. Pretraining helps but doesn’t fully close this gap.

Distribution shift is brutal. ECG quality varies enormously: motion artifact, poor electrode contact, patient demographics. Pretraining on clean hospital recordings and deploying on a wearable patch is a different problem. The model has seen the domain but not the noise distribution.

Labels from reports are noisy. Most large ECG datasets are labeled by mining clinical reports, not by cardiologist review of each tracing. Label noise at scale creates a ceiling on what pretraining can learn. This is not a problem unique to ECG — it’s the core challenge of clinical NLP in general.

The real bottleneck: evaluation

The hardest problem isn’t modeling — it’s knowing when the model is good enough to trust. Cardiac AI failures are not like ImageNet failures. A false negative on an MI-equivalent pattern isn’t just a wrong prediction; it’s a patient sent home who shouldn’t have been.

Current evaluation frameworks borrow too directly from ML benchmarks. We need:

• Prospective validation (not just held-out test sets from the same acquisition pipeline)
• Subgroup analysis by age, sex, comorbidities, and equipment type
• Calibration as a first-class metric, not an afterthought

What I’m watching

A few directions that look promising:

• Multi-modal pretraining (ECG + clinical notes + vitals) is underexplored. The signal from combining modalities should be much stronger than ECG alone.
• Personalised adaptation — using a patient’s own historical ECGs as context — could address the morphology variability problem. This is a retrieval problem as much as a modeling problem.
• Uncertainty quantification at inference time. If the model knows it doesn’t know, that’s clinically useful. Most deployed cardiac AI systems give you a score; almost none give you a calibrated confidence.

Foundation models for ECG aren’t a solved problem, but they’re a real one. The next few years will be determined less by architecture choices and more by whether the research community takes validation seriously enough.

I work on ECG and medical imaging AI at Tricog Health. The views here are my own.