12-lead ECG foundation model aligned to echo, chest X-ray, discharge and ECG reports, extending detection to structural heart and systemic disease.
No providers recorded yet. Browse all providers
A low left ventricular ejection fraction is diagnosed on an echocardiogram, not an electrocardiogram, yet the ECG recorded in the same hospital visit carries detectable traces of it. An ECG encoder trained only to match its machine interpretation report never learns those traces, because the report never mentions them: a contrastive objective learns the information two views share, and a rules-based ECG report shares only rhythm, conduction and waveform-morphology statements. The result is ECG representations that excel on arrhythmia benchmarks and can do little better than a randomly initialized encoder on structural heart disease.
MS-ECG-FM, developed at Apple with co-authors affiliated with MIT, Johns Hopkins and Princeton, widens that supervision. Each ECG is contrastively aligned to whichever clinical notes surround it in the medical record: machine and cardiologist ECG reads, echocardiography reports, chest X-ray reports and hospital discharge summaries. Because these notes describe cardiac structure, pulmonary status and systemic disease, the encoder is pushed to encode ECG features that predict them. The work was posted to arXiv in October 2026 by Robert Lewis and colleagues, with Andrew C. Miller and Joseph Futoma as equal-contribution senior authors.
It is a separate model from the Apple AHMS ECG model, which learns from single-lead Apple Watch recordings by self-supervision with no clinical text.
The encoder is a vision-transformer-style ECG model; the reported configuration has 38.28M parameters (tiny, small and base variants range from 5.40M to 86.06M). Pretraining uses the train folds of MIMIC-IV-ECG (720,102 of its 800,035 10-second, 500 Hz recordings from 161,352 patients), linked to 800,034 machine reports, 490,780 cardiologist ECG reads, 81,821 echocardiography reports, 715,465 chest X-ray reports and 331,793 discharge summaries. Echo and X-ray reports are matched to the nearest ECG within seven days; discharge summaries attach to every ECG from the visit. Training runs 75,000 steps with an InfoNCE loss, random lead masking and time-shift augmentation, and checkpoints are selected by an unsupervised effective-rank metric (RankMe) on validation embeddings.
Evaluation uses linear probes on frozen embeddings against MERL, D-BETA, MELP and ECGFounder, the last pretrained on over 10 million ECGs. Across 98 labels from PTB-XL, CPSC2018, Chapman-Shaoxing-Ningbo and EchoNext-Mini, MS-ECG-FM reaches a macro AUROC of 93.6 versus 91.5 for ECGFounder and 91.2 for MELP, with the largest margins on structural heart disease, myocardial infarction and conduction disturbances; arrhythmia performance matches the best baselines. Ablations show that discharge summaries and chest X-ray reports drive gains on structural and exploratory endpoints such as electrolyte imbalance, while echo reports, available for only 34,000 patients, add less.
The model is aimed at AI-enhanced ECG screening: using a cheap, routine test to flag conditions normally confirmed by imaging or laboratory work, including reduced ejection fraction, valvular disease, right heart failure and electrolyte imbalances. Its reduced-lead results point to wearable and bedside monitors that record only Lead I or a few limb leads, though those tests were simulated by dropping leads from clinical 12-lead recordings rather than run on ambulatory devices. Code and weights are not publicly released, so external groups can reproduce the method from the paper and the credentialed MIMIC-IV data but cannot download the encoder.
MS-ECG-FM reframes ECG representation learning around what supervision the encoder sees, not only how many recordings or parameters it has: it beats a model trained on fifteen times more ECGs using one public hospital dataset. Its practical lesson is that the clinical notes already sitting beside each ECG in an electronic health record are an underused training signal. The work is a preprint without peer review, pretrained on a single US hospital system, and its proposed diagnostic-group benchmark has yet to be adopted by other groups.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.