bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / biosignals / laef
Biosignals
University of CambridgeIndependent ResearcherReleased August 2026

LAEF

ECG foundation model that reads any subset of the 12 standard leads natively, encoding recordings as variable-size spatiotemporal graphs.

The short version

  • —Reads whatever leads a smartwatch or handheld recorder captured, with no zero-padding
  • —Fine-tune once on 12-lead data, then run at one or two leads with no further adaptation
  • —Encodes an ECG as a spatiotemporal graph that shrinks with the active lead count
  • —Stochastic lead sampling in pretraining builds robustness to lead configuration
  • —Per-lead accuracy holds steady across all 12 standard input leads
17Openness

Where to run it

No providers recorded yet. Browse all providers

Point-of-care cardiac devices — smartwatches, chest patches, handheld recorders — capture one or two ECG leads, while almost every ECG foundation model is built around the clinical 12-lead standard. The usual workaround is zero-padding: flat lines stand in for the missing leads so the input matches the fixed interface a model expects. That conflates a lead that was never recorded with a detached electrode, and it spends compute proportional to the full sensor array however few leads are present. LAEF (Lead-Agnostic ECG Foundation) treats the mismatch as an architectural problem rather than an input-formatting one.

Developed by researchers at the University of Cambridge and the University of Brescia, LAEF represents a recording as a variable-size spatiotemporal graph instead of a fixed-channel tensor. Each lead is partitioned into contiguous temporal segments that become graph nodes; intra-lead edges connect every segment within a lead, and inter-lead edges connect only time-aligned segments across leads. A Graph Attention Network processes whatever graph the available leads produce, so both the node count and the message-passing cost scale with the active lead count.

At 7.1M parameters LAEF is an order of magnitude smaller than the transformer-based ECG foundation models it is benchmarked against, among them ECG-JEPA, ECGFounder, and ECGFM-KED. Its central claim is that a model fine-tuned only on 12-lead recordings can run directly on one or two leads with no lead-specific retraining.

#Key Features

  • Native arbitrary-lead inference: Any subset of the 12 standard leads is processed end-to-end without zero-padding or architectural modification, and graph construction requires no lead identity.
  • Physiologically motivated topology: Intra-lead edges span a full lead to capture wave-to-wave and beat-to-beat structure, while time-aligned inter-lead edges encode that leads are simultaneous projections of one cardiac electrical field.
  • Stochastic lead sampling: Each recording in a minibatch is reduced to a lead subset whose size is drawn uniformly between 1 and 12, acting as a structured information bottleneck that forces representations to survive any configuration.
  • Multi-stage masked node modelling: Masked segments are predicted against a discrete codebook of prototypes; a second stage refits that codebook on the frozen encoder's own latents for finer supervision.
  • Efficiency at the edge: Peak memory is 384 MB, and at one or two leads LAEF records the best throughput and latency among the evaluated models on an NVIDIA A100.

#Technical Details

Pretraining draws on 11 12-lead ECG collections spanning five countries and 9.2M recordings, including CODE, MIMIC-IV, PTB-XL, CPSC, Chapman, Ningbo, and SPH. Signals are band-pass filtered to 0.05–47 Hz, downsampled to 100 Hz, and rescaled to [-1, 1]; each lead is cut into 20 segments embedded by a shared 1D convolutional module. An encoder ablation on a held-out 220k-subject subset favours the Graph Attention Network at 0.937 macro AUROC over a graph isomorphism network at 0.907 and a graph transformer at 0.900, and the spatiotemporal topology over a fully connected one (0.937 versus 0.854).

Evaluation covers 18 downstream datasets scored by macro AUROC after supervised fine-tuning on 12-lead data. At full lead availability no single model dominates and LAEF places in the top three on most datasets despite being up to 12× smaller than the largest baselines. Under a single randomly sampled lead it ranks first on 17 of 18 datasets, with a peak gain of +10.6 AUROC points; with two leads it ranks first on 14 of 18, and the average gain over zero-padded alternatives is +3.2 points. A lead-importance study across 164 cardiovascular conditions finds per-lead macro AUROC stable within 0.003 across all 12 leads while recovering established lead–condition associations, such as elevated importance of leads III and V5 for ischemia and infarction.

#Applications

LAEF is aimed at diagnostic screening on consumer and ambulatory hardware, where recording geometry is fixed by the device rather than by the clinic. A single model can be fine-tuned on abundant 12-lead hospital data and then deployed on a smartwatch or handheld recorder without a separate reduced-lead architecture — useful precisely because labelled point-of-care datasets remain scarce. The small parameter count, low memory footprint, and short fine-tuning time — 30 minutes on the largest evaluation corpus — also suit teams adapting the model to local cohorts on modest hardware.

#Impact

LAEF is a preprint awaiting peer review, and its results are self-reported. Code and checkpoints have not been released; the authors state they will follow acceptance. They also note that the model was evaluated on reduced-lead subsets of clinical 12-lead recordings rather than on signals acquired by point-of-care devices, leaving noise profiles and non-standard electrode placement untested, and that the lead-importance findings do not substitute for clinically established lead-specific protocols. What the work does establish is a concrete alternative to zero-padding for ECG foundation models: making the architecture itself indifferent to how many leads arrive, rather than forcing every recording into a 12-lead interface.

At a glance

Parameters
7.1 Million
Released
August 2026
Category
Biosignals
Organizations
University of Cambridge / Independent Researcher

Links

Research Paper

Tags

ecg_classificationelectrocardiogramfoundation_modelgraph_attention_networkself_supervised

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.