Loading…
Conditional flow matching model that reconstructs 4D left-ventricular geometry and motion from a 12-lead ECG and basic demographic variables.
No providers recorded yet. Browse all providers
A cardiac MR scan costs a magnet, a radiographer and a booked slot; a 12-lead ECG costs ten electrodes and ten seconds. Both interrogate the same myocardium, yet they are read for almost opposite things — the scan for chamber volume, wall thickness and contraction, the ECG largely for conduction abnormalities. Deep learning has narrowed that gap by training ECG models to emit a scalar or a binary label. What those models never return is the object the clinician reasons about: a heart with a shape and a beat.
visionECG, from Imperial College London and collaborating UK groups, reconstructs that object instead. It is a conditional latent flow matching framework mapping a resting 12-lead ECG plus eight demographic variables onto a 50-frame left-ventricular surface mesh sequence, 1,412 vertices per frame in fixed correspondence across frames and subjects. Two components do the work: visionECG-Flow synthesises motion latents conditioned on the encoded ECG, and visionECG-Motion fuses those latents into the resolved 4D mesh.
Because every generated sequence shares one mesh topology, the output is a reusable substrate rather than a prediction. Volumetric, geometric, strain-based and segmental measurements are computed from the same synthetic heart on demand, including quantities that were never training targets — adding a phenotype means writing a new measurement, not retraining a model.
Training runs in three stages. Two variational autoencoders are pretrained independently — a dilated convolutional VAE over the 12-lead waveform with a 1,024-dimensional latent, and a cardiac motion sequence VAE coupling mesh graph convolutions with a temporal Transformer into a 512-dimensional latent, optimised with Chamfer, Laplacian smoothing and KL terms. visionECG-Flow is then trained in the frozen motion latent space under a latent rectified flow matching loss, using a conditional U-Net of eight symmetric residual blocks with hidden width 1,024 and adaptive layer normalisation for conditioning. Inference integrates 100 Euler steps and averages the terminal latents of 100 sampled trajectories. Finally, visionECG-Motion is adapted on flow-generated latents with the flow held fixed.
The cohort is 71,132 UK Biobank participants with paired ECG and cardiac mesh sequences, split 56,906 / 14,226 at the participant level; the reference meshes come from DeepMesh, a multi-view cine CMR motion-tracking pipeline. Median absolute relative error across the sequence is 8.34% for endocardial volume and 9.54% for wall thickness, with median 90th percentile Hausdorff distance of 3.21 mm (endocardial) and average symmetric surface distance of 2.06 mm. Replacing each participant's waveform with a population-average ECG raised median volume error from 7.64% to 10.33%, and further in disease groups. Derived phenotypes reached AUC at or above 0.80 on six threshold-defined abnormality tasks and 0.83 for ICD-coded cardiac disease, 0.79–0.81 against echocardiographic ejection fraction in the external cohort, and a Harrell's C-index of 0.76 for incident heart failure against 0.63 for conventional ECG parameters.
The target setting is one where an ECG is available and imaging is not: pre-referral triage in primary care, opportunistic screening, and longitudinal tracking of cardiac remodelling from serial recordings. Because the output is a measured heart rather than a risk score, a clinician can inspect where the abnormality sits: hypertrophic cardiomyopathy and hypertension both raise ventricular mass, but separate on septal-to-lateral ratio and longitudinal strain.
visionECG is a preprint and has not been peer reviewed. Its code is released under the MIT licence, while the training corpus is UK Biobank and reaches researchers through that resource's own application process. The authors are explicit about the boundaries: training and primary validation sit in a volunteer cohort of predominantly European ancestry, the reconstruction covers the left ventricle only and so says nothing about valves, tissue characterisation or the remaining chambers, residual geometric error is larger at end-systole, and localised features are underestimated relative to global ones. What the work establishes is a template others can follow — treating a structured dynamic representation, rather than a task-specific label, as the primary output of a biosignal model.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.