A conditional flow-matching model called visionECG, trained on 71,132 paired ECG and cardiac mesh sequences from the UK Biobank and validated in 5,000 ECG-echocardiogram pairs, accurately reconstructs spatiotemporal 3D geometry of the left ventricle using only ECG waveforms and basic demographic data. The system enables discrimination of structural abnormalities, visualization of functional deficits, and quantification of both global and regional cardiac parameters — without any imaging hardware.
The implications here are potentially substantial. Cardiac MRI and echocardiography remain inaccessible across much of the world due to cost, equipment, and specialist scarcity. If ECG — a cheap, portable, 10-second test available in virtually every clinical setting globally — can serve as a proxy for structural cardiac assessment, the diagnostic reach expands enormously. This matters acutely for early detection of cardiomyopathies, heart failure phenotyping, and population-level cardiovascular risk stratification in resource-limited settings.
That said, critical caveats apply. The UK Biobank cohort skews toward healthier, predominantly white, middle-aged British adults — limiting generalizability to diverse or sicker populations where cardiac disease is most prevalent. The model reconstructs geometry probabilistically, not deterministically, raising questions about clinical reliability in individual patients. External validation in only 5,000 echocardiogram-paired cases is encouraging but modest for a deployment-grade tool.
As a preprint posted on medRxiv and not yet peer-reviewed, these findings require independent scrutiny before clinical translation. Still, as a proof-of-concept, visionECG represents a genuinely novel reframing of what ECG data can encode — less incremental than paradigm-adjacent.
