Preprint: VisionECG generates a three-dimensional model of the left ventricle from an ECG

Preprint: VisionECG generates a three-dimensional model of the left ventricle from an ECG

In a preprint on medRxiv posted on September 4, researchers described VisionECG, a model trained on 71 132 paired ECG and cardiac MRI examinations from UK Biobank. They then compared its estimates of ejection fraction with echocardiography in 5 000 patients.

An ECG records the heart’s electrical activity, while cardiac MRI shows the shape and movement of its chambers. To link these data, the authors first processed the MRI examinations using DeepMesh. This software converted each examination into a sequence of 50 three-dimensional surfaces representing the left ventricle, the heart’s main pumping chamber. VisionECG was trained to generate the same type of sequence from a standard 12-lead ECG and eight basic characteristics of the individual.

The generated sequence provides a complete picture of ventricular contraction.

It can be used to calculate volume, mass, wall thickness, ejection fraction, and myocardial strain. In 2 000 participants whose data were excluded from training, the median relative error across the full cardiac cycle was 8,34% for volume and 9,54% for wall thickness.

The authors also tested whether each person’s individual ECG contributed to the model’s performance. They replaced each person’s recording with the average ECG for the entire cohort while keeping the other inputs unchanged. In dilated cardiomyopathy, a condition in which the left ventricle enlarges and pumps blood less effectively, the volume estimation error increased from 12,59% to 20,29%. This indicates that the model uses the individual’s pattern of electrical activity in its calculations.

Ejection fraction is the proportion of blood that the left ventricle expels with each contraction. In 5 000 patients with both ECG and echocardiography data, the model distinguished reduced ejection fraction at three thresholds, 50%, 45%, and 35%, with an AUC of 0,79 to 0,81. AUC measures how well a model distinguishes one group of patients from another.

For 17 566 participants with follow-up data, the authors used these features to predict the five-year risk of heart failure. The C-index, which measures how accurately a prediction ranks people by risk, was 0,76, compared with 0,63 for a prediction based on conventional ECG parameters. A single generated sequence of ventricular motion thus provided the basis for both the measurements and the risk prediction.