By Dr. Oliver Jones
In this week’s Journal of the American College of Cardiology, Pierre Elias, MD, and colleagues presented ValveNet, a novel convolutional neural network for the detection of valvular heart disease (VHD) based on ECGs.
In a retrospective analysis, records from 77,163 patients with ECGs, followed within one year by an echocardiogram, were included. The data were divided, so that the model could be trained, validated, and finally, tested, for the detection of moderate or severe aortic stenosis (AS), aortic regurgitation (AR), or mitral regurgitation (MR), and a composite of all three.
Model performance was evaluated using area under the receiver-operating characteristic (AU-ROC) and precision-recall curves. ValveNet was able to accurately predict AS (AU-ROC = 0.88), AR (AU-ROC = 0.77), MR (AU-ROC = 0.83), and a composite (AU-ROC = 0.84). The authors compared this to “traditional” screening for VHD by physical examination, which has a reported sensitivity of 46% for moderate or greater aortic or mitral valve disease.
The authors suggest this technology could serve as an initial “rule-in” screening test, where patients scoring above a pre-determined threshold could be followed-up with echocardiography, as part of a cost-effective diagnostic cascade. In evaluating this, they noted that the precision, or positive predictive value (PPV), was highly dependent on the prevalence of VHD in testing datasets: at a prevalence of 5% and sensitivity threshold 50%, PPV = 22.8%, whereas at a prevalence of 10% with the same sensitivity threshold, PPV = 36.7%.
Importantly for any “black box” screening tool, the model performed equally well across sex, ethnicity, and race, and in validation using an independent dataset from another medical centre. Notably, the model was more accurate in younger patients, which may reflect the infrequency and hence predictive value of an abnormal ECG in younger patients.
The group is now actively enrolling for a prospective trial in 200 patients using the same ValveNet model to screen ECGs before inviting high predicted risk patients for echocardiography, with a composite primary endpoint of moderate or severe AS, AR, or MR.
Future steps include multi-centre validation in large prospective studies, outcomes-based randomised controlled trials, and experimental integration into existing clinical workflows.
Read more, including the paper and accompanying editorial, at: https://www.jacc.org/doi/10.1016/j.jacc.2022.05.029?utm_medium=email_newsletter&utm_source=jacc&utm_campaign=toc&utm_content=20220801#mmc1
