A deep learning framework for standardized interpretation of multiparameter cardiac ultrasound and disease classification

Zhihong Wen, Xiangpeng Liu, Yi Liu, Wenbo Tang, Kang An, Shengli Guo

Echocardiographic interpretation underlies a large share of cardiovascular diagnoses, yet specialist expertise remains unevenly distributed, and quantitative measurements show inter-observer variability of 15–17% that contributes to disagreement in borderline cases. We developed DeepCard, a multi-task deep learning system that produces standardized, reproducible interpretation of pre-measured echocardiographic parameters by jointly analyzing 39 quantitative measurements across 17 diagnostic tasks spanning valvular disease, ventricular dysfunction, and structural abnormalities. Trained on 400 patients, DeepCard reached 91% specificity for valvular assessment and 82% accuracy for ventricular evaluation, and reduced inter-observer interpretive variability for pre-measured parameters to 13.4%. On an independent external cohort of 102 patients from a separate institution, performance decreased by only 2.6%, indicating consistent generalization. By providing standardized interpretation of quantitative measurements, DeepCard may help clinicians achieve more consistent and reproducible cardiac assessment, particularly in settings where specialist availability is limited.

Also available via: iScience