Digital profile of children's hearts: automated echocardiogram strain analysis facilitates earlier detection of cardiac dysfunction.

Summary

A semi-supervised, vendor-agnostic deep learning pipeline delivered accurate and generalizable pediatric echocardiographic strain estimates with MAEs near 2% and strong correlations. Automated strain and motion features enabled high AUCs for predicting cardio-oncology dysfunction, late gadolinium enhancement, LVEF decline (outperforming manual strain), and myocardial infarction detection.

Key Findings

• Motion-Echo achieved MAEs of 2.099% (GLS) and 2.665% (GCS) with correlations of 0.799 and 0.781, respectively.
• Automated strain predicted cancer therapy-related cardiac dysfunction with AUC 0.906 and detected LGE with AUC 0.782.
• Forecasting LVEF decline outperformed manual strain (DeLong P < .001); motion flows improved MI detection to AUC 0.952.
• Framework trained on >22,000 echocardiograms across vendors and image qualities using minimal manual annotations.

Clinical Implications

Automated, robust pediatric strain analysis can enable earlier detection of subclinical dysfunction (e.g., cardio-oncology surveillance), support risk stratification, and reduce inter-operator variability across centers.

Limitations

• Retrospective design; prospective impact on clinical decision-making not yet demonstrated
• Generalizability to additional institutions and imaging protocols requires further validation

Future Directions

Prospective, multicenter trials to assess clinical workflow integration, patient outcomes, fairness across subpopulations, and regulatory-grade validation.