Skip to content
Patrick M. Boyle's CardSS Lab | UW Bioengineering

New paper in Echocardiography: deep learning on routine echo videos to predict cardiomyopathy in childhood cancer survivors

We’re pleased to share a new paper led by Surbhi Sharma, in collaboration with Lindsay Edwards (Duke) and colleagues across Seattle Children’s, Fred Hutch, and other children’s hospitals: “Early prediction of cardiomyopathy in pediatric cancer survivors using echocardiographic video-based deep learning” (also available on PubMed Central).

Cardiomyopathy (CM) is a leading cause of late death in childhood cancer survivors, but it is often caught only after irreversible cardiac remodeling. We asked whether a 3D convolutional neural network, applied to routine parasternal short-axis echo clips, could pick up subtle signs of CM up to five years before diagnosis.

Key findings:

  • On an internal hold-out set from one center, the model reached an AUROC of 0.87 and an AUPRC of 0.74.
  • Performance on two external centers was much weaker for single-study classification (AUROC 0.55 and 0.60).
  • Even so, model-defined high-risk survivors in the external cohort were more likely to develop CM over time than low-risk survivors (hazard ratio 2.78, 95% CI 1.52 to 5.10).

This is a proof-of-concept study, and the drop in external performance shows how much work remains on generalizability across institutions. We see it as a first step toward risk-guided surveillance that complements, rather than replaces, current care.

This work was supported by the NIH, the Rally Foundation, the Catherine Holmes Wilkins Charitable Foundation, the Unravel Pediatric Cancer Foundation, the Seattle Foundation, and the Seattle Children’s Research Institute.

Flow diagram with four boxes connected by arrows: Study Cohort, Data Preprocessing, Model Training, and Inference and Analysis. Study Cohort (at-risk pediatric cancer survivors): the development cohort from Center 1 has 104 cardiomyopathy-positive (CM+) echocardiograms from 28 subjects and 292 CM-negative echocardiograms from 46 subjects; the external validation cohort from Centers 2 and 3 has 146 CM+ echocardiograms from 39 subjects and 673 CM-negative echocardiograms from 114 subjects. Data Preprocessing: parasternal short-axis clips are imported from DICOM, frame-masked, downsampled and temporally subsampled to a 224 by 224 by 16 input, and histogram matched. Model Training: 3D CNN convolution layers extract spatiotemporal features, followed by dense layers and a softmax output giving a yes or no prediction, with patient-based train, calibration, and hold-out splits and 5-fold cross-validation. Inference and Analysis: Grad-CAM heatmaps for interpretability, plus external validation and survival analysis, shown as an ROC curve above the diagonal, a Grad-CAM heatmap of a short-axis echo frame, and a cumulative incidence plot in which the high-risk group (red) rises above the low-risk group (blue) over roughly 2,000 days.