Early Prediction of Cardiomyopathy in Pediatric Cancer Survivors Using Echocardiographic Video-Based Deep Learning

Cardiomyopathy (CM) is a leading cause of late mortality in childhood cancer survivors, but it is often identified only after irreversible cardiac remodeling has already occurred. In this proof-of-concept study, we assessed whether deep learning applied to routine echocardiographic videos can identify features predictive of future CM.

We analyzed parasternal short-axis clips from cancer survivors ≤21 years old at cancer diagnosis. 3D deep convolutional neural networks were trained to classify CM-positive versus CM-negative studies from echocardiograms conducted up to five years prior to CM diagnosis. The performance was assessed on internal hold-out validation (Center 1) and external validation cohorts (Centers 2, 3). The model's prognostic value was evaluated by stratifying patients into high- and low-risk groups for developing CM, and probability of future CM was assessed using cumulative incidence and proportional hazards regression analysis.

The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.87±0.05 and an area under the precision-recall curve (AUPRC) of 0.74±0.07 on the internal hold-out set. External validation showed near-chance single-study discrimination (AUROC 0.55±0.02 and 0.60±0.03; AUPRC 0.22±0.02 and 0.21±0.03 for Centers 2 and 3, respectively). Nevertheless, analysis of the external cohort showed significantly higher probability of CM in high- versus low-risk groups (hazard ratio 2.78, 95% CI = 1.52-5.10, p = 0.00056).

Deep learning models have the potential to identify subtle echocardiographic markers years before CM diagnosis in pediatric cancer survivors and could serve as a preliminary step toward risk-guided CM surveillance. Further refinement and validation are needed to determine the utility of deep learning for early identification of pathologic cardiac remodeling in this population.