Cardiovascular

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

Sharma S, Martinez Di Pietro B, et al. • Echocardiography (Mount Kisco, N.Y.) • 2026

TL;DR

In this proof-of-concept study, a 3D deep learning model applied to routine echocardiographic videos identified features predictive of future cardiomyopathy in pediatric cancer survivors, performing well on internal hold-out data but near chance on single-study external discrimination, though it still stratified external patients into high- and low-risk groups for CM; further refinement and validation are needed.

Key Findings

Results

The deep learning model discriminated CM-positive from CM-negative echocardiograms well on the internal hold-out set (Center 1).

AUROC was 0.87±0.05 on the internal hold-out set.
AUPRC was 0.74±0.07 on the internal hold-out set.
Models were 3D deep convolutional neural networks trained on parasternal short-axis echocardiographic clips from echocardiograms conducted up to five years before CM diagnosis.
The cohort comprised cancer survivors ≤21 years old at cancer diagnosis.

Results

Single-study discrimination on external validation cohorts was near chance.

Center 2: AUROC 0.55±0.02 and AUPRC 0.22±0.02.
Center 3: AUROC 0.60±0.03 and AUPRC 0.21±0.03.
This was a substantial drop compared with internal hold-out performance (AUROC 0.87, AUPRC 0.74).

Results

Despite poor external single-study discrimination, the model's risk stratification identified external-cohort patients at significantly higher risk of future CM.

Patients were stratified into high- and low-risk groups for developing CM.
Hazard ratio for CM in high- vs low-risk groups was 2.78 (95% CI = 1.52-5.10, p = 0.00056).
Analysis used cumulative incidence and proportional hazards regression in the external cohort.

Conclusions

The authors conclude deep learning could detect subtle echocardiographic markers years before CM diagnosis and might support risk-guided surveillance, but needs further refinement and validation.

The study is described as proof-of-concept.
CM is often identified only after irreversible cardiac remodeling has occurred.
The authors describe the approach as a preliminary step toward risk-guided CM surveillance.

What This Means

Childhood cancer survivors are at risk of heart muscle disease (cardiomyopathy) later in life, often because of cancer treatment. It is frequently detected only after heart damage has become irreversible. This study tested whether artificial intelligence applied to routine heart ultrasound videos could spot subtle signs of future cardiomyopathy up to five years before diagnosis. The AI model performed well on data from the same hospital where it was trained (AUROC 0.87), but when tested at two other hospitals its ability to correctly classify individual studies was close to chance (AUROC 0.55 and 0.60). However, when the researchers used the model to sort patients at the external centers into higher- and lower-risk groups and followed them over time, the high-risk group was about 2.8 times as likely to develop cardiomyopathy. This research suggests that AI analysis of routine echocardiograms might eventually help identify survivors who need closer heart monitoring. However, the weak performance at other centers shows the model is not ready for clinical use and needs more refinement and validation across different settings.

Have a question about this study?

Citation

Sharma S, Martinez Di Pietro B, Nandula A, Goplani T, Chen H, Armenian S, et al.. (2026). Early Prediction of Cardiomyopathy in Pediatric Cancer Survivors Using Echocardiographic Video-Based Deep Learning.. Echocardiography (Mount Kisco, N.Y.). https://doi.org/10.1111/echo.70643

Key Findings

What This Means

Have a question about this study?

Related Research

Citation