Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust AI-ECG for Predicting Left Ventricular Systolic Dysfunction in Pediatric Congenital Heart Disease

Yuting Yang, Lorenzo Peracchio, Joshua Mayourian, John K. Triedman, Timothy Miller, William G. La Cava

TL;DR

AI-ECG for LVSD screening in pediatric CHD is impeded by limited data and privacy constraints. The authors propose a robust, architecture-agnostic framework combining on-manifold adversarial perturbations with uncertainty-aware adversarial training to focus learning on near-boundary, uncertain samples, with latent-space perturbations constrained by a pre-trained autoencoder (ViT-MAE). On a real-world pediatric dataset, the approach delivers robust LVSD detection and shows substantial gains under data-scarce conditions (e.g., 10% of data), especially for challenging subgroups such as pacemakers, as confirmed by comprehensive ablations. The method’s generality and performance uplift suggest practical deployment in data-limited hospitals and potential extension to multi-modal cardiac tasks to advance democratization of AI in pediatric cardiology.

Abstract

Artificial intelligence-enhanced electrocardiogram (AI-ECG) has shown promise as an inexpensive, ubiquitous, and non-invasive screening tool to detect left ventricular systolic dysfunction in pediatric congenital heart disease. However, current approaches rely heavily on large-scale labeled datasets, which poses a major obstacle to the democratization of AI in hospitals where only limited pediatric ECG data are available. In this work, we propose a robust training framework to improve AI-ECG performance under low-resource conditions. Specifically, we introduce an on-manifold adversarial perturbation strategy for pediatric ECGs to generate synthetic noise samples that better reflect real-world signal variations. Building on this, we develop an uncertainty-aware adversarial training algorithm that is architecture-agnostic and enhances model robustness. Evaluation on the real-world pediatric dataset demonstrates that our method enables low-cost and reliable detection of left ventricular systolic dysfunction, highlighting its potential for deployment in resource-limited clinical settings.

Robust AI-ECG for Predicting Left Ventricular Systolic Dysfunction in Pediatric Congenital Heart Disease

TL;DR

AI-ECG for LVSD screening in pediatric CHD is impeded by limited data and privacy constraints. The authors propose a robust, architecture-agnostic framework combining on-manifold adversarial perturbations with uncertainty-aware adversarial training to focus learning on near-boundary, uncertain samples, with latent-space perturbations constrained by a pre-trained autoencoder (ViT-MAE). On a real-world pediatric dataset, the approach delivers robust LVSD detection and shows substantial gains under data-scarce conditions (e.g., 10% of data), especially for challenging subgroups such as pacemakers, as confirmed by comprehensive ablations. The method’s generality and performance uplift suggest practical deployment in data-limited hospitals and potential extension to multi-modal cardiac tasks to advance democratization of AI in pediatric cardiology.

Abstract

Artificial intelligence-enhanced electrocardiogram (AI-ECG) has shown promise as an inexpensive, ubiquitous, and non-invasive screening tool to detect left ventricular systolic dysfunction in pediatric congenital heart disease. However, current approaches rely heavily on large-scale labeled datasets, which poses a major obstacle to the democratization of AI in hospitals where only limited pediatric ECG data are available. In this work, we propose a robust training framework to improve AI-ECG performance under low-resource conditions. Specifically, we introduce an on-manifold adversarial perturbation strategy for pediatric ECGs to generate synthetic noise samples that better reflect real-world signal variations. Building on this, we develop an uncertainty-aware adversarial training algorithm that is architecture-agnostic and enhances model robustness. Evaluation on the real-world pediatric dataset demonstrates that our method enables low-cost and reliable detection of left ventricular systolic dysfunction, highlighting its potential for deployment in resource-limited clinical settings.

Paper Structure

This paper contains 18 sections, 11 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Methods
  4. On-Manifold Adversarial Example Generation
  5. Uncertainty-Aware Adversarial Training
  6. Experiments
  7. Dataset
  8. Outcomes
  9. Baselines
  10. ResNet
  11. ResNet+DA
  12. Metrics
  13. Implementation Details
  14. Evaluation Results
  15. Ablation Studies
  16. ...and 3 more sections

Figures (3)

  • Figure 1: The overall framework of the proposed approach. It identifies "Borderline ECGs", i.e., those near the classification boundary, augments them with on-manifold adversarial perturbations, and trains the model using a combination of original and adversarial samples to improve robustness.
  • Figure 2: Model performance of ResNet, ResNet+DA (data augmentation), and ResNet+ADV (adversarial training) on overall and pacemaker cohorts under full and 10% training data.
  • Figure 3: Model performance across congenital heart disease lesion subgroups (VSD = Ventricular septal defect, COA=Coarctation of the aorta, HLHS = Hypoplastic left heart syndrome, CAVC = Complete atrioventricular canal, DORV = Double outlet right ventricular, TAPVR = Total anomalous pulmonary venous return).