Scholarlink

<p>Automated detection of multiple cardiac arrhythmia classes from single-lead electrocardiographic (ECG) signals remains a critical clinical challenge: existing classifiers are hampered by training-set class imbalance, limited inter-patient diversity, and synthetic data generators that lack physiological grounding. To address these gaps, we propose a Physics-Informed Generative Digital Twin (PIGDT) framework that seamlessly integrates a hybrid Physics-Informed Neural Network (PINN) with a modified Windkessel hemodynamic model and a conditional denoising diffusion probabilistic model (DDPM) for realistic ECG synthesis across five cardiac pathology classes. The framework establishes a direct hemodynamic-to-ECG coupling via PINN-solved ventricular pressure trajectories that modulate diffusion sampling to yield physiologically consistent waveforms. A balanced dataset of 1,200 synthetic ECG signals was generated spanning normal sinus rhythm, atrial fibrillation, ventricular tachycardia, myocardial ischemia, and bundle branch block. Feature engineering yields a 58-dimensional representation comprising classical signal processing features (38), persistent homology topological descriptors (8), and R-peak graph spectral features (12). A three-layer Graph Attention Network (GAT) with Model-Agnostic Meta-Learning (MAML) enables rapid patient-level personalization. Cross-database validation on real MIT-BIH and PTB-XL recordings achieves 97.5% accuracy and macro-F1 of 97.3%, surpassing a Gaussian-synthetic baseline (92.1%). Ablation experiments confirm that removing topological features incurs a 6.0% accuracy drop and removing the PINN coupling causes a 10.2% decline, validating each novel component. Interpretability is achieved through SHAP (SHapley Additive exPlanations) analysis combined with persistence diagram visualization. The pipeline achieves 28.7 ms end-to-end inference on ARM Cortex-A78 hardware, satisfying real-time cardiac monitoring requirements; however, the ∼380&#xa0;MB memory footprint requires further optimization for ultra-low-power wearable deployment. These results establish PIGDT as a promising framework for personalized cardiac monitoring, with key limitations and future directions discussed herein.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Physics-informed generative digital twin with graph neural networks for personalized multi-class ECG-based heart disease detection

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Tito Waluyo Purboyo

摘要

Automated detection of multiple cardiac arrhythmia classes from single-lead electrocardiographic (ECG) signals remains a critical clinical challenge: existing classifiers are hampered by training-set class imbalance, limited inter-patient diversity, and synthetic data generators that lack physiological grounding. To address these gaps, we propose a Physics-Informed Generative Digital Twin (PIGDT) framework that seamlessly integrates a hybrid Physics-Informed Neural Network (PINN) with a modified Windkessel hemodynamic model and a conditional denoising diffusion probabilistic model (DDPM) for realistic ECG synthesis across five cardiac pathology classes. The framework establishes a direct hemodynamic-to-ECG coupling via PINN-solved ventricular pressure trajectories that modulate diffusion sampling to yield physiologically consistent waveforms. A balanced dataset of 1,200 synthetic ECG signals was generated spanning normal sinus rhythm, atrial fibrillation, ventricular tachycardia, myocardial ischemia, and bundle branch block. Feature engineering yields a 58-dimensional representation comprising classical signal processing features (38), persistent homology topological descriptors (8), and R-peak graph spectral features (12). A three-layer Graph Attention Network (GAT) with Model-Agnostic Meta-Learning (MAML) enables rapid patient-level personalization. Cross-database validation on real MIT-BIH and PTB-XL recordings achieves 97.5% accuracy and macro-F1 of 97.3%, surpassing a Gaussian-synthetic baseline (92.1%). Ablation experiments confirm that removing topological features incurs a 6.0% accuracy drop and removing the PINN coupling causes a 10.2% decline, validating each novel component. Interpretability is achieved through SHAP (SHapley Additive exPlanations) analysis combined with persistence diagram visualization. The pipeline achieves 28.7 ms end-to-end inference on ARM Cortex-A78 hardware, satisfying real-time cardiac monitoring requirements; however, the ∼380 MB memory footprint requires further optimization for ultra-low-power wearable deployment. These results establish PIGDT as a promising framework for personalized cardiac monitoring, with key limitations and future directions discussed herein.