Abstract
Training deep learning models for medical diagnostics, particularly in electrocardiography (ECG), is often hindered by data privacy, the scarcity of rare conditions, and the biological complexity of 12-lead signals. We present MultiODE-GAN, a generative framework that incorporates physiological constraints via Ordinary Differential Equations (ODEs) directly into the GAN loss function. By enforcing inter-lead mathematical dependencies and individual lead dynamics, MultiODE-GAN produces synthetic heartbeats that significantly enhance the specificity of diagnostic classifiers.
Introduction
The development of robust diagnostic tools for cardiac health requires large-scale, high-fidelity datasets. However, researchers face significant hurdles: strict privacy regulations (GDPR/HIPAA), the inherent rarity of certain cardiac pathologies, and the difficulty of capturing the intricate biological dependencies between the 12 leads of a clinical ECG. While mathematical simulators offer biological accuracy, they lack the diversity of real-world data. Conversely, standard deep generative models often produce signals that appear realistic in isolation but fail to maintain physiological consistency across multiple leads.
Methodology
MultiODE-GAN addresses these challenges by integrating an ECG Dynamical Model (EDM) into the training process of a WaveGAN-based architecture.
The core innovation is the Euler Loss, which utilizes a system of three coupled ODEs to define the expected biological trajectory of P, Q, R, S, and T waves. The generator's loss function is a weighted combination of the standard Wasserstein GAN loss and two Euler Loss components:
- Individual Lead Fidelity: Ensures each lead follows realistic cardiac dynamics.
- Inter-Lead Consistency: Enforces clinical mathematical relationships, such as Einthoven’s triangle and Goldberger’s central terminal, ensuring that the 12 leads are physically consistent with one another.
Experimental Setup
The model was trained and validated using the Georgia 12-Lead ECG Challenge (G12EC) dataset, comprising 10,344 ECGs across 27 diagnoses. The dataset was split 80/20 for training and testing. Synthetic data generated by MultiODE-GAN was used to augment the training set for a ResNet-based classifier to evaluate performance improvements.
Results and Discussion
MultiODE-GAN demonstrates superior performance compared to existing generative baselines, including DCGAN, WaveGAN, and diffusion-based models (SSSD-ECG).
A key finding is the δ trade-off, where a balance (optimized at δ = 0.6) must be struck between individual lead accuracy and global consistency. Over-optimizing for a single lead, as seen in previous approaches like SimGAN, tends to degrade the overall realism of the 12-lead set.
Notably, the inclusion of synthetic data primarily boosted specificity. For instance, in detecting Right Bundle Branch Block (RBBB), specificity rose from a baseline of 0.89 to 0.92. This suggests that the physics-informed synthetic data helps classifiers better distinguish "normal" physiological variations from actual abnormalities, thereby reducing false positives.
Conclusion
MultiODE-GAN successfully bridges the gap between mathematical modeling and deep learning. By "baking" physiological rules into the generative process, the model produces high-quality, consistent 12-lead ECG data. Future work will focus on extending this framework from segmented heartbeats to continuous, long-term ECG recordings and refining the modeling of rare, complex arrhythmias.
Summary updated