Hybrid Neural State-Space Modeling for Supervised and Unsupervised Electrocardiographic Imaging.

Area of Science:

• Medical Imaging
• Computational Electrophysiology
• Biomedical Signal Processing

Background:

• State-space modeling (SSM) is a versatile framework for image reconstruction.
• Inaccuracies in physiological knowledge can compromise SSM solutions.
• Deep learning methods offer potential but lack interpretability and require extensive labeled data.

Purpose of the Study:

• To develop a novel hybrid SSM framework for electrocardiographic imaging (ECGI).
• To integrate data-driven learning with state-space formulations for improved cardiac electrical activity reconstruction.
• To address limitations of traditional methods and deep learning in ECGI.

Main Methods:

• Developed a hybrid SSM framework combining physics-based forward operators with neural modeling.
• Introduced neural modeling for the transition function and a Bayesian filtering strategy.
• Applied the framework to reconstruct heart surface electrical activity from body-surface potentials.

Main Results:

• Demonstrated improved ECGI performance in unsupervised settings using limited ECG observations compared to fixed SSM.
• Achieved significant improvements (40.6% and 45.6%) in localizing ventricular activation origins with mixed supervised/unsupervised training.
• Outperformed traditional and supervised data-driven ECGI baselines on real data.

Conclusions:

• The hybrid SSM framework effectively reconstructs cardiac electrical activity from body-surface potentials.
• This approach enhances ECGI performance, particularly in data-limited and unsupervised scenarios.
• The framework offers a promising direction for interpretable and data-efficient ECGI.