Diffusion-based conditional ECG generation with structured state space models.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Data Science

Background:

• Sensitive health data distribution poses privacy challenges.
• Diffusion models and structured state space models are advanced generative techniques.
• Existing methods for synthetic electrocardiogram (ECG) generation lack robust baselines.

Purpose of the Study:

• To introduce SSSD-ECG, a novel method for generating synthetic 12-lead ECGs.
• To address privacy concerns in health data sharing.
• To establish reliable conditional generative models for ECG data.

Main Methods:

• SSSD-ECG combines diffusion models with structured state space models.
• Conditional variants of state-of-the-art unconditional generative models were developed.
• Evaluation involved assessing pre-trained classifiers on generated data and training classifiers solely on synthetic data.

Main Results:

• SSSD-ECG demonstrated superior performance compared to GAN-based competitors.
• Conditional class interpolation and a clinical Turing test validated the high quality of generated ECGs.
• The synthetic data generated by SSSD-ECG proved effective for training diagnostic classifiers.

Conclusions:

• SSSD-ECG successfully generates high-fidelity synthetic 12-lead ECGs.
• The proposed method enhances privacy in health data sharing.
• SSSD-ECG represents a significant advancement in synthetic medical data generation.