ECG Synthesis via Diffusion-Based State Space Augmented Transformer

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Signal Processing
• Cardiovascular Disease Research

Background:

• Cardiovascular diseases (CVDs) are a leading global health issue.
• Artificial intelligence (AI) and electrocardiogram (ECG) analysis show promise for CVD classification.
• Healthcare data privacy concerns hinder the development of data-driven CVD detection models.

Purpose of the Study:

• To address healthcare data confidentiality challenges for AI-driven CVD classification.
• To propose a novel method for synthesizing conditional 12-lead ECGs.
• To evaluate the quality and authenticity of the generated ECG data.

Main Methods:

• Developed a novel diffusion-based generative model.
• Integrated a State Space Augmented Transformer to capture long-term dependencies in time-series data.
• Synthesized conditional 12-lead ECGs from the PTB-XL dataset based on 12 heart rhythm classes.

Main Results:

• Generated synthetic 12-lead ECGs with assessed quality using Dynamic Time Warping (DTW) and Maximum Mean Discrepancy (MMD).
• Evaluated the authenticity of generated ECGs by comparing classifier performance on real and synthetic data.
• Demonstrated the potential of synthesized data for training AI models without compromising privacy.

Conclusions:

• The proposed diffusion model and State Space Augmented Transformer effectively synthesize realistic ECG data.
• Synthesized ECGs can mitigate privacy concerns associated with sharing sensitive patient data.
• This approach facilitates the development of robust AI tools for cardiovascular disease classification.