Exploring latent diffusion models for ECG generation on the minute scale.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Machine Learning

Background:

• Clinical electrocardiogram (ECG) datasets are often imbalanced, limiting the performance of AI interpretation models.
• Existing generative AI for biosignals produces short segments, hindering clinical utility.
• Techniques like inpainting for artifact removal remain underexplored in biosignal synthesis.

Purpose of the Study:

• To develop ECGEN, a latent diffusion model (LDM) for synthesizing long-duration, realistic ECGs.
• To enable data augmentation, rhythm-specific generation, and signal restoration for ECG analysis.
• To overcome limitations of existing AI models in handling diverse and rare ECG pathologies.

Main Methods:

• ECGEN was developed in three configurations (30-second, 90-second, and 320-second models) using a VQ-VAE and DDIM.
• Models were trained on real clinical ECGs from stroke patients.
• Evaluation metrics included heart rate (HR), heart rate variability (HRV), and morphological coherence.

Main Results:

• ECGEN-Small achieved high accuracy (AUC 0.98) in classifying atrial fibrillation (AFib) versus sinus rhythm.
• ECGEN-Medium effectively inpainted missing ECG segments, preserving plausible HR dynamics.
• ECGEN-Large generated long ECGs with consistent morphology, though HRV distributions showed shifts, indicating challenges in modeling long-range dependencies.

Conclusions:

• Latent diffusion models (LDMs) are feasible for generating long-duration ECGs, useful for data augmentation and signal restoration.
• Unsupervised biosignal synthesis presents challenges, including distributional mismatches and artefacts.
• Future research should focus on enhancing long-range temporal modeling and realism through advanced training techniques.