Denoising Motion-Corrupted Seismocardiogram Signals Using Score-Based Generative Diffusion Models

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Artificial Intelligence in Healthcare

Background:

• Noninvasive hemodynamic monitoring is crucial for cardiovascular health assessment and injury prevention.
• Current wearable sensors (ECG, PPG) have limitations in capturing cardiomechanical function.
• Seismocardiogram (SCG) signals offer insights into cardiomechanics but are prone to motion artifacts.

Purpose of the Study:

• To develop an effective motion-artifact reduction algorithm for SCG signals using a generative diffusion model.
• To enable reliable SCG signal acquisition in real-world, high-motion environments.
• To improve the accuracy of hemodynamic parameter estimation from wearable sensors.

Main Methods:

• Proposed a score-based generative diffusion model framework for SCG signal denoising.
• Leveraged SCG beat periodicity to learn a probability space for generating motion-free signals.
• Employed a multi-generation averaging approach to enhance signal quality.
• Evaluated performance on waveform and feature extraction accuracy using exercise data.

Main Results:

• Achieved low mean absolute errors for aortic valve opening (3.74 ms) and closing (7.67 ms).
• Outperformed existing signal processing and deep learning methods in feature extraction accuracy.
• Demonstrated effective denoising and generalizability on an unseen, real-world dataset.

Conclusions:

• The proposed diffusion model effectively reduces motion artifacts in SCG signals.
• This technology can enhance wearable systems for reliable hemodynamic monitoring.
• Potential to improve cardiovascular assessment and reduce injuries in high-risk individuals.