Synthetic seismocardiogram generation using a transformer-based neural network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiovascular Physiology

Background:

• Seismocardiogram (SCG) signals are vital for noninvasive cardiovascular monitoring.
• Limited SCG data hinders the development of advanced machine learning applications.
• Existing methods face challenges due to data scarcity.

Purpose of the Study:

• To design and validate a novel deep generative model for SCG dataset augmentation.
• To address the challenge of limited SCG data availability.
• To enable precise control over SCG signal features like aortic opening and closing.

Main Methods:

• A transformer neural network-based deep generative model was developed for SCG augmentation.
• The model allows control over specific SCG features (aortic opening, aortic closing, participant morphology).
• Generated SCG beats were compared to real human SCG using distribution distance metrics like Sliced-Wasserstein Distance (SWD).

Main Results:

• The model generated SCG signals with minimal distribution distance compared to real human SCG, outperforming other data augmentation methods.
• Input/output features showed minimal error, with 95% limits of agreement for pre-ejection period (PEP) and left ventricular ejection time (LVET) timings.
• Data augmentation using the model improved PEP estimation accuracy by 3.3% for every 10% augmentation.

Conclusions:

• The developed model generates physiologically diverse and realistic SCG signals with controllable features.
• This approach effectively overcomes data scarcity in SCG processing and machine learning.
• The model offers a unique solution for enhancing SCG datasets for improved cardiovascular monitoring.