Heartbeat Detection from Ballistocardiogram Signals Using a Transformer Network

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Signal Processing

Background:

• Longitudinal monitoring of heart rate (HR) and heart rate variability (HRV) is crucial for assessing cardiovascular diseases (CVDs), sleep quality, and autonomic nervous system activity.
• Current methods for HR/HRV monitoring may be invasive or lack long-term applicability in everyday settings.

Purpose of the Study:

• To develop and evaluate a transformer network for precise heartbeat timing detection from Ballistocardiogram (BCG) signals.
• To assess the performance of segment-based versus subject-based models for HR and HRV estimation using BCG data.

Main Methods:

• A transformer network was designed to predict electrocardiogram (ECG) signals from input BCG signals for heartbeat detection.
• Performance was evaluated using segment-based and subject-based models across three datasets: young adults, older adults, and a combined group.
• Correlation coefficients against ground truth ECG were calculated for HR and mean heart beat interval (MHBI).

Main Results:

• The segment-based model achieved superior performance compared to the subject-based model.
• Correlation coefficients of 0.97 were obtained for both HR and MHBI using the segment-based model against ground truth ECG.
• The model demonstrated effectiveness across different age groups.

Conclusions:

• A non-invasive transformer network approach using BCG signals can accurately detect heartbeat timing for HR and HRV monitoring.
• This method holds significant potential for long-term, at-home monitoring to aid in the early detection and prevention of cardiovascular issues.
• The segment-based approach is recommended for enhanced accuracy in HR/HRV estimation from BCG.