Heart rate estimation from ballistocardiographic signals using deep learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Ballistocardiography (BCG) offers an unobtrusive, cost-effective method for patient health monitoring.
• Accurate heart rate estimation is crucial for effective health surveillance.
• Traditional signal processing methods for BCG analysis face challenges with inter-patient variability.

Purpose of the Study:

• To evaluate deep learning models for heart rate estimation from BCG signals.
• To compare the performance of deep learning against established digital signal processing techniques.
• To investigate smaller, less complex deep learning architectures for BCG analysis.

Main Methods:

• Deep learning architectures including convolutional and recurrent neural networks were implemented.
• Models were trained and tested on BCG recordings from 42 patients using a pneumatic system.
• Performance was assessed using mean absolute error (MAE) and compared across different patient cohorts.

Main Results:

• Deep learning models significantly outperformed traditional methods, achieving an MAE of 2.07 beat/min versus 4.24 beat/min.
• Deep learning models demonstrated more consistent performance across patients, better handling inter-patient variability.
• A compact deep learning model with 8,283 parameters achieved an MAE of 2.32 beat/min, indicating efficiency.

Conclusions:

• Deep learning offers a substantial advancement over signal processing algorithms for heart rate estimation from BCG.
• Smaller deep learning models provide sufficient performance for biosignal processing, challenging the need for excessively large networks.
• This study highlights the potential of fully convolutional networks for 1D signal processing in BCG applications.