Generative adversarial networks with fully connected layers to denoise PPG signals.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Photoplethysmography (PPG) signals are crucial for non-invasive physiological monitoring.
• Motion artifacts frequently distort PPG signals, limiting their clinical utility.
• Existing methods for PPG reconstruction often rely on additional sensor data.

Purpose of the Study:

• To develop and evaluate a generative adversarial network (GAN) for reconstructing motion-corrupted PPG signals.
• To assess the performance of the GAN-based reconstruction method without using movement sensor data.
• To determine the accuracy of heart rate estimation from reconstructed PPG signals.

Main Methods:

• A generative adversarial network (GAN) with fully connected layers was designed for PPG signal reconstruction.
• Clean PPG signals from the BIDMC Heart Rate dataset were artificially corrupted to create training and testing data.
• The model was trained and validated using processed data from the MIMIC II waveform database.

Main Results:

• The proposed GAN model achieved a mean absolute error of 1.31 bpm in heart rate estimation for signals within the 70-115 bpm range.
• The model demonstrated effectiveness in reconstructing PPG signals irrespective of the length and amplitude of introduced corruption.
• Accurate heart rate extraction was achieved from the reconstructed PPG signals.

Conclusions:

• The developed GAN architecture is effective for reconstructing noisy PPG signals, even with significant motion artifacts.
• The model shows promise for real-time PPG signal processing without the need for accelerometer or gyroscope inputs.
• This approach offers a robust method for improving the reliability of PPG-based health monitoring.