Respiratory Rate Estimation using PPG: A Deep Learning Approach.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Respiratory rate (RR) is a critical vital sign, yet its measurement is often hindered by the lack of unobtrusive and convenient sensors.
• Photoplethysmogram (PPG) signals contain respiratory modulations that can be leveraged for RR estimation using signal processing techniques.
• Traditional methods rely on signal processing, waveform fiducial markers, and hand-crafted rules, which may not capture complex signal variations.

Purpose of the Study:

• To propose an end-to-end deep learning approach for accurate respiratory rate estimation from photoplethysmogram (PPG) data.
• To investigate the effectiveness of using a synthetic dataset to augment real-world data for training deep learning models.
• To compare the performance of the deep learning model against classical methods for RR estimation.

Main Methods:

• An end-to-end deep learning model based on the Residual Network (ResNet) architecture was developed to process time-series PPG data.
• A synthetic PPG dataset was generated and incorporated into the training process to address data scarcity issues.
• The model's performance was evaluated using 5-fold cross-validation on two public PPG datasets, with mean absolute error (MAE) as the primary metric.

Main Results:

• The inclusion of a synthetic dataset significantly improved the deep learning model's performance by 34%.
• The deep learning approach achieved a mean absolute error of 2.5±0.6 breaths per minute (brpm) for RR estimation.
• The developed deep learning model demonstrated performance comparable to a classical signal processing method.

Conclusions:

• Deep learning models, particularly when augmented with synthetic data, offer a viable and accurate method for respiratory rate estimation from PPG signals.
• This approach holds significant potential for unobtrusive vital sign monitoring, including RR, using PPG and other physiological signals.
• Further validation with large-scale, real-world data is recommended to fully realize the clinical utility of deep learning for vital sign monitoring.