Lightweight End-to-End Deep Learning Solution for Estimating the Respiration Rate from Photoplethysmogram Signal

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Respiratory ailments pose significant health risks, particularly for COVID-19 patients.
• Respiration rate (RR) is a critical vital sign for health monitoring.
• Current RR monitoring is often limited to intensive care units (ICUs).

Purpose of the Study:

• To develop a deep learning-based, end-to-end solution for estimating RR directly from photoplethysmogram (PPG) signals.
• To evaluate the feasibility of using PPG signals for non-invasive RR estimation.
• To create a lightweight model suitable for deployment on mobile devices for continuous patient monitoring.

Main Methods:

• A deep learning approach was employed to directly estimate RR from PPG signals.
• The proposed system utilized a lightweight ConvMixer model.
• The model was evaluated on the VORTAL and BIDMC public datasets.

Main Results:

• The ConvMixer model achieved a root mean squared error (RMSE) of 1.75 bpm and mean absolute error (MAE) of 1.27 bpm on the VORTAL dataset.
• On the BIDMC dataset, the model achieved an RMSE of 1.20 bpm and an MAE of 0.77 bpm, with a correlation coefficient (R) of 0.92 for both datasets.
• Fine-tuning the model on out-of-distribution data improved performance, yielding an average R of 0.81.

Conclusions:

• Deep learning, specifically the ConvMixer model, effectively estimates respiration rate from PPG signals.
• The lightweight nature of the model enables potential real-time patient monitoring applications on mobile devices.
• Fine-tuning strategies can enhance model robustness for diverse patient populations and data distributions.