Generating Breathing Patterns in Real-Time: Low-Latency Respiratory Phase Tracking From 25 Hz PPG

Ian Karman, Yue Sun, Rahil Soroushmojdehi, Jose A. Silva, Mostafa 'Neo' Mohsenvand

Abstract: This study presents a low-latency, real-time breathing cycle tracking system utilizing a conditional Generative Adversarial Network (GAN) with Wasserstein loss, with a low-powered, low sample rate Photoplethysmography (PPG) sensor. The aim is to provide a clinically accurate respiratory tool capable of tracking and visualizing the breathing cycle and rate in real-time for at-home and general ambulatory applications. To detect breathing activity in real-time, we used a wearable headband with a 25 Hz PPG sensor and an inductive respiratory sensor as ground truth. To meet the real-time and low latency constraints, the inputs were processed in 1-s windows. Signal processing and machine learning techniques were explored, and the proposed GAN-based method with Wasserstein loss and gradient penalty, outperformed others in accurately tracking the ground-truth breathing curve. Leveraging the GAN-generated breathing curve, a peak-detection algorithm calculated the respiratory rate (RR) with an average mean absolute error (MAE) of 1.47 breaths per minute (bpm) across 10 test subjects, comparable to high-sampling rate PPG literature (1 bpm), but with the advantage of 5 times faster real-time monitoring. The GAN-generated respiratory signal from a low-sampling rate wearable PPG sensor demonstrates potential as a viable alternative to traditional respiratory monitoring systems. This system offers valuable breathing monitoring, useful in various applications such as pain management.