Finger-to-Chest Style Transfer-assisted Deep Learning Method For Photoplethysmogram Waveform Restoration with Timing Preservation

TL;DR

This work addresses the challenge of restoring chest-worn PPG signals degraded by motion and low perfusion. It introduces a style-transfer, cycle-consistent GAN (starGAN) that translates finger PPG style to chest PPG while preserving timing, using three synchronized PPG channels from the Soundi sensors. Across 40 subjects and thousands of 5-second chunks, the method achieves ~${R}\approx0.89$ with substantial improvements in RMSE, SNR, and heart-rate–related metrics, demonstrating performance comparable to recent literature and strong potential for single-device wearable health monitoring. The multi-channel fusion approach and real-artifact data collection enhance generalization and clinical relevance, enabling more reliable cardiovascular assessments in daily life.

Abstract

Wearable measurements, such as those obtained by photoplethysmogram (PPG) sensors are highly susceptible to motion artifacts and noise, affecting cardiovascular measures. Chest-acquired PPG signals are especially vulnerable, with signal degradation primarily resulting from lower perfusion, breathing-induced motion, and mechanical interference from chest movements. Traditional restoration methods often degrade the signal, and supervised deep learning (DL) struggles with random and systematic distortions, requiring very large datasets for successful training. To efficiently restore chest PPG waveform, we propose a style transfer-assisted cycle-consistent generative adversarial network, called starGAN, whose performance is evaluated on a three-channel PPG signal (red, green,and infrared) acquired by a chest-worn multi-modal sensor, called Soundi. Two identical devices are adopted, one sensor to collect the PPG signal on the chest, considered to feature low quality and undergoing restoration, and another sensor to obtain a high-quality PPG signal measured on the finger, considered the reference signal. Extensive validation over some 8,000 5-second chunks collected from 40 subjects showed about 90% correlation of the restored chest PPG with the reference finger PPG, with a 30% improvement over raw chest PPG. Likewise, the signal-to-noise ratio improved on average of about 125%, over the three channels. The agreement with heart-rate computed from concurrent ECG was extremely high, overcoming 84% on average. These results demonstrate effective signal restoration, comparable with findings in recent literature papers. Significance: PPG signals collected from wearable devices are highly susceptible to artifacts, making innovative AI-based techniques fundamental towards holistic health assessments in a single device.

Figures (6)

• Figure 1: Procedure to label fPPG 5-s chunks.
• Figure 2: Architecture of the starGAN model.
• Figure 3: Restoration of high- (left panel) and low-quality (right panel) measured cPPG signals. In both cases the pulse rate can be easily detected from the restored cPPG signal (blue line). More specifically, in the restoration of high-quality signals, the R values between the restored cPPG signal and the fPPG signal reached values of 0.98, 0.99, and 0.99 for the red, infrared, and green channels, respectively. In contrast, for the low-quality signal restoration, the R values were 0.63, 0.62, and 0.70 for the red, infrared, and green channels, respectively.
• Figure 4: Scatterplot and Bland-Altman for test set. Without lack of generality the green channel was here considered. Top: plots for the PP intervals (s) of measured cPPG vs. fPPG, along with the corresponding PP distributions and the $p$ value. Bottom: plots for the PP intervals of restored cPPG vs. fPPG, along with the corresponding PP distributions and the $p$ value.
• Figure 5: Lag histogram for the restored cPPG against measured cPPG and fPPG.