Finger-to-Chest Style Transfer-Assisted Deep Learning Method for Photoplethysmogram Waveform Restoration With Timing Preservation

Chest-acquired photoplethysmogram (PPG) signals often suffer from severe degradation due to motion artifacts and poor perfusion, limiting their clinical utility. We propose a novel style transfer-assisted cycle-consistent generative adversarial network (starGAN) to restore chest PPG (cPPG) signals using high-quality finger PPG as a reference during training. Leveraging a dual-sensor acquisition protocol, we avoid simulated artifacts and train the model to preserve physiological timing while improving waveform quality across three PPG channels (red, green, and infrared). Evaluation on over 8000 5-s segments from 50 subjects showed a 30% improvement in waveform correlation and a 125% increase in signal-to-noise ratio (SNR) over raw cPPG. Pulse rate (PR) accuracy, compared to electrocardiogram (ECG), exceeded 84%. Multichannel input significantly outperformed single-channel restoration, and starGAN achieved up to fourfold improvement over variational mode decomposition (VMD) and other more advanced methods. Even under motion conditions (walking and stair climbing), the model improved signal quality by over 30%. These results highlight the effectiveness of cycle-consistent style transfer in restoring wearable PPG signals for reliable health monitoring from a single chest-worn sensor.