Signal completion using generative adversarial networks for enhanced photoplethysmography measurement accuracy

Despite the growing adoption of wearable photoplethysmography (PPG) devices in personal health management, their measurement accuracy remains limited due to susceptibility to noise. This paper proposes a novel signal completion technique using generative adversarial networks that ensures both global and local consistency. Our approach innovatively addresses both short- and long-term PPG variations to restore waveforms while maintaining waveform consistency within and between pulses. We evaluated our model by removing up to 50 % of segments from segmented PPG waveforms and comparing the original and reconstructed waveforms, including systolic peak information. The results demonstrate that our method accurately reconstructs waveforms with high fidelity, producing natural and seamless transitions without discontinuities at reconstructed boundaries. Additionally, the reconstructed waveforms preserve typical PPG shapes with minimal distortion, underscoring the effectiveness and novelty of our technique.