Advancing Stress Detection with Chaotic Attractor Informed Synthesis of PPG Signal

Kaiwen Hu, Sipo Zhang, Xiaowei Zhang, Qiqi Zhao, Guangyuan Gao, Tianzhi Wang, Jian Shen, Bin Hu

Abstract: Photoplethysmography (PPG) signals have been extensively used for monitoring stress level and improving human mental health. A major obstacle to improving PPG classification is the scarcity of real signals, necessitating the employment of signal synthesis techniques. Incorporating the dynamics of PPG into the generative adversarial network (GAN) helps model the physiological dynamics and improves synthesis quality. However, differential equations that can describe the dynamic characteristics of PPG are unavailable. To address this issue, we propose a novel generative adversarial network for PPG synthesis with data-driven attractor constraints (AC-GAN). Firstly, we design a recurrent neural network (RNN), which can continuously update its hidden states, to extract the chaotic motion characteristics of real PPG signals in a purely data-driven mode. Subsequently, the pre-trained attractor extraction network is used as a prior in the optimization process of a GAN to create PPG signals that conform to the underlying dynamics of physiological systems. Several experiments on three public datasets demonstrate that the signals generated by AC-GAN are optimal in terms of both similarity and usability compared to several state-of-theart methods.

External IDs:dblp:conf/bibm/HuZZZGWSH25