A CNN and diffusion based encoder-decoder model to extract fine-grained human vital waveforms from mmWave radar

• Main Author: Zuo, Jia
• Other Authors: Xie Lihua
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Computer and Information Science; Engineering; Medicine, Health and Life Sciences; Non-contact vital sign monitoring; Fine-grained ECG waveform recovery; mmWave radar; Diffusion model
• Online Access: https://hdl.handle.net/10356/182152
• Institution: Nanyang Technological University

Vital signs like electrocardiogram (ECG) waveforms, are critical for early disease monitoring and prevention, including the detection of cardiac abnormalities. Radio-frequency (RF) sensors like millimeter-wave (mmWave) radar, have emerged as significant tools for non-contact vital sign monitoring in healthcare applications. These sensors enable continuous monitoring while eliminating the movement restrictions and potential discomfort associated with contact-based methods. However, accurately reconstructing fine-grained ECG waveforms remains challenging, as the detection of specific ECG features, such as P-waves, T-waves, and R-peaks, is more complex than simple pulse detection, and traditional signal processing methods struggle to recover these waveforms effectively. In prior research, some studies have leveraged Long Short-Term Memory (LSTM) networks or one-dimensional Convolutional Neural Networks (1D CNNs) for waveform segmentation, with limited success in recovering high-resolution waveforms. In this work, we propose an end-to-end neural network model comprising a wavelet transform and CNN-based encoder paired with a diffusion model-based decoder to facilitate the fine-grained recovery of ECG waveforms from a non-contact mmWave radar monitoring system. Experimental results demonstrate that our model, leveraging the powerful signal recovery capabilities of the diffusion model, achieves state-of-the-art (SOTA) performance, with a 51.6% reduction in Mean Absolute Error (MAE), a 51.9% reduction in Mean Squared Error (MSE), and an 18.7% improvement in the temporal correlation coefficient compared to leading 1D segmentation models.