How Much Can We Salvage in Disrupted RF Vital Signs Monitoring? A Measurement Study of Post-Processing

Continuous monitoring of vital signs offers valuable insights into health status through patterns of changes over time. Radio Frequency (RF) solutions have gained significant attention due to their non-invasive and privacy-preserving nature. Nevertheless, the reliability of RF-based vital signs monitoring remains an ongoing challenge, as RF signals are inherently fragile and susceptible to disruptions, especially due to random body movements. Most of methods proposed to enhance the robustness of RF vital signs sensing reply on accessing and analyzing raw RF signals. However, the diversity of RF configurations and the restricted accessibility of commercial off-the-shelf (COTS) RF solutions make direct analysis of raw RF data impractical for improving accuracy. To address this gap, we propose a novel framework that focuses solely on post-processing to recover the vital signs from disrupted RF signals. Our approach implements a suite of classical smoothing and denoising algorithms, alongside representative data-driven techniques, to rectify noisy and disrupted RF vital signs estimations through data reconstruction. We evaluate these post-processing techniques using a dataset containing 58 hours collected from 3 subjects in cluttered, freeliving environments. Our results show that applying Temporal Convolutional Network (TCN) to RF heart rate (HR) estimations doubles the percentage of data below 5 bpm error against ground truth. We additionally find that RF respiration rate (RR) estimations is relatively robust and a simple moving average can increase the percentage of data below 2 bpm error by over 20%. We assess the generalizability of these methods through a leave-one-out evaluation and analyze their respective computational costs, shedding light on practical trade-offs between accuracy and resource requirements.