EdgeCaps: Pruned Capsule Networks for RF Vital Sign Measurement in Real-time

Abstract Radio frequency (RF) sensing enables both non-contact and non-invasive monitoring of vital signs, offering a promising solution for unobtrusive health assessment. Recent advances in machine learning (ML) and deep learning (DL) have significantly improved the accuracy and reliability of RF-based vital sign monitoring. However, despite their high performance, state-of-the-art DL models are often computationally intensive and energy-demanding, making them unsuitable for continuous use on low-cost, resource-constrained edge devices. Moreover, real-time deployment of such models on embedded platforms remains underexplored in the literature. To address this challenge, we propose EdgeCaps, a system that compresses a high-capacity capsule neural network (CapsNet) into a compact student model optimized for edge deployment. Specifically, EdgeCaps distills the knowledge of a 33.7 million-parameter teacher model into a compact student model with only 2.11 million parameters, while maintaining comparable representational capability. Using knowledge distillation (KD)-where a smaller model (student) learns to replicate the outputs of a larger model (teacher)-the student model achieves an accuracy of 85.5%, just two percentage points below the teacher, despite a 94% reduction in parameter count. We validate the proposed system through continuous respiration monitoring of five volunteers using a Raspberry Pi 4. A Monte Carlo-based evaluation strategy is employed to measure the generalization and robustness of the results. During deployment, the student model maintains a mean inference latency below 0.3 s, RAM usage around 18 %, and core temperature consistently under 44 °C, all within safe operational limits. In contrast, the teacher model exhibits latencies ranging from 0.4 s to 0.5 s, core temperatures up to 51 °C, and RAM usage up to 22%. These results demonstrate that EdgeCaps enables real-time, energy-efficient, and thermally stable RF-based vital sign monitoring on edge devices without reliance on off-device computation. To our knowledge, this work is the first to introduce a KD-based CapsNet for RF sensing and the first to provide a comprehensive end-to-end evaluation of continuous physiological monitoring on embedded hardware.