Cross-domain generalization in RF-based respiratory sensing: quantifying and closing the sensing-angle gap

The growth of remote and home-based care models in the United Kingdom, including NHS virtual wards for respiratory and frailty pathways, has increased the need for scalable, contactless vital-sign monitoring technologies suitable for the home environment. Contact-based respiratory monitoring is poorly suited to unsupervised home settings, motivating investigation of radio-frequency (RF) sensing as a contactless alternative that avoids wearable hardware and patient compliance burdens. This paper investigates a deep multilayer perceptron (MLP) for classifying human respiratory patterns from software-defined radio (SDR) signals as a candidate technology for remote respiratory monitoring. Using a publicly available USRP X310-based respiratory dataset from 5 participants, breathing activity was classified into three categories (Normal Respiration (NR), Fast Respiration (FR), and Sleep-Apnea-Related (SAR) events), directly from raw time-domain signal segments without hand-crafted features. The model, with three hidden layers (256, 128, 64 units) and dropout regularization, was compared against a logistic regression baseline and evaluated for robustness using leave-one-angle-out cross-validation across sensor-to-subject geometries. The MLP achieved 87.18% test accuracy under a random split (95% bootstrap CI: [86.36%, 87.94%]), pending confirmation of no participant-level leakage (a notable concern given the small cohort), versus 60% for logistic regression, with most errors occurring between FR and SAR. However, leave-one-angle-out validation revealed accuracy dropping from 93.47% to 53.45% on unseen angles, with mean cross-domain accuracy of only 73.03%. These results demonstrate the feasibility of contactless, deep-learning-based respiratory classification from SDR signals.