MS-MBSP: Multi-Scale Multi-Band Spectral Pooling for Device-Free Wi-Fi CSI Respiratory State Classification

This is the authors' abstract. We don't add key points for this paper.

Wi-Fi Channel State Information (CSI) is a contactless sensing modality that captures respiratory chest motion through subcarrier-wise amplitude variation. This paper presents MS-MBSP (Multi-Scale Multi-Band Spectral Pooling), a CNN-BiLSTM-based configuration for device-free five-class respiratory state classification, including NORMAL, SLOW, FAST, HOLD, and COUGH. MS-MBSP introduces two structural components: multi-band spectral pooling that separates respiratory and cough-related frequency regions, and a dual-scale branch that processes time-aligned 5 s and 10 s windows. An auxiliary respiratory-rate regression head based on PSD-peak pseudo-labels is additionally used as a mild training-time regularizer; removing it changes the 90-fold mean Macro F1 by only Δ =+0.006 (paired Wilcoxon $p=0.333$ ), so it is not a performance-driving component. The method is evaluated on a 10-subject Wi-Fi CSI dataset using a subject-disjoint validation rotation protocol, where the training, validation, and test subjects are mutually separated across 90 folds. Under identical training settings, we compare MS-MBSP with three time-series baselines and four spectral-pooling variants, including spec, spec_large, multi_band_5s, and multi_band_10s. MS-MBSP achieves a 90-fold mean Macro F1 of 0.7612 and an accuracy of 0.8129. The parameter-controlled spec_large variant shows little improvement over spec, indicating that the gain mainly comes from the proposed structural design rather than parameter count. Per-class and confusion-matrix analyses show that the multi-band and multi-scale components improve COUGH and SLOW classification while residual errors remain mainly between NORMAL, SLOW, and HOLD. Classical spectral-feature classifiers (random forest, RBF-SVM, and gradient-boosted trees) reach at most 0.664 Macro F1 under the identical protocol, indicating that the task is not trivially separable. Because the 90 folds are not independent, significance is assessed with the subject as the unit ( n=10 ), which places the MS-MBSP mean in a 95\,% confidence interval of [0.701,\,0.821] . All data were collected from ten healthy adults seated in a single laboratory environment under instructed breathing, so these results should be read as a controlled structural comparison of CSI-based respiratory classifiers rather than as evidence of deployment-ready generalization.