Classification of human breathing patterns and poses using millimeter wave (MmWave) radar signals for remote healthcare monitoring

Abstract

This thesis explores the classification of human breathing patterns and static body poses using physiological signals captured through millimeter wave (mmWave) radar. It addresses the increasing need for contactless and privacy-preserving health monitoring systems, which are particularly relevant in elderly care and intensive care settings, where continuous observation of vital signs and behavioral states can enhance patient safety and reduce the burden on healthcare professionals.

The study focuses on two core classification tasks: the detection of four breathing patterns, namely normal breathing, guided breathing, reading, and breath holding (simulating apnea), and the recognition of three static poses: standing, sitting, and lying down. In addition to classification, the thesis investigates regression tasks aimed at estimating age and body mass index from radar-derived physiological signals. These estimations support the development of personalized monitoring systems that adapt to individual characteristics. The methodology follows a complete signal processing pipeline, beginning with a theoretical overview of mmWave radar and its applications in physiological monitoring. Raw signals are preprocessed through filtering and segmentation using various windowing strategies to evaluate their influence on model performance. Two main modeling approaches are examined: traditional machine learning, which involves extracting handcrafted features from segmented windows, and deep learning, which employs one-dimensional convolutional neural networks to process raw signals directly.

Experiments were conducted on the OMuSense 23 dataset, comprising respiratory and cardiac signals from 50 participants engaged in different breathing and posture activities. The evaluation includes multiple configurations, varying segmentation strategies and modeling techniques, and uses both Leave One Subject Out and K-Fold Cross-Validation protocols. The results provide a comparative analysis of traditional and deep learning methods for classification and regression tasks. The findings demonstrate the potential of the mmWave radar as a viable tool for non-invasive health monitoring and offer insights into how preprocessing and modeling decisions influence overall performance in physiological signal analysis.