Wearable sensors and intelligent algorithms for continuous monitoring of respiratory health

Abstract

The estimation of respiratory variables is essential for monitoring respiratory health, particularly in patients who require continuous assessment of respiratory function. For example, the detection of thoracoabdominal asynchrony, the noncoincident motion of the ribcage and abdomen during breathing, can be used to indicate respiratory distress. The assessment of the tidal volume, the volume of air that is inhaled and exhaled with each breath, provides a direct measure of a subject’s ventilatory status. However, traditional spirometry-based methods are precise but impractical for daily use due to their invasive nature, discomfort, and limitations on mobility, particularly for populations such as children or individuals with neuromuscular disorders who may struggle with standard measurement protocols. This thesis develops a wearable system for non-invasive respiratory monitoring using low-cost inertial measurement units (IMUs), aiming to provide accurate and continuous estimation of respiratory kinematics and volumetric variables outside the clinical setting. The system is designed to estimate three-dimensional thoracoabdominal displacements and derive key respiratory parameters, including respiratory rate, thoracoabdominal asynchrony, and tidal volume, by integrating nonlinear observer theory with machine learning and deep learning techniques.The first step in the research is the development of a switched-gain nonlinear observer to estimate thoracoabdominal tilt angles using IMU data. The observer is designed to reject both gravity-induced errors and integration drift while maintaining global asymptotic stability. Experimental validation against a reference optical motion capture system shows that displacement estimates typically remain within ±1 mm of ground truth across a range of breathing conditions. These displacement signals are then used to compute respiratory variables such as breathing frequency and thoracoabdominal phase angle. Respiratory asynchrony is evaluated using both visual Lissajous plots and cross-correlation-derived phase angles, demonstrating the capability to detect various breathing patterns. To move from kinematics to volumetric estimation, the system incorporates a series of models to estimate tidal volume (TV). Initially, three regression approaches—Linear Regression (LR), Support Vector Regression (SVR), and Random Forest Regression (RFR)—are trained using displacement features derived from the chest and abdomen. Among them, SVR achieves the best overall performance, with an average RMSE of 54.6 mL across postures and subjects. However, due to its sensitivity to sensor placement variability, an alternate hybrid deep learning model is developed, integrating raw IMU signals with estimated displacements as inputs to a CNN-LSTM network, a combination of convolutional neural networks (CNN) and long short-term memory (LSTM) networks. This architecture captures spatial and temporal dynamics of respiration and achieves high prediction accuracy even under repeated sensor removal and reattachment. In a small-cohort, IRB-approved study, the CNN-LSTM model yields an average RMSE of 41.30 mL, approaching the accuracy of clinical spirometry and highlighting its potential for real-world deployment. Altogether, this thesis contributes a novel, validated framework for wearable respiratory monitoring by fusing nonlinear estimation with modern artificial intelligence. The proposed system is accurate, unobtrusive, and scalable, thus offers a promising alternative to traditional respiratory monitoring tools. By enabling long-term tracking of various key respiratory metrics, this work lays the foundation for continuous monitoring, early detection of respiratory deterioration, and more personalized respiratory care.