Respiration Tracking Using Wearable Inertial Measurement Unit System and CNN-Transformer-Based Hybrid Model for Healthcare Applications

Respiration analysis is essential for the identification and diagnosis of cardiovascular disorders, including sleep apnea and chronic obstructive pulmonary disease (COPD). Although deep learning models for respiratory identification in everyday activities have garnered significant attention. Many existing algorithms are often too complex and inefficient for practical use. This highlights the pressing need to develop more advanced methodologies that can provide accurate and reliable results without the complexity and computational burden of current solutions. Such improvements would enhance the usability of these models in real-world applications, particularly for continuous monitoring in wearable devices. This research presents a novel convolution and transformer-based hybrid model (RePair-Net) for predicting respiratory parameters from inertial measurement unit (IMU) data. The research is conducted on ten healthy participants without any neurological or pulmonary conditions. The RePair-Net model was tested across three self-paced breathing conditions-slow, normal, and fast to evaluate its consistency and robustness. The estimated respiration rate (RR) was then compared against measurements obtained from a gold-standard commercial sensor belt (SB). The mean absolute error (MAE) of RePair-Net compared to SB for normal, fast, and slow breathing conditions was 1.12±1.51, 1.61±1.65, and 1.50±1.02, respectively, across subjects. The average root mean square error (RMSE) among subjects was 1.32±0.54 for normal breathing, 2.75±0.87 for fast breathing, and 1.43±0.73 for slow breathing. The proposed model also quantifies many temporal characteristics of the respiratory signal, aiding in the early diagnosis, prediction, and treatment of COPD-related disorders.