Non-invasive tidal volume estimation with wearable sensors using a high-gain observer and deep learning

Non-invasive tidal volume (TV) estimation can be valuable for respiratory monitoring, particularly for patients needing continuous assessment. Traditional spirometry-based methods are precise but impractical for daily use due to their invasive nature, discomfort and limitations on mobility. This study integrates a nonlinear high-gain observer (HGO) with a convolutional neural network long short-term memory network (CNN-LSTM) to estimate TV using wearable inertial measurement unit (IMU) sensors. The HGO provides reliable thoracoabdominal displacements by mitigating sensor drift and removing gravity components measured by the accelerometer. Combined with raw IMU data, these displacements serve as inputs for a deep learning CNN-LSTM network, which captures spatial and temporal dependencies to improve prediction accuracy. The CNN-LSTM model trained with both data sources demonstrated superior accuracy and also a high degree of robustness to sensor placement variations. Experimental results in an IRB approved study with 6 subjects show that the method achieved an averaged RMS error of 40.38 mL even with repeated taking off and re-wearing of the sensors. These findings underscore the potential of replacing invasive spirometry with convenient wearable sensors when coupled with reliable estimation algorithms.