Estimation of respiratory effort using tracheal sound and movement

Nasim Montazeri 1, 2; Bojan Gavrilovic 2; Shumit Saha 1, 2; Babak Taati 1, 2, 3; Hisham Alshaer 2; Azadeh Yadollahi 1, 2

 1. Institute of Biomaterials and Biomedical Engineering, University of Toronto; 2. Toronto Rehabilitation Institute, University Health Network; 3. Department of Computer Science, University of Toronto

Background: Respiratory effort (RE) is commonly used for diagnosis of respiratory sleep disorders including hypopnea and apnea which are associated with reduction and cessation of airflow, respectively. Dual thoraco-abdominal belts used for measuring RE can be inconvenient. In previous studies, tracheal sound has been suggested as an informative source for assessing respiration system. The purpose of this study is to develop a convenient and reliable method to estimate RE using tracheal sound and movement.    

Methods: Participants with suspected sleep disorder were recruited to undergo overnight polysomnography (PSG). Tracheal sound and movement were recorded simultaneously with the PSG using a small wearable device called The Patch over their neck. For our reference gold standard, RE measured using by the conventional belts were extracted and the average peak-to-peak changes were calculated within a 10-second moving window. The same value was calculated for tracheal movement in subsequent to low-pass filtering with 0.7 Hz cut-off frequency to capture the breathing pattern. The tracheal sound during sleep is usually contaminated by snoring, thus an unsupervised algorithm based on wavelet decomposition was employed to remove the effect of snoring from sound. Then, the average sound energy was calculated within the 10-second moving window. The baseline of the signals varies depending on factors such as sleep stages and body posture. Thus, using an adaptive normalization algorithm, long-term baseline variability was detected and removed. Finally, average sound energy and peak-to-peak value of tracheal movement were used as features to train three different predictive models including Kalman Filter (KF), linear regression (LR) and Random Forest (RF). The performance of the models was investigated using repeated measure correlation between estimated and reference RE during hypopneas, apneas, normal segments as well as over the whole night data.

Results: Data from 10 subjects (6 females), age of 51.8±9.9 years, BMI of 28.6±4.1 kg/m2 and apnea/hypopnea index of 23.8±26.7 events/hour were analyzed.  There was a significant positive correlation between estimated RE from the models (KF, LR, RF) and reference RE during apnea (r=0.68, 0.55, 0.33), hypopnea (r=0.52, 0.46, 0.34), normal segment (r=0.54, 0.46, 0.34) and over whole data (r=0.56, 0.47, 0.35) over all subjects.

Conclusion: Our results indicate that by analyzing the tracheal sound and movement, the proposed algorithm can estimate RE with acceptable correlation in comparison with the RE measured over the chest and abdomen. After validation over a larger patient cohort, the proposed model can be used to estimate RE that can be used for detecting sleep apnea.