Development of Robust Algorithm for Prediction of Cardiac Abnormalities from Heart Sounds
Kabir, Muammar 1, 2 ; Gavrilovic, Bojan 2 ; Saha, Shumit 1, 2 ; Yadollahi, Azadeh 1, 2
1. Institute of Biomaterials and Biomedical Engineering, University of Toronto; 2. Toronto Rehabilitation Institute, University Health Network
Background: Obstructive sleep apnea (OSA) is a disorder of breathing that is characterized by repetitive complete (apnea) or partial (hypopnea) closure of the pharyngeal airway during sleep. OSA adversely affects autonomic nervous system activity, altering the autonomic modulation of heart rate through elevated sympathetic nerve activity which can lead to long-term cardiovascular disorders. Therefore, OSA patients will significantly benefit from technologies that can identify apnea/hypopnea events and monitor any symptoms or risks of cardiac arrhythmias. Currently available techniques for monitoring pharyngeal airway collapse such as polysomnography and endoscopy are expensive and inconvenient. We have developed a small wearable device (The Patch) which can record respiratory and heart sounds over the neck. We have shown that Patch can detect apnea and hypopnea events during sleep using respiratory sounds analysis. The purpose of this study is to leverage our current acoustic technologies and develop robust algorithms based on heart sounds analysis to investigate the effects of apneas and hypopneas during sleep on cardiac abnormalities in OSA patients.
Methods: Tracheal respiratory and heart sounds will be recorded in 50 subjects using The Patch. An algorithm to detect severity of pharyngeal collapses is being developed by our group which will be used to identify apnea/hypopnea events. Further, to extract cardiac features, an efficient and robust algorithm based on adaptive filtering, wavelet de-noising and time-spectral features will be developed to localize and separate heart sounds from breathing, snoring, swallowing, background noises and movement artefacts. Techniques such as power-spectral analysis, empirical mode decomposition and the Hilbert transform will be used to quantify heart sounds, and extract acoustic features (sound intensity, spectral power and pitch) and hemodynamics (heart rate, blood pressure, cardiac output and systolic time interval). To predict abnormalities in cardiac function, heart rate variability and heart sound features such as temporal features of second heart sound, presence and intensity of third and fourth heart sounds will be investigated. Accuracy of the developed algorithm in detecting cardiac arrhythmias will be validated against electrocardiogram recording from overnight polysomnography.
Expected Results: We anticipate that the developed algorithm will be robust in: 1) separating heart sounds from tracheal sounds; 2) predicting hemodynamics; 3) identifying heart sound features that could predict abnormalities in cardiac function.
Significance: The significant outcome of this study is the development of a robust and convenient tool for long-term monitoring and comprehensive assessment of cardiac function non-invasively during sleep and daytime activities. The algorithm can be used in future applications such as assessing the risk of heart disease, hypertension, and atrial fibrillation.