Hisham Alshaer

Validation of an automated algorithm for detecting apneas and hypopneas by acoustic analysis of breath sounds

BACKGROUND Sleep-disordered breathing (SDB) is common and is associated with increased risk for cardiovascular disease. However, most patients remain undiagnosed due to lack of access to sleep laboratories. We therefore tested the validity of a single-channel monitoring setup that captures and analyzes breath sounds (BSs) to detect SDB. METHODS BS were recorded from 50 patients undergoing simultaneous polysomnography (PSG). Using custom-designed automatic software, BS were subjected to a set of pattern recognition rules to identify apneas and hypopneas from which the acoustic apnea-hypopnea index (AHI-a) was calculated. Apneas and hypopneas from PSG were scored blindly by three technicians according to two criteria; one relying solely on the drop of the respiratory signal by >90% for an apnea and by 50% to 90% for a hypopnea (TV50 criteria), and another that also required a desaturation or an arousal for a hypopnea (American Association of Sleep Medicine [AASM] criteria). PSG AHI (AHI-p) was calculated for each technician according to both criteria. RESULTS There was no significant difference between AHI-p scores according to TV50 and AASM criteria. AHI-a was strongly correlated with AHI-p according to both TV50 (R=94%) and AASM criteria (R=93%). Bland-Altman plot analysis revealed that 98% and 92% of AHI-a fell within the limits of agreement for AHI-p according to TV50 and AASM criteria, respectively. Based on a diagnostic cutoff of AHI-p≥10 for SDB, overall accuracy of AHI-a reached 88% and negative predictive value reached 100%. CONCLUSION Acoustic analysis of BS is a reliable method for quantifying AHI and diagnosing SDB compared to simultaneous PSG.

Phase tracking of the breathing cycle in sleeping subjects by frequency analysis of acoustic data

We tested the hypothesis that the inspiratory and expiratory phases of breathing could be identified from breath sound recordings during sleep. Breath sounds were digitally recorded from 10 subjects during sleep. Frequency spectra of inspiration and expiration were determined. The ratio of frequency magnitude bins between 400-1000 Hz to frequency bins between 10-400 Hz was calculated for inspiration (Ri) and expiration (Re) for each breath. The Ri/Re ratio was significantly greater than the thresholds of 1.5 (p < 0.001) and 2-fold (p < 0.001). Breathing phases were correctly identified in 90% and 73% of cases using the 1.5 and 2.0 thresholds, respectively.

Detection of upper airway narrowing via classification of LPC coefficients: Implications for obstructive sleep apnea diagnosis

The similarities between unvoiced speech sounds and turbulent breath sounds were used to detect change in sound characteristics caused by narrowing of the upper airway (UA), similar to that occurring in obstructive sleep apnea (OSA). In 18 awake subjects, UA resistance (RAU), an index of UA narrowing, was measured simultaneously with breath sounds recording. Linear Prediction Coding was applied on turbulent inspiratory sounds drawn from low and high RAU conditions and K-means was used to cluster the resulting coefficients. The resulting 2 clusters were tested for agreement with the underlying RAU status. Distinct clusters were formed when RUA increased relatively high but not in cases with lower rise in RUA (P&#60;0.01 for all indicators.) This is the first work to show the utility of LPC in breath sounds analysis confirmed by an objective indicator or UA narrowing.

Comparison of in-laboratory and home diagnosis of sleep apnea using a cordless portable acoustic device

BACKGROUND AND OBJECTIVES Sleep apnea (SA) is a common, serious, but underdiagnosed condition. There is a need for more accessible and economic means of diagnosing SA in the home. The aim of this study was to test the validity of a cordless acoustic portable device (BresoDx™) for home diagnosis of SA compared with standard polysomnography (PSG). METHODS A total of 135 subjects underwent full overnight PSG and simultaneous recording of breath sounds by BresoDx in the sleep laboratory. Acoustic data extracted from BresoDx were analyzed using validated computer acoustic algorithms. The PSG-derived apnea-hypopnea index (AHI-p) and the acoustic AHI (AHI-a) were calculated and compared. A subset of 100 subjects used the device in a subsequent night in their home from which home AHI (AHI-h) was determined. RESULTS The correlation between AHI-a and simultaneous AHI-p was 95.2% and diagnostic accuracy of BresoDx ranged between 88.9% and 93.3% around AHI cutoffs of 5-15. In the home, AHI-h did not differ significantly from AHI-p (p = 0.60). Using an AHI-p cutoff ≥ 10 BresoDxs accuracy was 81%. Of the 100 subjects, 81 (81%) had low inter-night variability measured by a difference between home AHI-h and PSG AHI-p < 10 event/h, while 19% had higher inter-night variability. CONCLUSION AHI determined using BresoDx was in excellent agreement with simultaneous AHI-p. The majority of patients had a consistent AHI in their subsequent home study with very good overall diagnostic accuracy. We conclude that BresoDx is a reliable device for diagnosing SA that can be used by subjects, unattended in their own homes.

Relationship of respiratory sounds to alterations in the upper airway resistance

Respiratory sound analysis is a simple and noninvasive way to study the pathophysiology of the upper airway (UA). Recently, it has been used to diagnose partial or complete UA collapse in patients with obstructive sleep apnea (OSA). In this study, we investigated whether fluid accumulation in the neck alters the properties of respiratory sounds in temporal and spectral domains and whether the respiratory sounds analysis can be used to monitor variations in the physiology of the UA, as rejected by UA resistance (RU A). We recorded respiratory sounds and RU A from 19 individuals while awake. We applied lower body positive pressure (LBPP) to shift fluid out of the legs and into the neck, which increased RU A. We calculated first and second formants and energy of inspiratory sound segments. Our results show that during both control (no LBPP) and LBPP arms of the study, the extracted features were different for the sound segments corresponding to low and high RU A. Also, the features were different during control and LBPP arms of the study. With the application of support vector machine (SVM) based classifier, we were able to classify the sound segments into two groups of high/low resistance during control and LBPP arms and into two groups of control/LBPP when including all sound segments. The accuracies of non-linear SVM classifier were 74.5 ± 19.5%, 75.0 ± 15.4% and 77.1 ± 12.3% for the control arm, LBPP arm and between the arms, respectively. We also showed that during the LBPP arm, the variations in first formant of the sound segments corresponding to low and high RU A was much less than during the control arm. This indicates that with application of LBPP and accumulation of fluid in the neck, there are less variations in the morphology of the UA in response to changes in RU A, than during the control arm. These results indicate that acoustic analysis of respiratory sounds can be used to investigate physiology of the UA and how interventions can alter UA properties.

Adaptive segmentation and normalization of breathing acoustic data of subjects with obstructive sleep apnea

Breath sounds in patients with obstructive sleep apnea are very dynamic and variable signals due to their versatile nature. In this paper, we present an adaptive segmentation algorithm for these sounds. The algorithm divides the breath sounds into segments with similar amplitude levels. As the first step, the proposed scheme creates an envelope of the signal characterizing its long term amplitude variations. Then, K-means clustering is iteratively applied to detect borders between different segments in the envelope, which will then be used to segment and normalize the original signal.

The effect of sitting and calf activity on leg fluid and snoring

Prolonged sitting may promote leg fluid retention that redistributes to the neck during sleep and contributes to snoring. This could be attenuated by calf activity while sitting. In 16 healthy non-obese subjects we measured leg fluid volume (LFV) below the knees using bioelectrical impedance while sitting for 4h, snoring using a portable BresoDx™ device, and Mallampati grade. Using a double cross-over study design, subjects were randomized to one of two arms and crossed-over one week later: control arm - no calf exercise while sitting; intervention arm - calf contraction against a pedal resistance while sitting. The effects of sitting±calf activity on LFV and snoring were compared. We found that LFV increased by 216±101.0ml (p<0.0001) after sitting. Calf activity while sitting attenuated LFV by 53.8ml (p<0.0001) and, in all five subjects with severe upper airway narrowing (Mallampati grade IV), reduced snoring duration (from 357±132.9 to 116.2±72.1s/h, p=0.02) suggesting reduced overnight rostral fluid shift to the neck.

Subject independent identification of breath sounds components using multiple classifiers

Breath sounds have been shown very valuable for diagnosis of obstructive sleep apnea. In this study, we present a subject independent method for automatic classification of breath and related sounds during sleep. An experienced operator manually labelled segments of breath sounds from 11 sleeping subjects as: inspiration, expiration, inspiratory snoring, expiratory snoring, wheezing, other noise, and non-audible. Ten features were extracted and fed into 3 different classifiers: näıve Bayes, Support Vector Machine, and Random Forest. Leave-one-out method was used in which data from each subject, in turn, is evaluated using models trained with all other subject. Mean accuracy for concurrent classification of all 7 classes reached 85.4%. Mean accuracy for separating data into 2 classes, snoring and non-snoring, reached 97.8%. To our knowledge, these are the highest accuracies achieved in automatic classification of all breath sounds components concurrently and for snoring, in a subject independent model.

In-hospital diagnosis of sleep apnea in stroke patients using a portable acoustic device

PurposeSleep apnea (SA) is highly prevalent in post-stroke patients. Due to physical disability and relative inaccessibility of polysomnography (PSG) to test for SA, patients with stroke frequently remain undiagnosed and untreated. Portable SA monitoring can facilitate at-home or in-hospital testing for SA. However, portable SA monitoring is not recommended in those with complex medical conditions, such as stroke, due to the lack of validation of portable monitoring in such patients.MethodsThe objective of our study was to test the accuracy and feasibility of a portable single-channel acoustic device, BresoDx™ for quantifying the apnea-hypopnea index (AHI) and diagnosing SA in a post-stroke population. Patients who recently suffered a stroke and were undergoing rehabilitation in a stroke rehabilitation unit (SRU) underwent testing with BresoDx both simultaneously during attended PSG and unattended on the SRU.ResultsWe studied 23 stroke patients of whom 78% had SA (defined by AHI ≥15) on PSG. All of the patients tolerated the BresoDx. Using cutoff AHI of ≥15 by PSG to diagnose SA, BresoDx had sensitivity of 90.0%, specificity of 84.6%, and overall accuracy of 87.0% in the laboratory.ConclusionsThis study demonstrates that BresoDx is well tolerated and feasible to use in the post-stroke population where it was found to have excellent positive and negative predictive values for the diagnosis of SA.

Poster: WearCOPD - Monitoring COPD Patients Remotely using Smartwatches

Chronic Obstructive Pulmonary Disease (COPD) is a chronic lung disease that is characterized by airway obstruction, coughing, shortness of breath and increased sputum production. An acute exacerbation of COPD is a sudden worsening of the disease. Acute exacerbations result in more frequent and severe coughing and increased difficulty breathing. If not treated quickly, hospitalization may be required which is expensive and decreases patients quality of life. If untreated, an acute exacerbation can lead to death. We present WearCOPD, an application that uses a smartwatch and smartphone to continuously monitor physiological signs from patients with the goal of predicting exacerbations before they happen.