Access the full text.
Sign up today, get DeepDyve free for 14 days.
Loading next page...
/lp/de-gruyter/sleep-related-breathing-biomarkers-as-a-predictor-of-vital-functions-B9YvQCOoP0
- Publisher
- de Gruyter
- Copyright
- Copyright © 2017 by the
- ISSN
- 1895-9091
- eISSN
- 1896-530X
- DOI
- 10.1515/bams-2017-0003
- Publisher site
- See Article on Publisher Site
Abstract
Because an average human spends one third of his life asleep, it is apparent that the quality of sleep has an important impact on the overall quality of life. To properly understand the influence of sleep, it is important to know how to detect its disorders such as snoring, wheezing, or sleep apnea. The aim of this study is to investigate the predictive capability of a dual-modality analysis scheme for methods of sleep-related breathing disorders (SRBDs) using biosignals captured during sleep. Two logistic regressions constructed using backward stepwise regression to minimize the Akaike information criterion were extensively considered. To evaluate classification correctness, receiver operating characteristic (ROC) curves were used. The proposed classification methodology was validated with constructed Random Forests methodology. Breathing sounds and electrocardiograms of 15 study subjects with different degrees of SRBD were captured and analyzed. Our results show that the proposed classification model based on selected parameters for both logistic regressions determine the different types of acoustic events during sleep. The ROC curve indicates that selected parameters can distinguish normal versus abnormal events during sleep with high sensitivity and specificity. The percentage of prediction for defined SRBDs is very high. The initial assumption was that the quality of result is growing with the number of parameters included in the model. The best recognition reached is more than 89% of good predictions. Thus, sleep monitoring of breath leads to the diagnosis of vital function disorders. The proposed methodology helps find a way of snoring rehabilitation, makes decisions concerning future treatment, and has an influence on the sleep quality. Introduction Humans and all living beings need sleep. In recent years, many people have been suffering from sleep disorders caused by work-related stress, irregular lifestyle, or mental health problems [1]. The symptoms of sleep apnea sometimes appear unnoticeable that the person affected by sleep disorders may not note this as a serious issue. Sleepiness and sleep disorders occur because the brain wakes up when significant drops in airflow arrive [2]. As a result, the brain does not have long, uninterrupted periods of sleep, which are very important for a healthy lifestyle. Sleep is said to be "fragmented". Most patients with sleep apnea fall asleep quickly. Some, however, may complain of insomnia. It has to be mentioned that snoring is a common sign of obstructive sleep apnea (OSA) [3]. The influence of sleep quality and quantity on human health and well-being is not fully understood and remains still an open question. The proposed paper is focused on these issues. In practice, sleep disorders frequently overlap or coexist with each other. Most common disorders evaluated and managed in health care are sleep-related breathing disorders (SRBDs) [4]. Despite recent advances in its medical management, SRBD still leads to high mortality and morbidity. Sleep disorders (e.g. snoring, wheezing, and sleep apnea) are widespread among people of all ages although prevalent amid the male population, obese people with a high value of body mass index (BMI), people who suffered from a stroke, and people who have hypertension or other heart diseases [5]. Patients with SRBD are showing a higher prevalence of heart diseases [6]; thus, sleep monitoring often leads also to the diagnosis of cardiac diseases [2]. In addition, it is well known that chronic heart failure with sleep-disordered breathing is related to the appearance of atrial fibrillation [7]. Sleepdisordered breathing may impair ventricular repolarization and modulate the autonomic nervous system, which contributes to sudden cardiac death [8]. To improve the survival of cardiac patients, the identification of SRBD *Corresponding author: Klaudia Proniewska, Jagiellonian University Medical College, witej Anny 12, 31-008 Krakow, Poland, E-mail: klaudia.proniewska@uj.edu.pl Agnieszka Pregowska: Institute of Fundamental Technological Research, Polish Academy of Sciences, Pawinskiego 5B, 02-106 Warsaw, Poland Krzysztof Piotr Malinowski: Jagiellonian University Medical College, witej Anny 12, 31-008 Krakow, Poland 44Proniewska et al.: Sleep-related breathing biomarkers that contributes to increased mortality is required [9]. Simultaneously acquired acoustic and electrocardiographic (ECG) signals can be used to quantify respiratory obstructions during sleep. Previous research was focused on either sleep acoustic signal or ECG analysis [6]. Some preliminary results of the analysis of cardiac arrhythmia and SRBD were reported [10]. Mendez et al. have developed an autoregressive method model to screen sleep apneas from a single ECG lead [11]. Their methodology was based on RR intervals and the area under the QRS complex analysis. However, according to one group, the identification of sleep apnea can be also achieved using variations on heart rate (HR) variability (HRV) signals [12]. These studies [10, 12] were the motivation for our current work by reemphasizing the importance of accurate automatic classification of snore-related sounds for both clinical and research problems using both acoustic and ECG signal analysis [13]. The aim of the present study was to propose valid statistical models for the automatic detection of patients with SRBD suffering from ailments such as snoring and wheezing using acoustic derived biosignals correlated to these subjects' ECG signal acquired during sleep. The obtained results contribute to finding a correlation between two particular physiological signals, that is, vital functions such as acoustic signal of breathing and ECG signal represented by heart activity parameters acquired during the same period of breathing during sleep. The evaluation of the simultaneously recorded signals using different methods enables the assessment of variability parameters during sleep. We conducted a hypothesis that sleep-related breathing biomarkers are effective predictors of vital function disorders. Thus, three main expectations were taken into account. The first one was to derive a set of descriptors, that is, parameters and measurements of registered signals using relevant methods of biosignal analysis. The second one was to select a specific set of chosen parameters expected to provide high efficiency of breathing disorder recognition. The third one was to develop and validate predictive statistical models generally sensitive to signal variability. The paper [13] discusses the integration of these information related to the laboratory sleep scoring results made by the subject and supported by artificial intelligence [14]. Supported sleep evaluation system may aid in snoring rehabilitation and in making decisions concerning future treatment and have influence on the quality of night sleep. The developed methodology, based on two logistic regressions, divided breathing events into two groups: normal breathing against all breathing events and severe snoring against other breathing events. The standard method for the diagnosis of sleep disorders is polysomnography (PSG). Its main disadvantages are a full-night hospital stay in a specifically equipped bedroom. During this stay, the patient's physiological signals are recorded by attaching over 15 sensor channels. PSG is an expensive examination and thus is not suitable for large-scale application to screen large populations also because of limited PSG facilities around the world. Therefore, a clinical need for a reliable diagnosis of OSA at home is needed. Patients with heart failure often suffer from sleepdisordered breathing [15]. It was noticed that snoring is mostly accompanied by OSA and is recognized as its early symptom [16]. The major advantage of detecting snoring abnormalities is the application of low-cost noncontact equipment (e.g. in-air microphones). The results [17] suggest that OSA alters autonomic cardiovascular variability during sleep and wakefulness. However, a quantitative analysis of snoring sounds is currently not a common practice for OSA detection. Therefore, the potential of snoring analysis in the diagnosis of OSA is still not fully exploited. Several methods of snore segmentation and classification have been reported [18]. The method presented by these authors used Mel-frequency-cepstral coefficients (MFCCs). These features were applied in a hidden Markov model (HMM)-based calcification (8289% accuracy in identifying snore recordings). Another study was able to classify snores with an accuracy ranging from 86.8% to 97.3%. However, this study did not take into account the signalto-noise ratio of the signals on segmentation [19]. In this area, the development of new diagnostic indicators and risk assessment methods is particularly important. Moreover, the combination of features derived from the HRV and the EDR signals provides good classification results for sleep apnea detection [20]. The diagnosis of sleep disorders based on using two complementary modalities (i.e. acoustic and ECG signals collected simultaneously during sleep) may lead to specify the system scoring of SRBDs. Specific measurement procedures are described in "Materials and methods". The results are shown and discussed in "Results". "Discussion and conclusions" contains a brief discussion, concludes the paper, and gives future perspectives. Materials and methods Breathing sounds and ECGs of 15 subjects with different degrees of SRBD were captured and analyzed. A Proniewska et al.: Sleep-related breathing biomarkers45 microphone placed over the mouth, simultaneously with the ECG device placed on the thorax, recorded the biomedical signals over the entire night. Finally, the database of normal/disorder-related breathing sounds (different level of obstruction) with corresponding ECG signal was created. This database contained approximately 650 normal and disordered breathing short recordings called sound strips. Each recording duration included one breath episode. ECG signals The HR was derived from the Holter ECG method by an algorithm detecting a single heartbeat. The heartbeat is represented in the ECG by a QRS complex. The inverse of the time difference between the so-called normal heartbeats, QRS complexes resulting from sinus node depolarization, gives the HR. Then, the HR is sampled between consecutive intervals (NN intervals). In the next step, proper statistic parameters of the ECG signal are derived and used in recognition models of breathing disorders [21]. To acquire these data, the following equipment was used: Holter recorder AsPEKT 702, Aspel ECG recording - threechannel, battery-operated personal recorder with 12-bit 128 Hz, which is designed for long-record-term ECG data. Acoustic signals The registration of acoustic signals used the sampling rate of 44,100 Hz. Before saving the analog audio signal in digital form, anti-aliasing filtering was performed. Acoustic analysis techniques give information on the mechanism, loudness, and intensity and it is possible to extract relevant parameters to describe the signal. Filtering was also applied to remove unwanted noise from the signal frequency spectrum. The next step consisted of dividing the recorded acoustic signal, that is, sound track into smaller chunks called sound strips. Then, in each strip, the signal amplitude was normalized. Sequentially, each audio strip was divided into smaller segments of data called frames. Its purpose was to obtain the signal characteristics that can best describe the strip. The final stage of the study was the classification of acoustic signals. The registered acoustic signal was divided into sound strips, aiming at the size reduction of the signal and thus the reduction of the amount of data, computation time, and memory needed. For this reason, it is assumed that the shortest audio strips had a duration of one breath, typically 24 s. The use of segmentation creates frames used later with methods such as fast Fourier transform (FFT). Here, it should be noted that, for consecutive frames, the FFT method is contiguous. In the present study, the Hamming window was used. The Hamming window size is equal to the size of the frame; for the FFT algorithm, it was 32,768 samples with a sampling frequency of 44,100 Hz. This window size was sufficiently precise to identify the breathing acoustic signal. Snoring and other breathing events were analyzed in the frequency and time domain where these can be defined with a set of quantitative parameters [22]. The transformation of data from the time domain to the frequency domain was carried out by the short-time Fourier transform (STFT) algorithm. The sampling frequency of the analog-to-digital converter determines the maximum time duration of snoring sound. The frequency range of 12 kHz can completely describe the snoring phenomena. In this paper, we proposed a new support method for the interpretation of acoustic recordings, that is, acoustic strips correlated with the ECG signal for events that appeared during sleep at night. It is based on selected features from the conducted analysis. Collected ECGs and acoustic recordings were manually divided into breathing events. Breathing events during sleep were grouped in two types of acoustic events. In the first case, breathing events were divided into two groups: normal breathing versus all breathing events. In the second one, disordered breathing events were divided into different two groups: severe snoring versus other breathing disorders. All groups of acoustic events were subjected to statistical analysis to create models of disorders. One can observe that selected acoustic and ECG parameters change in the presence of breathing events. We considered 650 normal and disordered breathing short recordings. The vector of features, which was used in finding the best descriptors of breathing disorder identification, is presented in Table 1, Table 1:Descriptors of breathing disorders extracted from ECG and acoustic biosignals. Recording type ECGs Acoustic recordings Parameters Statistical time-domain parameters: SDNN, RMSSD, p50nn, SDi, and SDANN Spectral moments: M0, M1, M2 Formants: F1, F2, F3, F4 Frequency of formants: FF1, FF2, FF3, FF4 Power coefficients: W1, W2, W3 Cepstral coefficients: C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, C11, C12 Fundamental frequency: F0 Jitter and Shimmer 46Proniewska et al.: Sleep-related breathing biomarkers where SDNN is the standard deviation of all NN intervals (ms; normal-to-normal RR intervals); RMSSD is the square root of the mean of the squares of differences between adjacent NN intervals (ms); p50nn is the percentage of differences between adjacent NN intervals that are greater than 50 ms, a member of the larger pNNx family (%); SDi is the mean of the standard deviations of NN intervals in all 5-min segments of a 24-h recording (ms); and SDANN is the standard deviation of the averages of NN intervals in all 5-min segments of a 24-h recording (ms). Table 2:Probabilistic statistical classification model based on selected parameters for the first logistic regression. Parameter SDNN RMSSD pNN50 M1 M2 F3 F4 FF2 FF3 W1 W3 C2 C3 C4 C7 C8 C9 C10 C11 C12 Estimate value -0.032 0.019 -0.091 -1.54 0.013 0.0055 -0.0038 -0.09 0.75 109.2 37.2 -1.91 -2.78 -0.43 0.034 4.50 -6.29 -2.68 -6.88 -8.47 Standard error 0.015 0.014 0.035 0.82 0.006 0.0016 0.0011 0.13 0.24 32.3 31.3 0.51 0.57 0.33 0.68 0.87 1.37 1.61 2.17 2.24 2a 4.24 1.70 6.84 3.54 5.40 12.19 11.22 0.48 10.20 11.45 1.41 14.06 23.76 1.74 0.00 26.85 21.02 2.78 10.06 14.27 Prob>2b 0.0396c 0.1925 0.0089c 0.0599 0.0201c 0.0005c 0.0008c 0.4895 0.0014c 0.0007c 0.2356 0.0002c <0.0001c 0.1869 0.9601 <0.0001c <0.0001c 0.0955 0.0015c 0.0002c Results To designate the main parameters of information collected during sleep using two methods (i.e. acoustic signal and ECG signal dedicated analysis), the descriptive statistics was used for the prevalence of SRBDs. For the estimation of the parameters of a qualitative response model, the categorical dependent variable based on 34 predictor variables, cardiac and acoustic parameters, is described in Table 1. Based on this classification, two logistic regressions were proposed to show the types of probabilistic statistical classification models. In the case of the first logistic regression, breathing events were divided in two groups: normal breathing versus all breathing events. In second one, the events were divided into two groups: severe snoring versus other breathing disorders. In both logistic regressions, the Akaike information criterion (AIC), which is a measure of the relative quality of a statistical model, for a given set of data was used. Each logistic regression was constructed using backward stepwise regression to minimize the AIC. The probabilistic statistical classification model based on selected parameters is presented in Table 2 for the first logistic regression and Table 3 for the second one. These classifiers were evaluated using a receiver operating characteristic (ROC) curve [13, 23] and area under the graph. The ROC curves were used to evaluate the goodness-of-fit for binary classifiers if they correctly classified those events. The closer the curve follows the left border and then the top border of the ROC space, the more accurate is the classification. The closer the curve comes to the 45° diagonal of the ROC space, the classification tends towards random. Classification quality was measured by the area under the ROC curve, where an area of 1.0 represents a perfect classification and an area of 0.5 represents a worthless (random) classification. For performing the statistical analysis using methods described in the previous section, it was assumed that there are two classifications of breathing, as in Table 4. Computed by taking twice the difference in negative log-likelihoods between the fitted model and the reduced model that has only intercepts. bProbability of obtaining a greater 2 value by chance alone if the specified model fits no better than the model that includes only intercepts. cStatistically significant. Table 3:Probabilistic statistical classification model based on selected parameters for the second logistic regression. Parameter SDNN pNN50 M1 M2 F2 FF1 FF2 FF4 W1 C1 C4 C6 C7 C8 C9 C10 C11 Jitter Estimate value -0.03 -0.06 -4.64 0.029 0.007 -0.58 0.48 0.40 -60.49 -0.95 -4.21 3.73 -2.29 -6.5 2.32 9.1 3.05 -5.75 Standard error 0.01 0.02 0.84 0.005 0.002 0.14 0.17 0.12 21.88 0.26 0.86 0.83 0.73 1.3 1.02 1.8 1.56 5.89 2a 7.78 9.12 30.63 28.75 13.35 16.80 7.63 10.76 7.64 12.88 23.93 20.00 9.87 25.06 5.16 26.29 3.85 0.95 Prob>2b 0.0053c 0.0025c <0.0001c <0.0001c 0.0003c <0.0001c 0.0058c 0.0010c 0.0057c 0.0003c <0.0001c <0.0001c 0.0017c <0.0001c 0.0231c <0.0001c 0.0499c 0.3287 a Computed by taking twice the difference in negative log-likelihoods between the fitted model and the reduced model that has only intercepts. bProbability of obtaining a greater 2 value by chance alone if the specified model fits no better than the model that includes only intercepts. cStatistically significant. Proniewska et al.: Sleep-related breathing biomarkers47 Table 4:Classifications of breathing. First regression (normal breathing and disordered breathing) Negative: not having disordered breathing Positive: having disordered breathing Second regression (severe snoring and other disordered breathing) Negative: not having severe snoring Positive: having severe snoring Table 5:Percentage of prediction for defined SRBDs. Prediction for classes [%] First acoustic event Second acoustic event Third acoustic event Fourth acoustic event 92.86% 33.33% 84.00% 91.67% 96.55% Fifth acoustic event 73.91% 87.63% 88.31% 93.98% 92.71% Sixth acoustic event 100.00% 86.96% 100.00% 88.89% 89.47% Seventh acoustic event 80.00% 18.75% 94.74% 84.21% 89.47% All parameters (cardiac and acoustic) Best model 94.23% 94.83% 66.67% Cardiac parameters Best model 88.24% 71.31% 68.00% Acoustic parameters Best model 97.78% 96.12% 92.59% Mixed parameters (cardiac and acoustic without cepstral coefficients) Best model 94.38% 97.54% 80.00% Mixed parameters (cardiac and acoustic only with cepstral coefficients) Best model 100.00% 96.27% 96.43% The results of such analysis in both cases showed that the areas under the ROC curves are close to 1.0. In the case of normal breathing versus all observed breathing events, the area under the curve is equal to 0.99453. For severe snoring versus all other observed breathing events, the area under the curve is equal to 0.99025. This suggests that the selected parameters can recognize normal and abnormal breathing during sleep with high sensitivity and specificity [13]. The cutoff point is the best point of classification, maximizing sensitivity and specificity. There is also the optimum point: the minimization of error 1 and type 2. For the first ROC curve, the value of cutoff point is equal to 0.37; for the second ROC curve, the value of cutoff point is equal to 0.35. To validate our methodology, data mining classification with constructed Random Forests was applied. The Random Forests method was used in database with the observations of acoustic events: raw database with seven acoustic events (638 acoustic strips with correlated ECG signal). Statistic predictive models were created and compared to results with raw number of observations in the database. The Random Forests method takes eight properties such as numbers of predictor, numbers of tress, random test data proportion, subsample proportion, minimum number of cases, minimum number in child node, maximum number of levels, and maximum number of nodes, which were changed in different models. Initially, the Random Forests method was fed with all the data acquired and grouped into the type of acoustic event. In Table 5, data indicate that the obtained results are satisfactory. The percentage of prediction for defined SRBDs is very high. The initial assumption was that the quality of result is growing with the number of parameters included in the model. Based on the presented results, we conclude that the use of mixed parameters (cardiac and acoustic, only with cepstral coefficients) allows for best recognition, with more than 89% of good predictions. Discussion and conclusions Sleep monitoring has been evaluated by many researchers [2426]. The monitoring of lucid dreaming was presented [27]. The patients led a dream diary for the entire week before the time of the experiment and were trained nightly to access lucidity while asleep and to note their lucid dreams in the diary. Then, subjects spent a single night in the sleep laboratory. Sleep monitoring used EEG and a prolonged multiple sleep latency test. Moreover, the Sonomat system (i.e. a contactless monitoring system) used for the diagnosis of sleep-disordered breathing was described [28]. It records the two essential variables needed to detect and classify respiratory 48Proniewska et al.: Sleep-related breathing biomarkers events: breathing movement and airflow in the form of breath sounds. Jafarian et al. [25] developed a prototype of an acoustic respiratory monitor consisting of a series of microphones of various designs, a custom-built six-channel amplification system, a USB-based data acquisition system, and the ability to display information in the time domain and the frequency domain. In the paper [29], the management of sleep apnea by three nights of portable sleep monitoring was presented. The authors determined the diagnostic efficiency, consequent therapeutic decision-making, and costs of OSA diagnosis using PSG versus three consecutive studies of portable sleep monitoring in patients with mild to moderate suspicion of sleep apnea or with comorbidity that can mask OSA symptoms. This method includes a tracking algorithm to compensate the effect of head movements on the thermal imaging waveform extraction. On the contrary, the combination of ECG and respiration parameters provides good classification results for sleep apnea detection [30]. Since many years, the aims of continuous ECG monitoring were known and have expended from single lead rhythm monitoring to detection and classification of morphological disorders of ECG signal such as cardiac arrhythmias and dynamic assessment of ST-segment changes. The monitoring of HR behavior and response is especially important for cardiac patients, because an early autonomic nervous system dysfunction plays a dominant role in the progression and prognosis of their disease [31]. A comparison of five different algorithms for EEG signal analysis in artifact rejection for monitoring depth of anesthesia was made [32]. It was suggested that the complex index is a relatively more reliable indicator to design a real-time depth of anesthesia monitoring system for clinical application in comparison to the single indices. A monitoring system for cardiac care [33] could play a vital role in the early detection of a wide range of cardiac ailments, from a simple arrhythmia to life-threatening conditions such as myocardial infarction. In this study, we proposed a new methodology that is dedicated to the detection of respiratory disorders associated with patient's sleep using two complementary methods: simultaneous acquisition of ECG signals and acoustic signal at home. Prototype statistical models for the automatic detection of sleep disorders have been elaborated and experimentally validated. In the case of patients with SRBD, acoustic signal of breathing and ECG signal were observed across a 24-h period. For predicting the outcome of a categorical dependent variable based on 34 predictor variables, two logistic regressions were proposed. Our qualitative and quantitative results indicate that the probabilistic statistical classification model based on selected parameters for both logistic regressions determine different types of acoustic events (i.e. breathing disorders during sleep). The other issue we explored was ROC curves. The obtained results showed that the areas under the ROC curves are close to 1. This suggests that the selected parameters can distinguish normal versus abnormal events during sleep with high sensitivity and specificity. Moreover, it can be observed that selected acoustic and ECG parameters change in the presence of breathing events. A statistical analysis of ROC curves was performed to check a visible strong association between parameters. This suggests that the proposed parameters have a very high predictive value in differentiating the causes and severity of respiratory events during sleep. Present-day patient home-care monitoring systems are still in development. ECG acquisition modules integrated into a monitoring network are also able to store ECG data for retrospective analysis. Still, the optimal way to monitor sleep is electroencephalography, electrooculography, ECG, blood oxygen levels using pulse oximetry, and snoring using a microphone; finally, restlessness using a camera is complicated and long-term. The future development of integrated sleep monitoring could be very well based on the analysis of signals from both Holter monitoring and acoustic recording and employ the proposed models. Also, a more in-depth analysis of the correlation between sleep disorders and cardiovascular diseases prognostics could be undertaken. Acknowledgments: This paper has been partially supported by the Lesser Enterprise Center. The authors thank the AGH University of Science and Technology for outstanding technical assistance. Author contributions: The authors have accepted responsibility for the entire content of this submitted manuscript and approved submission. Research funding: The Lesser Enterprise Center. Employment or leadership: None declared. Honorarium: None declared. Competing interests: The funding organization(s) played no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the report for publication.
Journal
Bio-Algorithms and Med-Systems – de Gruyter
Published: Mar 1, 2017