- Email: [email protected]
Electrocardiogram-derived respiration in screening of sleep-disordered breathing
Available online at www.sciencedirect.com
Journal of Electrocardiology 44 (2011) 700 – 706 www.jecgonline.com
Electrocardiogram-derived respiration in screening of sleep-disordered breathing Saeed Babaeizadeh, PhD, a,⁎ Sophia H. Zhou, PhD, a Stephen D. Pittman, MSBME, b David P. White, MD c, d a
Advanced Algorithm Research Center, Philips Healthcare, Andover, MA, USA Philips Respironics, Philips Home Healthcare Solutions, Brighton, MA, USA c Sleep Medicine, Harvard Medical School, Boston, MA, USA d Division of Sleep Medicine, Brigham and Women's Hospital, Boston, MA, USA Received 9 May 2011 b
Abstract
Methods for assessment of sleep-disordered breathing (SDB), including sleep apnea, range from a simple questionnaire to complex multichannel polysomnography. Inexpensive and efficient electrocardiogram (ECG)–based solutions could potentially fill the gap and provide a new SDB screening tool. In addition to the heart rate variability (HRV)–based SDB screening method that we reported a year ago, we have developed a novel method based on ECG-derived respiration (EDR). This method derives the respiratory waveform by (a) measuring peak-to-trough QRS amplitude in a single-channel ECG, (b) removing outlier introduced by noise and artifacts, (c) interpolating the derived values, and (d) filtering values within the respiration rates of 5 and 25 cycles per minute. Each 30 seconds of the respiratory waveform is then classified as normal, SDB, or indeterminate epoch. The previously reported HRV-based method, applied at the same time, is based on power spectrum of heart rate over a sliding 6-minute time window to classify the middle 30-second epoch. We then combined the EDR- and HRV-based techniques to optimize the classification of each epoch. The combined method further improved the accuracy of SDB screening in an independent test database with annotated SDB epochs. The development database was from PhysioNet (n = 25 polysomnograms). The test database was from Sleep Health Centers in Boston (n = 1907 polysomnogram) where the SDB epochs (n = 1 538 222 epochs) were scored using American Academy of Sleep Medicine criteria. The first test was to classify every epoch in the evaluation data set. The combined EDR and HRV method classified 78% of the epochs as either normal or SDB and 22% as indeterminate, with a total accuracy of 88% for scored epochs (not indeterminate). The second test was to evaluate the SDB status for each patient. The algorithm correctly classified 71% of patients with either moderate-to-severe SDB or mild-to-no SDB. We believe that the ECG-based methods provide an efficient and inexpensive tool for SDB screening in both home and hospital settings and make SDB screening feasible in large populations. © 2011 Elsevier Inc. All rights reserved.
Keywords:
Sleep-disordered breathing (SDB); Sleep apnea screening; ECG-derived respiration (EDR)
Introduction Sleep-disordered breathing (SDB) refers to a group of disorders characterized by abnormalities of respiratory pattern or the quantity of ventilation during sleep. It is a highly prevalent disease that remains underdiagnosed. 1 Obstructive sleep apnea (OSA) is the most common such disorder. It is characterized by repeated episodes of ⁎ Corresponding author. Advanced Algorithm Research Center, Philips Healthcare, 3000 Minuteman Rd, MS0220, Andover, MA 01810, USA. E-mail address: [email protected] 0022-0736/$ – see front matter © 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.jelectrocard.2011.08.004
complete or partial collapse of the pharyngeal airway during sleep and, generally, the need to arouse to resume ventilation. The signs and symptoms of OSA include sleep fragmentation, hypoxemia, hypercapnia, and marked swings in intrathoracic pressure with associated increased sympathetic activity. 2 The high prevalence and wide spectrum of severity of OSA in adults have been well documented by several population-based cohort studies conducted in the United States, Europe, Australia, and Asia 3 that have shown that almost 1 in 5 adults has at least mild OSA and 1 in 15 has moderate or severe OSA. However, more than 85% of
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
patients with clinically important and treatable OSA have never been diagnosed. 4,5 Central sleep apnea (CSA) is another type of SDB characterized by the intermittent loss of all respiratory effort during sleep often leading to decreases in blood oxygen saturation. Central sleep apnea prevalence is increased in heart failure, left ventricular dysfunction, and stroke. 3 For both OSA and CSA, if the cessation of airflow lasts longer than 10 seconds, the episode is called apnea. If there is reduction in but not complete cessation of airflow (generally to b50% of normal), usually in association with a reduction in oxyhemoglobin saturation, the episode is called a hypopnea. 3 The standard method used to definitively diagnose patients with suspected apnea is called polysomnography. It generally requires spending at least 1 night in a sleep laboratory during which multiple physiological parameters are continuously recorded. Using these signals, disordered breathing and its effect on sleep and oxygenation can be precisely quantified. 6,7 The high cost and relative scarcity of diagnostic sleep laboratories may have contributed to the fact that sleep apnea is widely underdiagnosed. Hence, techniques to screen and diagnose patients for SDB without the need for a specialized sleep laboratory with fewer and simpler physiological signals, if effective, can be expected to have potential clinical utility. Several such techniques have been proposed including Epworth Sleepiness Scale, 8 the Berlin Questionnaire, 9 overnight oximetry, and combined limited respiratory assessment, electrocardiogram (ECG), and oximetry. 10 Specialized analysis of 24-hour ECG recordings also has been proposed as a possible screening tool. 11 In a joint initiative of Physionet 12 and the organizers of the 2000 Computers in Cardiology (CinC) conference, a competition was conducted to determine the efficacy of ECG-based methods for apnea detection using a representative database. A comparison of different algorithms participated in the challenge to detect sleep apnea from ECG recordings alone is described in Penzel et al. 13 We have previously reported an SDB screening method based on heart rate variability (HRV) 14 that was developed and evaluated on the same database used in the CinC challenge. This study is to report our continued effort in ECG-based SDB screening. We have developed an ECG-derived respiration (EDR) technique and a method to combine the EDR-based method with HRV-based methods for improved epoch classification. We report the performance of our HRV- and EDR-based and combined techniques on a large database collected in a clinical sleep laboratory. Compared with the database used in the CinC challenge and what was used in our previous work on HRV, 14 in this study, we have the additional value of a larger more diverse data set that likely better represents what is being obtained in standard sleep laboratories.
Materials and methods Database The database we used for developing our EDR-based algorithm is the publicly available St Vincent's University Hospital/University College Dublin sleep apnea database
701
15
from PhysioNet. This database contains 25 full overnight polysomnograms (PSGs) from adult subjects with suspected SDB. Subjects were randomly selected from patients referred to the sleep disorders clinic for possible diagnosis of OSA, CSA, or primary snoring. Subjects were older than 18 years, without known cardiac disease, autonomic dysfunction, and not on medication known to interfere with heart rate. For the development of our algorithm, we extracted the single-lead ECG measurements from the PSG recordings. The ECG signal was recorded at 128 samples per second (SPS) with the amplitude resolution of 0.2 μV per least significant bit. The standard sleep laboratory ECG electrode position was used. The subjects were 21 men and 4 women between 28 and 68 years old (50 ± 10 years) and has a body mass index between 25.1 and 42.5 kg/m 2 (31.6 ± 4.0 kg/m 2). The database included criterion standard SDB scoring with visual marking of the onset time and duration of respiratory events (obstructive, central, and mixed apneas and hypopneas and periodic breathing episodes) based on respiration and oxygen saturation signals, using amplitude criteria for airflow and desaturation. The database we used for testing was collected in sleep health centers in Boston. This database includes 1907 subjects with overnight PSGs. There was no available patient information related to cardiac disease, autonomic dysfunction, and medication. Polysomnogram recordings include 1 channel of ECG sampled at either 100 or 200 SPS with the amplitude resolution of 1 μV per least significant bit. Every recoding was scored for apnea and hypopnea episodes by a registered polysomnographic technician. For each PSG recording, based on scoring annotations, we calculated an apnea-hypopnea index (AHI) as the number of 30-second apnea and hypopnea epochs (epochs containing an apnea or hypopnea) per hour. Patients with an AHI of 15 and above were considered apneic. Patients with an AHI of less than 15 were considered normal (mild-to-no apnea). From 1907 patients in the evaluation data set, 989 were normal, and 918 were apneic. Electrocardiogram-derived respiration The algorithm to screen SDB by analyzing ECG involves signal conditioning, beat detection, and beat classification. The ECG algorithm analyzes the single-lead ECG to distinguish normal sinus beats from ectopic beats and erroneous measurements. In this study, we used Philips ST/AR arrhythmia algorithm 16 to detect and classify normal beats in the single-lead ECG recordings. One EDR value is then calculated for each normal QRS complex by measuring the peak to QRS trough amplitude. To adjust the EDR values that may be abnormally large or small because of noise and artifact, in a nonoverlapping, sliding 30-second window, EDR values were restricted to remain within the range of mean value ± 2 SDs. After adjusting the outliers, the EDR measurements are interpolated using cubic splines to derive a signal uniformly sampled at 8 SPS. To generate a smooth EDR waveform, the resulting signal is then filtered within respiration rates of 5 and 25 cycles per minute using a Chebyshev Type I bandpass filter. The sequential steps in the
702
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
process are illustrated in Fig. 1. An example of the derived EDR waveform with a flat pattern in a patient with CSA is shown in Fig. 2. The cessation of breathing because of CSA as seen in the airflow signal is remarkably visible in the EDR waveform as well. The EDR waveform's minima and maxima values were searched in each breath and used to derive a breath swing signal. An example of such signal is shown in Fig. 3. A supervised training technique was applied to derive classification features based on the shape of the breath swing waveform. A subset of such features with the strongest correlation with the presence or absence of SDB is selected. An exhaustive search technique was then applied to optimize the classifier performance by combining the classification features in the selected subset. By applying this feature selection scheme on the learning data set, 3 features were used for SDB classifications: (a) the number of times that the breath swing signal passes a threshold, (b) the percentage of time that the breath swing signal is below that threshold, and (c) respiration rate. The bottom panel of Fig. 3 shows the breath swing signal for the CSA episode shown in Fig. 2. A classifier, trained on the learning set, uses the 3 classification features mentioned above to classify the 30-second epoch as SDB, normal, or indeterminate. This classifier is basically a set of rules using the 3 features. For example, an epoch is classified as apnea if the number of threshold crossing is between 1 and 4 and more than 20% of the time the breath swing in below that threshold. The threshold for each epoch is half the range of the breath swing signal. Furthermore, if the respiration rate, at any point in an epoch, has dropped to
less than half the median respiration rate in that epoch, the epoch will be classified as apnea. An epoch will be classified as indeterminate if, because of noise and artifact, there is no reliable signal in the 30-second epoch to analyze. In addition to this EDR-based technique, we have developed and previously published an HRV-based SDB screening method. 14 The next step is to combine these 2 independent techniques, EDR and HRV, to improve classification of each patient as normal or as having SDB in addition to obtain reliable counts of individual epochs. We applied the EDR- and HRV-based techniques independently to classify epochs, combined epoch classifications, calculated an algorithm-based SDB index (ASI) as the number of 30second SDB epochs per hour, and used the index to classify the patients as SDB or normal. The key in this technique is the combination method. There are several possible ways to combine the output of EDR- and HRV-based methods. The best performance in the learning set was obtained for the combination technique summarized in Table 1.
Results We tested the performance of our 3 algorithms (EDR alone, HRV alone, and combined EDR and HRV) for both quantitative and qualitative SDB screening. Quantitative assessment of SDB Table 2 summarizes the performance of the algorithms in classifying all 1 538 222 epochs of the 1907 ECG recordings
Fig. 1. Sequential steps used for derivation of the EDR signal from single-lead ECG. Peak-to-trough QRS amplitude (EDR value) is measured for each normal QRS complex. Electrocardiogram-derived respiration values that are clustered and deviated over a set threshold are excluded. The EDR values are then interpolated using cubic splines, and the interpolated EDR signal is filtered within respiration rates of 5 to 25 cycles per minute to retain the respiratory component.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
703
Fig. 2. Electrocardiogram-derived respiration waveform with a flat pattern due to SDB in a patient with CSA. The top panel shows the ECG recording of the subject; the middle panel, the recorded airflow; and the bottom panel, the calculated EDR waveform. This snapshot includes a 10-second episode of CSA with its onset and offset annotated on the airflow waveform. The cessation of breathing due to CSA is remarkably visible in the EDR waveform.
in the evaluation data set. The algorithm classifies each epoch as SDB, normal, or indeterminate. It is a known fact that the ECG-based algorithms have difficulty detecting hypopnea epochs because such events usually do not significantly alter the ECG. Including hypopnea epochs in the test and punishing the algorithm for classifying them as normal, the combined method classified 78% of the epochs as normal or SDB (not indeterminate) with 88% accuracy.
Excluding the hypopnea epochs, meaning neither reward nor punish the algorithm for classifying them, the combination technique classified 80% of the epochs as normal or SDB with an accuracy of 98%. Screening SDB For this test, the algorithm uses the ECG signal to determine if important SDB is present or not to classify the patient as normal or as having SDB (using an AHI of 15 as the threshold). Using a threshold for ASI derived from the Table 1 Statistically optimal decision rules to combine EDR- and HRV-based techniques
Fig. 3. The breath swing signal (bottom panel) is calculated by using the minima and maxima of each breath from the EDR waveform (top panel). Three features are extracted from the breath swing for EDR-based SDB epoch classification: (a) the number of times that the breath swing signal passes a threshold, (b) the percentage of time that the breath swing signal is below the threshold, and (c) respiratory rate.
EDR and HRV combined
EDR alone
HRV alone
SDB SDB Normal Normal Indeterminate Indeterminate Indeterminate Indeterminate Indeterminate
SDB SDB Normal Normal Indeterminate Normal SDB Indeterminate Indeterminate
SDB Indeterminate Normal Indeterminate Indeterminate SDB Normal Normal SDB
For instance, if both EDR and HRV classify a 30-second epoch as normal, SDB, or indeterminate, then combined EDR and HRV classification is the same as the individual EDR and HRV determinations. If EDR marks an epoch as normal or SDB and if HRV classifies the same epoch as indeterminate, then the combined method agrees with EDR classification. If EDR and HRV have controversial classifications, the final classification done by the combined method is indeterminate.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
Table 2 Performance for quantitative assessment of SDB Method
Classified (%)
Accuracy (%)
EDR alone HRV alone EDR and HRV combined
96 99 78
79 82 88
The evaluation data set includes a total of 1 538 222 epochs. Hypopnea epochs are included as apnea. The column “Classified” shows the percentage of the epochs classified as normal or SDB (and not indeterminate), and the column “Accuracy” shows the percentage match between the algorithms' output and manual scoring done by human experts.
receiver operating characteristic (ROC) curve for the learning set, the combination technique classified recordings with an ASI of larger than 1.5 as SDB and others as normal. As seen from Table 3, in the evaluation data set, the algorithm correctly classified 1355 of 1907 subjects with a total accuracy of 71%. In this test, the hypopnea epochs were included as they contributed to the AHI used for creating the criterion standard annotation for each patient. Discussion In classifying every epoch in the evaluation data set, the combined EDR and HRV techniques achieved accuracy of 88% for scoring 78% of the epochs as either normal or SDB. Testing the SDB status for each patient, this algorithm correctly classified 71% of patients with either moderate-tosevere SDB or mild-to-no SDB. In this article, we summarized our effort to provide a reliable screening technique for SDB based solely on analysis of a single-lead ECG recording. Our approach was based on the expectation that SDB affects respiratory pattern and that respiration pattern affects both heart rate and ECG morphology. Few techniques to detect sleep apnea using HRV have been previously proposed by other researchers. 17-19 Respiration effects on ECG morphology have been studied by several groups to derive the respiratory signal by estimating the heart electrical axis using 2, 20 3, 21 and even 8 22 ECG leads. Our EDR technique requires only 1 channel of ECG. This single-lead EDR method works best if the lead axis is significantly different from the mean electrical axis to obtain a relatively large signal. Compared with measuring QRS area for estimating EDR as introduced by others, 20 our technique
has an advantage of being insensitive to ECG baseline wander. Therefore, removal of baseline wander, which requires additional processing and could be challenging at times, is unnecessary. For classifying each epoch, the EDR algorithm needs only 30 seconds of signal at 1 time, and the HRV-based algorithm requires 6 minutes of ECG to mark the 30second epoch in the middle of the 6-minute window. 14 Because a short 30-second window is more likely to be corrupted completely by noise or artifact than the long 6minute window, the EDR-based method is more prone to noise compared with the HRV-based method. Therefore, more epochs are classified as indeterminate by the EDRbased method as compared with the HRV-based method. It is important to know that SDB assessment can be difficult even for human experts. There is considerable variability in SDB scoring when 2 qualified technologist score the same record or 1 expert scores the same record in different times. The basis of our method is that the altered respiratory patterns in patients with SDB affect both heart rate and ECG morphology. If such alteration is not detectable or eliminated by other factors, such as medication or cardiac disease or autonomic dysfunction, the ECG-based method will not work. The SDB screening techniques introduced in this study rely on a method for ECG beat detection and classification from single-lead ECG recordings. The quality of ECG beat detector and classifier impacts the performance of ECGbased techniques. Thus, the limitations in the evaluation data set should be pointed out. One major limitation is the relatively low quality of ECG recordings. The 1-channel ECG was recorded as part of the PSG study, and the quality of the ECG recording was rather poor. The HRV method does not work on patients who either have irregular heart rhythm or are on rhythm- and rate-control drugs or
Receiver Operating Characteristic (ROC) 1
Sensitivity
704
0.60
ASI threshold=1.5
0.5
Table 3 Classification performance for SDB screening by EDR- and HRV-based methods when used alone or combined Method
Sensitivity Specificity Positive Negative Accuracy predictive predictive value value
EDR alone 45 HRV alone 74 EDR and HRV 60 combined
76 62 82
64 64 75
60 72 69
61 68 71
The evaluation data set included 989 normal and 918 apneic patients marked by using AHI threshold of 15 using the criterion standard SDB annotations from full PSG. The combined method has significantly higher specificity, positive predictive value, and the total accuracy. All the numbers are reported in percentage (%).
0
0
0.18
0.5 1-Specificity
1
Fig. 4. Receiver operating characteristic curve for the performance of combined EDR and HRV algorithm on the evaluation data set. It is noted that the ASI threshold of 1.5 corresponds to the sensitivity of 60% and specificity of 82% as reported in Table 3. For hospital use, one can choose a different threshold to achieve higher sensitivity and lower specificity. For home screening, it may be desirable to change the threshold to make the algorithm less sensitive but more specific.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
devices. We did not exclude subjects with irregular heart rate, other arrhythmias, or any cardiac disease. The HRV method depends on the sampling frequency, and it is better to have sampling rate of 500 SPS or higher. In our evaluation data set, ECGs have low sampling rates of 100 or 200 SPS. The EDR method is sensitive to amplitude resolution. The evaluation data set has poor amplitude resolution. Furthermore, an accurate estimated EDR signal can only be obtained if heart rate is high enough to prevent sub-Nyquist frequency undersampling. Therefore, extreme bradycardia can result in frequency aliasing; hence, unreliable EDR waveform. Similar undersampling can occur if, due to arrhythmia or other cardiac diseases, the patient has many abnormal QRS complexes for which an EDR value will not be calculated. It is also known that the ECG axis, hence, EDR morphology, changes as a function of body position and posture. This change, however, if limited, may not significantly affect the epoch classifications because the 30-second epochs are being classified independently. We did not exclude these potentially challenging ECG recordings from our database. A low sampling rate may produce a jitter in the estimated R-wave fiducial point, which alters the spectrum considerably. The optimal range is 250 to 500 SPS or perhaps even higher, 23 which is above the 100 SPS sampling rate of most studies assessed here. The same jitter in the R-wave fiducial point potentially affects the EDR values measured as beat swings that may also be affected by poor amplitude resolution of ECG recordings. More study is needed to investigate the potential effects of ECG recordings quality on our SDB monitoring algorithms. The current performance results may improve on higher quality ECG data, in particular, with higher sampling rate and amplitude resolution. However, our techniques performed reasonably well even under the current challenging conditions. In the clinical setting, the outcome of the algorithm can be adjusted by using different ASI limits. It may be better to select 2 separate ASI limits for normal and SDB and mark the group falling in between as indeterminate or borderline. This may miss some patients with milder SDB but will usually identify those with moderate to severe disordered breathing. When the algorithm is used for SDB screening, it is possible to change its sensitivity and specificity by adjusting the ASI threshold. Fig. 4 shows the ROC curve for the combined EDR and HRV on the evaluation data set. It is noted that the ASI threshold of 1.5 corresponds to the sensitivity of 60% and specificity of 82% as reported in Table 3. For hospital SDB screening where higher sensitivity is desired, a different threshold can be used to achieve a higher sensitivity. For home SDB screening where a higher specificity is desired, an adjusted threshold can certainly lead to higher specificity with reduced sensitivity. Another way to improve the detection accuracy could be to combine the ECG signal with other easy-to-obtain signals such as pulse oximetry and/or limited respiratory measurements. The multiparameter approach for automatic screening of SDB will be the subject of ongoing research.
705
Conclusion Our 1-channel ECG-only SDB screening algorithm provides reasonably accurate SDB detection and quantification. Because ECG is readily available, noninvasive, and low cost, this approach will make it possible for SDB screening in a large population in both hospital and home care environments. This would be quite convenient for cardiac patients who have ECG monitoring in the hospital or at home. Thus, SDB screening can be done without additional cost. References 1. Cartwright R. Obstructive sleep apnea: a sleep disorder with major effects on health. Dis Mon 2001;47:109. 2. Epstein LJ, Kristo D, Strollo Jr PJ, et al. Clinical guideline for the evaluation, management and long-term care of obstructive sleep apnea in adults. J Clin Sleep Med 2009;5:263. 3. Somers VK, White DP, Amin R, et al. AHA/ACCF Scientific Statement. Sleep apnea and cardiovascular disease. Circulation 2008;118:1080. 4. Kapur V, Strohl KP, Redline S, Iber C, O'Connor G, Nieto J. Underdiagnosis of sleep apnea syndrome in U.S. communities. Sleep Breath 2002;6:49. 5. Young T, Evans L, Finn L, Palta M. Estimation of the clinically diagnosed proportion of sleep apnea syndrome in middle-aged men and women. Sleep 1997;20:705. 6. Iber C, Ancoli-Israel S, Chesson A, Quan SF. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications. 1st ed. Westchester, IL: American Academy of Sleep Medicine; 2007. 7. Caples SM, Somers VK, Rosen CL, et al. The scoring of cardiac events during sleep. J Clin Sleep Med 2007;3:147. 8. Miletin MS, Hanly PJ. Measurement properties of the Epworth sleepiness scale. Sleep Med 2003;4:195. 9. Netzer NC, Stoohs RA, Netzer CM, Clark K, Strohl KP. Using the Berlin Questionnaire to identify patients at risk for the sleep apnea syndrome. Ann Intern Med 1999;131:485. 10. Abraham WT, Trupp RJ, Phillips B, et al. Validation and clinical utility of a simple in-home testing tool for sleep-disordered breathing and arrhythmias in heart failure: results of the Sleep Events, Arrhythmias, and Respiratory Analysis in Congestive Heart Failure (SEARCH) study. Conges. Heart Fail 2006;12:241. 11. Roche F, Gaspoz JM, Court-Fortune I, et al. Screening of obstructive sleep apnea syndrome by heart rate variability analysis. Circulation 1999;100:1411. 12. Goldberger AL, Amaral LA, Glass L, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 2000;101:E215. 13. Penzel T, McNames J, Murray A, de Chazal P, Moody G, Raymond B. Systematic comparison of different algorithms for apnoea detection based on electrocardiogram recordings. Med Biol Eng Comput 2002;40:402. 14. Babaeizadeh S, White DP, Pittman SD, Zhou SH. Automatic detection and quantification of sleep apnea using heart rate variability. J Electrocardiol 2010;43:535. 15. http://www.physionet.org/physiobank/database/ucddb/. 16. http://www.healthcare.philips.com/main/products/patient_monitoring/ products/st_ar/. 17. Guilleminault C, Connolly S, Winkle R, et al. Cyclical variation of the heart rate in sleep apnoea syndrome. Mechanisms, and usefulness of 24 h electrocardiography as a screening technique. Lancet 1984;1:126. 18. Penzel T, Amend G, Meinzer K, Peter JH, Von Wichert P. MESAM: a heart rate and snoring recorder for detection of obstructive sleep apnea. Sleep 1990;13:175. 19. Hilton MF, Bates RA, Godfrey KR, Chappell MJ, Cayton RM. Evaluation of frequency and time-frequency spectral analysis of heart rate variability as a diagnostic marker of the sleep apnoea syndrome. Med Biol Eng Comput 1999;37:760. 20. Moody GB, Mark RG, Zoccola A, Mantero S. Derivation of respiratory signals from multi-lead ECGs. Proceedings of Computers in Cardiology 1985. IEEE Press 1985;12:113.
706
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
21. Pinciroli F, Pozzi G, Rossi R. Processing electrocardiograms towards respiratory signals. Proc. of the 10th IEEE Engineering in Medicine & Biology Society Int. Conf. New Orleans, USA; 1988. p. 139. 22. Travaglini A, Lamberti C, De Bie J. Respiratory of signal derived from eight-lead ECG. Proceedings of Computers in Cardiology 1998. IEEE Press 1998;2:65.
23. Task Force of the European Society of Cardiology the North American Society of Pacing Electrophysiology. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Circulation 1996; 93:1043.
Journal of Electrocardiology 44 (2011) 700 – 706 www.jecgonline.com
Electrocardiogram-derived respiration in screening of sleep-disordered breathing Saeed Babaeizadeh, PhD, a,⁎ Sophia H. Zhou, PhD, a Stephen D. Pittman, MSBME, b David P. White, MD c, d a
Advanced Algorithm Research Center, Philips Healthcare, Andover, MA, USA Philips Respironics, Philips Home Healthcare Solutions, Brighton, MA, USA c Sleep Medicine, Harvard Medical School, Boston, MA, USA d Division of Sleep Medicine, Brigham and Women's Hospital, Boston, MA, USA Received 9 May 2011 b
Abstract
Methods for assessment of sleep-disordered breathing (SDB), including sleep apnea, range from a simple questionnaire to complex multichannel polysomnography. Inexpensive and efficient electrocardiogram (ECG)–based solutions could potentially fill the gap and provide a new SDB screening tool. In addition to the heart rate variability (HRV)–based SDB screening method that we reported a year ago, we have developed a novel method based on ECG-derived respiration (EDR). This method derives the respiratory waveform by (a) measuring peak-to-trough QRS amplitude in a single-channel ECG, (b) removing outlier introduced by noise and artifacts, (c) interpolating the derived values, and (d) filtering values within the respiration rates of 5 and 25 cycles per minute. Each 30 seconds of the respiratory waveform is then classified as normal, SDB, or indeterminate epoch. The previously reported HRV-based method, applied at the same time, is based on power spectrum of heart rate over a sliding 6-minute time window to classify the middle 30-second epoch. We then combined the EDR- and HRV-based techniques to optimize the classification of each epoch. The combined method further improved the accuracy of SDB screening in an independent test database with annotated SDB epochs. The development database was from PhysioNet (n = 25 polysomnograms). The test database was from Sleep Health Centers in Boston (n = 1907 polysomnogram) where the SDB epochs (n = 1 538 222 epochs) were scored using American Academy of Sleep Medicine criteria. The first test was to classify every epoch in the evaluation data set. The combined EDR and HRV method classified 78% of the epochs as either normal or SDB and 22% as indeterminate, with a total accuracy of 88% for scored epochs (not indeterminate). The second test was to evaluate the SDB status for each patient. The algorithm correctly classified 71% of patients with either moderate-to-severe SDB or mild-to-no SDB. We believe that the ECG-based methods provide an efficient and inexpensive tool for SDB screening in both home and hospital settings and make SDB screening feasible in large populations. © 2011 Elsevier Inc. All rights reserved.
Keywords:
Sleep-disordered breathing (SDB); Sleep apnea screening; ECG-derived respiration (EDR)
Introduction Sleep-disordered breathing (SDB) refers to a group of disorders characterized by abnormalities of respiratory pattern or the quantity of ventilation during sleep. It is a highly prevalent disease that remains underdiagnosed. 1 Obstructive sleep apnea (OSA) is the most common such disorder. It is characterized by repeated episodes of ⁎ Corresponding author. Advanced Algorithm Research Center, Philips Healthcare, 3000 Minuteman Rd, MS0220, Andover, MA 01810, USA. E-mail address: [email protected] 0022-0736/$ – see front matter © 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.jelectrocard.2011.08.004
complete or partial collapse of the pharyngeal airway during sleep and, generally, the need to arouse to resume ventilation. The signs and symptoms of OSA include sleep fragmentation, hypoxemia, hypercapnia, and marked swings in intrathoracic pressure with associated increased sympathetic activity. 2 The high prevalence and wide spectrum of severity of OSA in adults have been well documented by several population-based cohort studies conducted in the United States, Europe, Australia, and Asia 3 that have shown that almost 1 in 5 adults has at least mild OSA and 1 in 15 has moderate or severe OSA. However, more than 85% of
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
patients with clinically important and treatable OSA have never been diagnosed. 4,5 Central sleep apnea (CSA) is another type of SDB characterized by the intermittent loss of all respiratory effort during sleep often leading to decreases in blood oxygen saturation. Central sleep apnea prevalence is increased in heart failure, left ventricular dysfunction, and stroke. 3 For both OSA and CSA, if the cessation of airflow lasts longer than 10 seconds, the episode is called apnea. If there is reduction in but not complete cessation of airflow (generally to b50% of normal), usually in association with a reduction in oxyhemoglobin saturation, the episode is called a hypopnea. 3 The standard method used to definitively diagnose patients with suspected apnea is called polysomnography. It generally requires spending at least 1 night in a sleep laboratory during which multiple physiological parameters are continuously recorded. Using these signals, disordered breathing and its effect on sleep and oxygenation can be precisely quantified. 6,7 The high cost and relative scarcity of diagnostic sleep laboratories may have contributed to the fact that sleep apnea is widely underdiagnosed. Hence, techniques to screen and diagnose patients for SDB without the need for a specialized sleep laboratory with fewer and simpler physiological signals, if effective, can be expected to have potential clinical utility. Several such techniques have been proposed including Epworth Sleepiness Scale, 8 the Berlin Questionnaire, 9 overnight oximetry, and combined limited respiratory assessment, electrocardiogram (ECG), and oximetry. 10 Specialized analysis of 24-hour ECG recordings also has been proposed as a possible screening tool. 11 In a joint initiative of Physionet 12 and the organizers of the 2000 Computers in Cardiology (CinC) conference, a competition was conducted to determine the efficacy of ECG-based methods for apnea detection using a representative database. A comparison of different algorithms participated in the challenge to detect sleep apnea from ECG recordings alone is described in Penzel et al. 13 We have previously reported an SDB screening method based on heart rate variability (HRV) 14 that was developed and evaluated on the same database used in the CinC challenge. This study is to report our continued effort in ECG-based SDB screening. We have developed an ECG-derived respiration (EDR) technique and a method to combine the EDR-based method with HRV-based methods for improved epoch classification. We report the performance of our HRV- and EDR-based and combined techniques on a large database collected in a clinical sleep laboratory. Compared with the database used in the CinC challenge and what was used in our previous work on HRV, 14 in this study, we have the additional value of a larger more diverse data set that likely better represents what is being obtained in standard sleep laboratories.
Materials and methods Database The database we used for developing our EDR-based algorithm is the publicly available St Vincent's University Hospital/University College Dublin sleep apnea database
701
15
from PhysioNet. This database contains 25 full overnight polysomnograms (PSGs) from adult subjects with suspected SDB. Subjects were randomly selected from patients referred to the sleep disorders clinic for possible diagnosis of OSA, CSA, or primary snoring. Subjects were older than 18 years, without known cardiac disease, autonomic dysfunction, and not on medication known to interfere with heart rate. For the development of our algorithm, we extracted the single-lead ECG measurements from the PSG recordings. The ECG signal was recorded at 128 samples per second (SPS) with the amplitude resolution of 0.2 μV per least significant bit. The standard sleep laboratory ECG electrode position was used. The subjects were 21 men and 4 women between 28 and 68 years old (50 ± 10 years) and has a body mass index between 25.1 and 42.5 kg/m 2 (31.6 ± 4.0 kg/m 2). The database included criterion standard SDB scoring with visual marking of the onset time and duration of respiratory events (obstructive, central, and mixed apneas and hypopneas and periodic breathing episodes) based on respiration and oxygen saturation signals, using amplitude criteria for airflow and desaturation. The database we used for testing was collected in sleep health centers in Boston. This database includes 1907 subjects with overnight PSGs. There was no available patient information related to cardiac disease, autonomic dysfunction, and medication. Polysomnogram recordings include 1 channel of ECG sampled at either 100 or 200 SPS with the amplitude resolution of 1 μV per least significant bit. Every recoding was scored for apnea and hypopnea episodes by a registered polysomnographic technician. For each PSG recording, based on scoring annotations, we calculated an apnea-hypopnea index (AHI) as the number of 30-second apnea and hypopnea epochs (epochs containing an apnea or hypopnea) per hour. Patients with an AHI of 15 and above were considered apneic. Patients with an AHI of less than 15 were considered normal (mild-to-no apnea). From 1907 patients in the evaluation data set, 989 were normal, and 918 were apneic. Electrocardiogram-derived respiration The algorithm to screen SDB by analyzing ECG involves signal conditioning, beat detection, and beat classification. The ECG algorithm analyzes the single-lead ECG to distinguish normal sinus beats from ectopic beats and erroneous measurements. In this study, we used Philips ST/AR arrhythmia algorithm 16 to detect and classify normal beats in the single-lead ECG recordings. One EDR value is then calculated for each normal QRS complex by measuring the peak to QRS trough amplitude. To adjust the EDR values that may be abnormally large or small because of noise and artifact, in a nonoverlapping, sliding 30-second window, EDR values were restricted to remain within the range of mean value ± 2 SDs. After adjusting the outliers, the EDR measurements are interpolated using cubic splines to derive a signal uniformly sampled at 8 SPS. To generate a smooth EDR waveform, the resulting signal is then filtered within respiration rates of 5 and 25 cycles per minute using a Chebyshev Type I bandpass filter. The sequential steps in the
702
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
process are illustrated in Fig. 1. An example of the derived EDR waveform with a flat pattern in a patient with CSA is shown in Fig. 2. The cessation of breathing because of CSA as seen in the airflow signal is remarkably visible in the EDR waveform as well. The EDR waveform's minima and maxima values were searched in each breath and used to derive a breath swing signal. An example of such signal is shown in Fig. 3. A supervised training technique was applied to derive classification features based on the shape of the breath swing waveform. A subset of such features with the strongest correlation with the presence or absence of SDB is selected. An exhaustive search technique was then applied to optimize the classifier performance by combining the classification features in the selected subset. By applying this feature selection scheme on the learning data set, 3 features were used for SDB classifications: (a) the number of times that the breath swing signal passes a threshold, (b) the percentage of time that the breath swing signal is below that threshold, and (c) respiration rate. The bottom panel of Fig. 3 shows the breath swing signal for the CSA episode shown in Fig. 2. A classifier, trained on the learning set, uses the 3 classification features mentioned above to classify the 30-second epoch as SDB, normal, or indeterminate. This classifier is basically a set of rules using the 3 features. For example, an epoch is classified as apnea if the number of threshold crossing is between 1 and 4 and more than 20% of the time the breath swing in below that threshold. The threshold for each epoch is half the range of the breath swing signal. Furthermore, if the respiration rate, at any point in an epoch, has dropped to
less than half the median respiration rate in that epoch, the epoch will be classified as apnea. An epoch will be classified as indeterminate if, because of noise and artifact, there is no reliable signal in the 30-second epoch to analyze. In addition to this EDR-based technique, we have developed and previously published an HRV-based SDB screening method. 14 The next step is to combine these 2 independent techniques, EDR and HRV, to improve classification of each patient as normal or as having SDB in addition to obtain reliable counts of individual epochs. We applied the EDR- and HRV-based techniques independently to classify epochs, combined epoch classifications, calculated an algorithm-based SDB index (ASI) as the number of 30second SDB epochs per hour, and used the index to classify the patients as SDB or normal. The key in this technique is the combination method. There are several possible ways to combine the output of EDR- and HRV-based methods. The best performance in the learning set was obtained for the combination technique summarized in Table 1.
Results We tested the performance of our 3 algorithms (EDR alone, HRV alone, and combined EDR and HRV) for both quantitative and qualitative SDB screening. Quantitative assessment of SDB Table 2 summarizes the performance of the algorithms in classifying all 1 538 222 epochs of the 1907 ECG recordings
Fig. 1. Sequential steps used for derivation of the EDR signal from single-lead ECG. Peak-to-trough QRS amplitude (EDR value) is measured for each normal QRS complex. Electrocardiogram-derived respiration values that are clustered and deviated over a set threshold are excluded. The EDR values are then interpolated using cubic splines, and the interpolated EDR signal is filtered within respiration rates of 5 to 25 cycles per minute to retain the respiratory component.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
703
Fig. 2. Electrocardiogram-derived respiration waveform with a flat pattern due to SDB in a patient with CSA. The top panel shows the ECG recording of the subject; the middle panel, the recorded airflow; and the bottom panel, the calculated EDR waveform. This snapshot includes a 10-second episode of CSA with its onset and offset annotated on the airflow waveform. The cessation of breathing due to CSA is remarkably visible in the EDR waveform.
in the evaluation data set. The algorithm classifies each epoch as SDB, normal, or indeterminate. It is a known fact that the ECG-based algorithms have difficulty detecting hypopnea epochs because such events usually do not significantly alter the ECG. Including hypopnea epochs in the test and punishing the algorithm for classifying them as normal, the combined method classified 78% of the epochs as normal or SDB (not indeterminate) with 88% accuracy.
Excluding the hypopnea epochs, meaning neither reward nor punish the algorithm for classifying them, the combination technique classified 80% of the epochs as normal or SDB with an accuracy of 98%. Screening SDB For this test, the algorithm uses the ECG signal to determine if important SDB is present or not to classify the patient as normal or as having SDB (using an AHI of 15 as the threshold). Using a threshold for ASI derived from the Table 1 Statistically optimal decision rules to combine EDR- and HRV-based techniques
Fig. 3. The breath swing signal (bottom panel) is calculated by using the minima and maxima of each breath from the EDR waveform (top panel). Three features are extracted from the breath swing for EDR-based SDB epoch classification: (a) the number of times that the breath swing signal passes a threshold, (b) the percentage of time that the breath swing signal is below the threshold, and (c) respiratory rate.
EDR and HRV combined
EDR alone
HRV alone
SDB SDB Normal Normal Indeterminate Indeterminate Indeterminate Indeterminate Indeterminate
SDB SDB Normal Normal Indeterminate Normal SDB Indeterminate Indeterminate
SDB Indeterminate Normal Indeterminate Indeterminate SDB Normal Normal SDB
For instance, if both EDR and HRV classify a 30-second epoch as normal, SDB, or indeterminate, then combined EDR and HRV classification is the same as the individual EDR and HRV determinations. If EDR marks an epoch as normal or SDB and if HRV classifies the same epoch as indeterminate, then the combined method agrees with EDR classification. If EDR and HRV have controversial classifications, the final classification done by the combined method is indeterminate.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
Table 2 Performance for quantitative assessment of SDB Method
Classified (%)
Accuracy (%)
EDR alone HRV alone EDR and HRV combined
96 99 78
79 82 88
The evaluation data set includes a total of 1 538 222 epochs. Hypopnea epochs are included as apnea. The column “Classified” shows the percentage of the epochs classified as normal or SDB (and not indeterminate), and the column “Accuracy” shows the percentage match between the algorithms' output and manual scoring done by human experts.
receiver operating characteristic (ROC) curve for the learning set, the combination technique classified recordings with an ASI of larger than 1.5 as SDB and others as normal. As seen from Table 3, in the evaluation data set, the algorithm correctly classified 1355 of 1907 subjects with a total accuracy of 71%. In this test, the hypopnea epochs were included as they contributed to the AHI used for creating the criterion standard annotation for each patient. Discussion In classifying every epoch in the evaluation data set, the combined EDR and HRV techniques achieved accuracy of 88% for scoring 78% of the epochs as either normal or SDB. Testing the SDB status for each patient, this algorithm correctly classified 71% of patients with either moderate-tosevere SDB or mild-to-no SDB. In this article, we summarized our effort to provide a reliable screening technique for SDB based solely on analysis of a single-lead ECG recording. Our approach was based on the expectation that SDB affects respiratory pattern and that respiration pattern affects both heart rate and ECG morphology. Few techniques to detect sleep apnea using HRV have been previously proposed by other researchers. 17-19 Respiration effects on ECG morphology have been studied by several groups to derive the respiratory signal by estimating the heart electrical axis using 2, 20 3, 21 and even 8 22 ECG leads. Our EDR technique requires only 1 channel of ECG. This single-lead EDR method works best if the lead axis is significantly different from the mean electrical axis to obtain a relatively large signal. Compared with measuring QRS area for estimating EDR as introduced by others, 20 our technique
has an advantage of being insensitive to ECG baseline wander. Therefore, removal of baseline wander, which requires additional processing and could be challenging at times, is unnecessary. For classifying each epoch, the EDR algorithm needs only 30 seconds of signal at 1 time, and the HRV-based algorithm requires 6 minutes of ECG to mark the 30second epoch in the middle of the 6-minute window. 14 Because a short 30-second window is more likely to be corrupted completely by noise or artifact than the long 6minute window, the EDR-based method is more prone to noise compared with the HRV-based method. Therefore, more epochs are classified as indeterminate by the EDRbased method as compared with the HRV-based method. It is important to know that SDB assessment can be difficult even for human experts. There is considerable variability in SDB scoring when 2 qualified technologist score the same record or 1 expert scores the same record in different times. The basis of our method is that the altered respiratory patterns in patients with SDB affect both heart rate and ECG morphology. If such alteration is not detectable or eliminated by other factors, such as medication or cardiac disease or autonomic dysfunction, the ECG-based method will not work. The SDB screening techniques introduced in this study rely on a method for ECG beat detection and classification from single-lead ECG recordings. The quality of ECG beat detector and classifier impacts the performance of ECGbased techniques. Thus, the limitations in the evaluation data set should be pointed out. One major limitation is the relatively low quality of ECG recordings. The 1-channel ECG was recorded as part of the PSG study, and the quality of the ECG recording was rather poor. The HRV method does not work on patients who either have irregular heart rhythm or are on rhythm- and rate-control drugs or
Receiver Operating Characteristic (ROC) 1
Sensitivity
704
0.60
ASI threshold=1.5
0.5
Table 3 Classification performance for SDB screening by EDR- and HRV-based methods when used alone or combined Method
Sensitivity Specificity Positive Negative Accuracy predictive predictive value value
EDR alone 45 HRV alone 74 EDR and HRV 60 combined
76 62 82
64 64 75
60 72 69
61 68 71
The evaluation data set included 989 normal and 918 apneic patients marked by using AHI threshold of 15 using the criterion standard SDB annotations from full PSG. The combined method has significantly higher specificity, positive predictive value, and the total accuracy. All the numbers are reported in percentage (%).
0
0
0.18
0.5 1-Specificity
1
Fig. 4. Receiver operating characteristic curve for the performance of combined EDR and HRV algorithm on the evaluation data set. It is noted that the ASI threshold of 1.5 corresponds to the sensitivity of 60% and specificity of 82% as reported in Table 3. For hospital use, one can choose a different threshold to achieve higher sensitivity and lower specificity. For home screening, it may be desirable to change the threshold to make the algorithm less sensitive but more specific.
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
devices. We did not exclude subjects with irregular heart rate, other arrhythmias, or any cardiac disease. The HRV method depends on the sampling frequency, and it is better to have sampling rate of 500 SPS or higher. In our evaluation data set, ECGs have low sampling rates of 100 or 200 SPS. The EDR method is sensitive to amplitude resolution. The evaluation data set has poor amplitude resolution. Furthermore, an accurate estimated EDR signal can only be obtained if heart rate is high enough to prevent sub-Nyquist frequency undersampling. Therefore, extreme bradycardia can result in frequency aliasing; hence, unreliable EDR waveform. Similar undersampling can occur if, due to arrhythmia or other cardiac diseases, the patient has many abnormal QRS complexes for which an EDR value will not be calculated. It is also known that the ECG axis, hence, EDR morphology, changes as a function of body position and posture. This change, however, if limited, may not significantly affect the epoch classifications because the 30-second epochs are being classified independently. We did not exclude these potentially challenging ECG recordings from our database. A low sampling rate may produce a jitter in the estimated R-wave fiducial point, which alters the spectrum considerably. The optimal range is 250 to 500 SPS or perhaps even higher, 23 which is above the 100 SPS sampling rate of most studies assessed here. The same jitter in the R-wave fiducial point potentially affects the EDR values measured as beat swings that may also be affected by poor amplitude resolution of ECG recordings. More study is needed to investigate the potential effects of ECG recordings quality on our SDB monitoring algorithms. The current performance results may improve on higher quality ECG data, in particular, with higher sampling rate and amplitude resolution. However, our techniques performed reasonably well even under the current challenging conditions. In the clinical setting, the outcome of the algorithm can be adjusted by using different ASI limits. It may be better to select 2 separate ASI limits for normal and SDB and mark the group falling in between as indeterminate or borderline. This may miss some patients with milder SDB but will usually identify those with moderate to severe disordered breathing. When the algorithm is used for SDB screening, it is possible to change its sensitivity and specificity by adjusting the ASI threshold. Fig. 4 shows the ROC curve for the combined EDR and HRV on the evaluation data set. It is noted that the ASI threshold of 1.5 corresponds to the sensitivity of 60% and specificity of 82% as reported in Table 3. For hospital SDB screening where higher sensitivity is desired, a different threshold can be used to achieve a higher sensitivity. For home SDB screening where a higher specificity is desired, an adjusted threshold can certainly lead to higher specificity with reduced sensitivity. Another way to improve the detection accuracy could be to combine the ECG signal with other easy-to-obtain signals such as pulse oximetry and/or limited respiratory measurements. The multiparameter approach for automatic screening of SDB will be the subject of ongoing research.
705
Conclusion Our 1-channel ECG-only SDB screening algorithm provides reasonably accurate SDB detection and quantification. Because ECG is readily available, noninvasive, and low cost, this approach will make it possible for SDB screening in a large population in both hospital and home care environments. This would be quite convenient for cardiac patients who have ECG monitoring in the hospital or at home. Thus, SDB screening can be done without additional cost. References 1. Cartwright R. Obstructive sleep apnea: a sleep disorder with major effects on health. Dis Mon 2001;47:109. 2. Epstein LJ, Kristo D, Strollo Jr PJ, et al. Clinical guideline for the evaluation, management and long-term care of obstructive sleep apnea in adults. J Clin Sleep Med 2009;5:263. 3. Somers VK, White DP, Amin R, et al. AHA/ACCF Scientific Statement. Sleep apnea and cardiovascular disease. Circulation 2008;118:1080. 4. Kapur V, Strohl KP, Redline S, Iber C, O'Connor G, Nieto J. Underdiagnosis of sleep apnea syndrome in U.S. communities. Sleep Breath 2002;6:49. 5. Young T, Evans L, Finn L, Palta M. Estimation of the clinically diagnosed proportion of sleep apnea syndrome in middle-aged men and women. Sleep 1997;20:705. 6. Iber C, Ancoli-Israel S, Chesson A, Quan SF. The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specifications. 1st ed. Westchester, IL: American Academy of Sleep Medicine; 2007. 7. Caples SM, Somers VK, Rosen CL, et al. The scoring of cardiac events during sleep. J Clin Sleep Med 2007;3:147. 8. Miletin MS, Hanly PJ. Measurement properties of the Epworth sleepiness scale. Sleep Med 2003;4:195. 9. Netzer NC, Stoohs RA, Netzer CM, Clark K, Strohl KP. Using the Berlin Questionnaire to identify patients at risk for the sleep apnea syndrome. Ann Intern Med 1999;131:485. 10. Abraham WT, Trupp RJ, Phillips B, et al. Validation and clinical utility of a simple in-home testing tool for sleep-disordered breathing and arrhythmias in heart failure: results of the Sleep Events, Arrhythmias, and Respiratory Analysis in Congestive Heart Failure (SEARCH) study. Conges. Heart Fail 2006;12:241. 11. Roche F, Gaspoz JM, Court-Fortune I, et al. Screening of obstructive sleep apnea syndrome by heart rate variability analysis. Circulation 1999;100:1411. 12. Goldberger AL, Amaral LA, Glass L, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 2000;101:E215. 13. Penzel T, McNames J, Murray A, de Chazal P, Moody G, Raymond B. Systematic comparison of different algorithms for apnoea detection based on electrocardiogram recordings. Med Biol Eng Comput 2002;40:402. 14. Babaeizadeh S, White DP, Pittman SD, Zhou SH. Automatic detection and quantification of sleep apnea using heart rate variability. J Electrocardiol 2010;43:535. 15. http://www.physionet.org/physiobank/database/ucddb/. 16. http://www.healthcare.philips.com/main/products/patient_monitoring/ products/st_ar/. 17. Guilleminault C, Connolly S, Winkle R, et al. Cyclical variation of the heart rate in sleep apnoea syndrome. Mechanisms, and usefulness of 24 h electrocardiography as a screening technique. Lancet 1984;1:126. 18. Penzel T, Amend G, Meinzer K, Peter JH, Von Wichert P. MESAM: a heart rate and snoring recorder for detection of obstructive sleep apnea. Sleep 1990;13:175. 19. Hilton MF, Bates RA, Godfrey KR, Chappell MJ, Cayton RM. Evaluation of frequency and time-frequency spectral analysis of heart rate variability as a diagnostic marker of the sleep apnoea syndrome. Med Biol Eng Comput 1999;37:760. 20. Moody GB, Mark RG, Zoccola A, Mantero S. Derivation of respiratory signals from multi-lead ECGs. Proceedings of Computers in Cardiology 1985. IEEE Press 1985;12:113.
706
S. Babaeizadeh et al. / Journal of Electrocardiology 44 (2011) 700–706
21. Pinciroli F, Pozzi G, Rossi R. Processing electrocardiograms towards respiratory signals. Proc. of the 10th IEEE Engineering in Medicine & Biology Society Int. Conf. New Orleans, USA; 1988. p. 139. 22. Travaglini A, Lamberti C, De Bie J. Respiratory of signal derived from eight-lead ECG. Proceedings of Computers in Cardiology 1998. IEEE Press 1998;2:65.
23. Task Force of the European Society of Cardiology the North American Society of Pacing Electrophysiology. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Circulation 1996; 93:1043.