Automatic Detection of Respiration Rate From ... - Semantic Scholar

890

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 13, NO. 6, NOVEMBER 2009

Automatic Detection of Respiration Rate From Ambulatory Single-Lead ECG Justin Boyle, Niranjan Bidargaddi, Member, IEEE, Antti Sarela, Member, IEEE, and Mohan Karunanithi, Senior Member, IEEE

Abstract—Ambulatory electrocardiography is increasingly being used in clinical practice to detect abnormal electrical behavior of the heart during ordinary daily activities. The utility of this monitoring can be improved by deriving respiration, which previously has been based on overnight apnea studies where patients are stationary, or the use of multilead ECG systems for stress testing. We compared six respiratory measures derived from a single-lead portable ECG monitor with simultaneously measured respiration air flow obtained from an ambulatory nasal cannula respiratory monitor. Ten controlled 1-h recordings were performed covering activities of daily living (lying, sitting, standing, walking, jogging, running, and stair climbing) and six overnight studies. The best method was an average of a 0.2–0.8 Hz bandpass filter and RR technique based on lengthening and shortening of the RR interval. Mean error rates with the reference gold standard were ±4 breaths per minute (bpm) (all activities), ±2 bpm (lying and sitting), and ±1 breath per minute (overnight studies). Statistically similar results were obtained using heart rate information alone (RR technique) compared to the best technique derived from the full ECG waveform that simplifies data collection procedures. The study shows that respiration can be derived under dynamic activities from a single-lead ECG without significant differences from traditional methods. Index Terms—Cardiovascular system, electrocardiography, exercise, respiratory system.

I. INTRODUCTION ONITORING of respiration rate in the home and the influence of activity on respiration is important for sufferers of a variety of chronic diseases, such as heart disease, stroke, and chronic obstructive pulmonary disease. The prevalence of these disorders coupled with ageing populations and diminished workforces place pressures on health care models. Technology solutions are needed that can provide monitoring and assessment information in home or community care environments, i.e., away from the hospital. Body sensor networks comprising wearable sensors and data storage/processing units are becoming ubiquitous in such ambulatory health-monitoring

M

Manuscript received February 11, 2009; revised July 7, 2009. First published September 22, 2009; current version published November 4, 2009. J. Boyle, A. Sarela, and M. Karunanithi are with the Australian E-Health Research Centre, Commonwealth Scientific and Industrial Research Organisation, Information and Communication Technologies (CSIRO ICT) Centre, Royal Brisbane and Women’s Hospital, Herston, Qld. 4029, Australia (e-mail: [email protected]; [email protected]; [email protected]). N. Bidargaddi was with the Australian E-Health Research Centre, CSIRO ICT Centre, Royal Brisbane and Women’s Hospital, Herston, Qld. 4029, Australia. He is now with Country Health South Australia, 100 Waymouth Street, Adelaide, S.A. 5000, Australia (e-mail: niranjan.bidargaddi@health. sa.gov.au). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TITB.2009.2031239

applications. We are currently developing such a home-based framework that would benefit from ambulatory respiration measurement [1]. Respiration rate from ambulatory patients is traditionally obtained using strain gauges or piezoelectric transducer devices strapped to the chest or abdomen, or pressure transducers to measure nasal/mouth air pressure, each with associated burdens of wear. An alternative is to derive respiration from ambulatory electrocardiography, which is increasingly being used in clinical practice to detect and characterize occurrences of abnormal cardiac electrical behavior of the heart during ordinary daily activities [2]. The appearance of the respiratory cycle in the heart rate signal (respiratory sinus arrhythmia) is not new [3] and many techniques have been developed to derive the respiration rate from the ECG. However, the majority of related work is from overnight apnea studies where patients are stationary. A significant challenge is to accurately monitor respiration during exercise, capturing a range of respiratory effort. The development of portable devices to record single-lead ECGs has led to the need to develop ECG-derived respiration (EDR) techniques that are robust to noise from activities of daily living. The study presented in this paper expands on related work that focuses on either sleep (stationary body movements) or multilead ECG systems for stress testing, and the next section briefly describes some of these techniques used in previous studies. II. RELATED WORK A useful review of ECG-derived respiration techniques is provided by Bailon et al. [4], which groups algorithms into categories based on beat morphology, heart rate, or a combination of both. A section is included on algorithm evaluation, which stresses the importance of comparing the derived respiratory information with a simultaneous recording of the respiration signal. The authors cite their algorithm that uses the standard 12-lead ECG to estimate respiration during stress testing [5]. The stress testing consisted of 14 volunteers and 15 patients using a bicycle ergometer for approximately 10 min, and excluded a recovery phase of lying on a bed after the ergometer due to reliability difficulties with the reference signal. Algorithm error performance was 5.9% (mean) ± 4% [standard deviation (SD)]. Grossman et al. [6] compared respiration and cardiac parameters measured by the Vivometrics Lifeshirt ambulatory monitor in 40 subjects during normal daily activities for when heart rate was less than 110 beats per minute. This heart rate range was chosen as it was claimed to be the primary range of cardiac parasympathetic control. They found that ambulatory

1089-7771/$26.00 © 2009 IEEE Authorized licensed use limited to: CSIRO LIBRARY SERVICES. Downloaded on March 15,2010 at 19:49:54 EDT from IEEE Xplore. Restrictions apply.

BOYLE et al.: AUTOMATIC DETECTION OF RESPIRATION RATE FROM AMBULATORY SINGLE-LEAD ECG

respiratory sinus arrhythmia (high-frequency heart rate variability calculated using a peak–valley algorithm) was significantly associated with variations in respiratory rate and tidal volume (average R = 0.8). Some approaches use the amplitude of the R-wave of the ECG signal to calculate EDR. Khaled and Farges [7] first filtered the ECG with an eighth-order 2.5–25Hz bandpass filter before detecting R peaks. The same process has also been proposed in magnetic resonance applications, involving R peak detection, sample and hold where the output is a step function representing the respiratory signal [8]. Other approaches use both R- and S-wave amplitudes. Dobrev and Daskalov [9] applied a high-pass filter with 5 Hz cutoff, then a 40-Hz low-pass filter, followed by QRS detection. The derived respiratory signal was the sum of the absolute Rand S-wave amplitudes (smoothed with a second order Butterworth 2-Hz cutoff low-pass filter). Mason and Tarassenko [10] applied high-pass filtering to remove baseline wander, and then determined the R-wave amplitude with respect to the baseline (R-EDR) and also with respect to the S-wave amplitude (RS-EDR). The S-wave was defined as the minimum value in a 0.1-s window after the R-wave peak. Amplitudes were plotted at each R-wave and successive points were linked. Not all peaks joined in this way were breaths, but only those where a peak–trough amplitude was above a threshold. Evaluation on a polysomnography (sleep) database found higher sensitivity with the RS-EDR method (77%) compared to the R-EDR method (68%). Similar techniques involve measuring the area of each QRS complex (proportional to amplitude) in a fixed window. Source code for the projection of amplitude variation [11], including a precompiled executable program is freely available on the National Institute of Health sponsored online resource www.PhysioNet.org. Furman et al. [12] found QRS area, R-wave amplitude, and duration to have correlation above 0.85 with simultaneously recorded respiration for nighttime sleep recordings of 24 subjects. Although performance figures are not stated, Travaglini et al. [13] report strong correlation with respiration derived from eight ECG leads with convention measurements of respiration when tested on ten volunteers. Mazzanti et al. [14] describe their validation of the same technique with 11 patients using simultaneous eight-lead ECG and oral–nasal airflow, thorax and abdominal movements. Respiration-cycle detection sensitivity/specificity was 98%/90% and apnea detection 87%/85%. A similar approach by Zhao et al. [15] calculated EDR from the angle of two orthogonal amplitudes (using two ECG channels). Nine subjects undertook a controlled protocol consisting of resting, paced breathing, exercise, and recovery. Correlation between observed and derived respiration measured on the frequency spectrum ranged from 0.82 to 0.99. The use of the wavelet transform for ECG-derived respiration is apparent in the literature. Yi and Park [16] applied the discrete wavelet transform to the ECG signal to obtain two subsignals at each level: a detailed signal representing the upper half of the frequency components, and an approximation signal representing the other half. The derived respiration was a reconstruction of the ninth decomposition detail signal, corre-

891

sponding to frequencies of 0.2–0.4 Hz, and respiration periods were found from this by detecting zero crossings in a falling direction. Correlation with nasal airflow was reported to be above 0.9 when monitoring ECG during sleep [16]. The wavelet transform has also been widely applied to single-lead ECG signals and RR interval time series to identify diagnostic markers for sleep apnea [17]–[20]. It is relevant to describe approaches used to detect apnea episodes from nighttime recordings of single-lead ECG recordings that were submitted as a part of an Apnea Challenge conducted by the IEEE-sponsored conference Computers in Cardiology (CINC) and the National Institute of Health’s online physiological library PhysioNet. This challenge has been reviewed previously [21], [22] and involved scoring submissions identifying the number of subjects out of 30 with and without apnea, and to identify each minute (percentage of the total recording) demonstrating apnea. The results indicate that at least most of the information necessary to recognize sleep apnea is contained in the ECG. Some methods derived a respiration signal prior to apnea classification [18], while other techniques looked for apnea indicators without requiring an estimation of respiration [23]. The algorithms that performed the best used frequency domain parameters of heart rate variability or derived respiration with R-wave morphology [18], [24]–[26]. The results of the top three algorithms were also combined using majority voting, yielding an accuracy of 93%. Sleep apnea detection using cardiac signals is reported in the literature beyond the CINC challenge using approaches such as cyclic variations in heart rate [27] and power spectral density of the interbeat interval increment of the very low frequencies [28]. The results of these studies indicate that most of the information necessary to detect respiration is contained in the ECG—a challenge remains to derive this information using only a singlelead ECG and be tolerant of movement artefacts. III. AIM The aim of our study was to derive a measure of respiration rate from a single-lead ECG recording that was robust against noise from activities of daily living. The desired performance was no significant difference with more cumbersome traditional methods for quantifying respiration rate. IV. METHODS A. ECG-Derived Respiration Techniques From a review of related work, it is apparent that the mechanical action of respiration causes detectable frequency content changes in the ECG spectrum. We chose the following methods based on indications in the literature of their potential merits and preliminary review of their implemented results. 1) {EDR}: ECG wavelet decomposition—detailed signal, scale 9. 2) {EDR2}: Sum of (approximation signal, scale 8) and (detailed signal, scale 9). 3) {EDR3}: ECG bandpass—0.2-0.8 Hz. 4) {EDR4}: ECG bandpass—0.2–0.4 Hz.

Authorized licensed use limited to: CSIRO LIBRARY SERVICES. Downloaded on March 15,2010 at 19:49:54 EDT from IEEE Xplore. Restrictions apply.

892


Fig. 1. (Left) Wavelet decomposition refers to representing a signal as different frequency bands. (Right) Wavelet decomposition tree for example ECG signal showing detailed signal, scale 9.

5) {RR}: RR interval: shortening RR interval with inspiration and lengthening RR intervals with expiration. 6) {AvgeEDR3 + RR}: Average of EDR3 and RR methods. While the RR interval changes and bandpass methods can be easily understood, it may assist to provide some description on the wavelet decomposition methods. Most raw signals are represented in the time domain, meaning that they would be represented as time–amplitude when plotted. This representation is not always best, as often the most distinguished information is hidden in the frequency content of the signal. Popular transforms to the frequency domain include the Fourier, Hilbert, Wigner distributions, Radon, and wavelet transform. The wavelet transform is a time–frequency representation of the signal, and wavelet decomposition refers to breaking the signal into a low-frequency (approximation) and high-frequency (detailed) part at several scales, as illustrated in Fig. 1. After splitting we obtain a vector of approximation coefficients and a vector of detail coefficients, both at a coarser scale. The information lost between two successive approximations is captured in the detail coefficients. We used biorthogonal spline wavelets (“bior5.5”) for which symmetry and exact reconstruction are possible with finite-impulse response (FIR) filters. B. Validation Ambulatory ECG signals were obtained at 300 Hz using a single-lead portable ECG monitor (Alive Heart Monitor; Alive Technologies; Gold Coast, Australia). Electrodes were placed on the right shoulder just below the clavicle (RA position) and the fifth intercostal space at the left midaxillary line (V6 position). The ECG-derived respiration signals were compared with simultaneously measured respiration air flow at 100 Hz obtained from an ambulatory nasal cannula respiratory monitor (ApneaLink; ResMed; GmbH). This device records patient respiratory nasal pressure and is intended for use as a screening device to determine the need for clinical evaluation by polysomnography based on output from accompanying software. Nasal pressure was measured directly at the nostrils and is linearized to air flow by the analysis software [29]. Both devices were powered by rechargeable batteries and fitted with internal memory cards from which data were retrospectively downloaded. The respiration signal was resampled

Fig. 2.

Protocol for 1-h controlled respiration studies.

to match ECG at 300 Hz and both signals were synchronized by aligning the end of the recordings when both devices were switched off simultaneously. Measurements were a convenience sample of ten controlled 1-h recordings in an office environment and six overnight recordings, collected from healthy young-tomiddle aged adults with no known cardiac or respiratory diseases. Participation was on a consented voluntary basis and the study was subjected to Internal Review. The controlled 1-h recordings followed a prescribed protocol: 1) 15-min normal office activity (mostly sitting). 2) 15-min lying. 3) Treadmill: a) 2 min slow walk (2.5 km/h); b) 2 min medium walk (4.5 km/h); c) 2 min fast walk/jog (6.9 km/h); d) recovery period (8 min approximately). 4) Stair climbing: a) 6 min stair climbing (down 40 steps, up 40, down, etc.); b) recovery period (8 min approximately). Fig. 2 illustrates data from one subject collected in the controlled study, and also shows triaxial accelerometer data simultaneously collected at 10 Hz. C. Statistical Analysis The accuracy of each method was expressed by determining the average percentage error of the measurements when compared with the true value, calculated by subtracting the test value from the reference value and dividing the result by the reference value of each measurement, and then taking the average of this [30]. Absolute values were used in our study to avoid errors cancelling each other out Error =

|(Derived − Observed)| . Observed


(1)


Fig. 3.

893

Error for all activities shown in (left) percentage and (right) bpm showing the best derivation performance across all activities was around ±4 bpm.

To determine if there were significant differences between the derived and observed number of breaths, we used a two-tailed Dunnett’s test for comparing a controlled mean to each group mean. To determine if there were significant differences between the derived methods, we used Tukey multiple comparison tests to perform pairwise comparisons of the error data. Where two datasets were compared (for example, with and without ECG baseline correction), we performed two sample t-tests for each individual method. The significance level chosen was α = 0.05. V. RESULTS The results were designed to answer a number of questions, which are as follows. A. What Was the Overall Performance Across All Activities? Fig. 3 depicts the error between the derived and observed respiration signals for all activities tested during the 1-h controlled test, and includes 95% confidence intervals around the mean error for each of the methods across the ten subjects. Methods {EDR3} and {AvgeEDR3 + RR} had the least mean error with the observed signal at around ± 4 breaths per minute (bpm) (∼17%) difference with the observed signal across the range of activities tested. Most variation across the sample was seen with the {RR} method. Multiple comparison testing indicated that there were no significant differences with the observed number of breaths for the {EDR3}, {RR}, and {AvgeEDR3 + RR} methods. Thus, considering overall performance across all activities, these three methods gave equivalent performance in deriving respiration. B. What is the Value of Collecting the Full ECG Signal Compared to Just Heart Rate With Respect to Respiratory Monitoring? A comparison was made between the best ECG-derived method {AvgeEDR3 + RR} with the {RR} method. Collection of the full ECG waveform usually requires adhesive dots that are stuck to a patient’s chest, whereas heart rate alone can be obtained from a simpler chest strap. Many of the patients involved in other studies undertaken by the authors are elderly with thin fragile skin, and have cited skin irritation problems associated

with the use of stick-on ECG dots. This is a particularly important problem for studies that extend for long periods of time, where the application of the dots has to be repeated after daily showering, for example. It was desired to test the stated assumption by others that respiration derived from the RR interval time series is less reliable than morphology-based ECG-derived respiration [31]. A pairwise comparison of the error data indicated no significant differences between the {AvgeEDR3 + RR} and {RR} methods. Thus, it is possible to obtain statistically similar results using heart rate information alone compared to the full ECG signal. C. Were There Any Differences Between the Activities? The dataset was partitioned into the following activity categories: office, lying, slow treadmill, medium treadmill, fast treadmill, recovery 1, stairs, recovery 2. Although the protocol was timed, exact start and finish times were assigned by inspection of the simultaneously recorded accelerometer data, which at 10 Hz, allowed precise annotations to be made. Fig. 4 depicts the 1-h test activities along the horizontal axis in increasing order of mean observed respiration rate. Linear trends have been fitted to the mean derived respiration rates and a hypothetical ideal method is shown by a dashed line (y = x). The figure indicates the bias and magnitude of the errors. It can be seen that the bandpass methods {EDR3} and {EDR4} do not follow changes in the observed respiration, with {EDR3} mostly constant around 15 bpm (0.25 Hz) and {EDR4} around 9 bpm (0.15 Hz). While these bandpass methods have low error at low observed respiration rates, they do not perform well at higher rates experienced during strenuous exercise. As respiration rate increases, the wavelet decomposition methods {EDR} and {EDR2} perform better. The error associated with the other methods generally increase with respiration rate. It should be noted that there are several variables relating to this performance in addition to increasing respiration rate. For example, fast treadmill and stair climbing activities exhibit much movement of the ECG leads that creates a lot of noise. Statistical testing shows no significant difference between the {AvgeEDR3 + RR} method and the observed rate for all activities (see Table I) and for this reason it is the recommended method. The wavelet decomposition method {EDR} had the best


894


Fig. 4.

Trends in derived respiration as a function of respiration rate. TABLE I DIFFERENCES IN DERIVED RESPIRATION FOR CONTROLLED 1-h TEST

accuracy and was statistically similar to the observed rate only at high respiration rates. {AvgeEDR3 + RR} is the preferred method in free-living environments that include a range of respiration rates. Table I also indicates mean absolute differences between each method and the observed rate. For stationary activities (e.g., lying and sitting) derived respiration rate was around ±2 bpm (∼17%) of the reference gold standard. D. What are the Findings From the Overnight Study? Simultaneous respiratory and ECG monitoring was also performed for six subjects for overnight sleep monitoring. The data was meticulously assessed, as the respiratory monitor software that was used as the gold standard reference signal excluded periods of data where the observed signal was deemed too small for analysis, and hence, no reference signal was available although ECG was still recorded. It is unclear what circumstances trigger these occurrences and whether it is related to nasal passage shape, slippage of cannula prongs, etc. It does highlight a disadvantage of the commercial respiratory monitor in addition to the obvious burden of wear. Error rates for overnight analysis are presented in Fig. 5, which includes 95% confidence intervals around the mean error for each of the methods across the six

subjects. Method {AvgeEDR3 + RR} had the least mean error with the observed signal at 9%, which equates to an average of around 1 bpm difference with the observed signal across the night. Fig. 6 depicts algorithm performance with respect to actual observed respiration rate. For a more detailed analysis, it was desired to pool all overnight data, partitioning overnight measurements into blocks of maximum 1 h length and discarding data below 2 min duration, which resulted in 57 observations. Statistical testing showed that there were no significant differences with the observed number of breaths for the {EDR3}, {RR}, and {AvgeEDR3 + RR} methods. VI. DISCUSSION It is often beneficial to assess respiration during task-based activities, where respiratory effort can be compared to perceived difficulty. Thus, a technique that offers high accuracy over a dynamic range of respiration rate is desirable. This paper has described the validation of a method derived from the ECG signal that is robust against movement noise and statistically similar to traditional techniques. The high respiration rates in this study were observed during activities exhibiting large



Fig. 5.

895

Error rates for overnight study in %error and bpm.

Fig. 6. Average observed and derived respiration rates for overnight study. (Left) Error bars showing variance across sample. (Right) Error bars showing pooled variance across comparison test; shaded = no significant difference with observed signal.

amounts of movement with corresponding high artefact noise. High performance was maintained in the presence of this noise, which is inherent in lead-based cardiac monitoring systems. There were several limitations to this study that present opportunities for further work. The study used healthy participants with no known cardiac or respiratory diseases. It would be interesting to include data from these groups in the analysis. Also the activities of daily life were represented by lying, sitting, standing, walking, jogging, running, and stair climbing. There are difficulties in arriving at a standard range of test activities that are representative of activities at a population level. Another point is that respiration rates were averaged across the sample; improved performance may be reported if algorithms were tuned to an individual’s fitness levels. Large variance in respiration were observed for example in the observed stair-climbing respiration (mean = 30 bpm, SD = 7 bpm) for which respiratory fitness would have an impact. Similarly, it would be useful to tune and correlate derived respiration with reference to metabolic expenditure, obtainable directly from simultaneous VO2 testing or derived from accelerometer output. The wavelet decomposition method {EDR} had the highest accuracy at high respiration rates, and could be the preferred option during high-performance activities, such as exercise, sports, etc., where a better resolution is required. Finally, the methods evaluated in the study were processed on a regular PC running MATLAB (v7.2.0.232 R2006a,

The MathWorks, Inc.) and mobile, real-time implementations would have different computational limitations. However, we believe the results reported in the paper will remain valid in such real-time or mobile platforms based on implementation of similar algorithms in real-time applications [32]. We are currently developing a mobile-phone-based web portal to capture such ambulatory data and incorporation of the algorithms is a consideration of future work by our research group. VII. CONCLUSION The following conclusions are drawn from this study. 1) For a range of common activities of daily life tested in the study (lying, sitting, walking, jogging, stair climbing, and transitions), it was possible to derive respiration rate from a single-lead ECG to within ±4 bpm (mean absolute percentage error ∼17%) of a reference gold standard. 2) For stationary activities (e.g., lying and sitting) derived respiration rate was within ±2 bpm (mean absolute percentage error ∼17%) of a reference gold standard. 3) For overnight studies (sleep) derived respiration was within ±1 bpm (mean absolute percentage error ∼9%) of a reference gold standard. 4) The best method was an average of EDR3 and RR rate, which was statistically similar to the observed rate across all test activities and when analyzed across each individual


896

5) 6) 7)

8) 9)


activity. The EDR3 technique was a 0.2–0.8 Hz bandpass filter applied to the ECG signal, and the RR technique was based on lengthening and shortening of the RR interval. The accuracy of wavelet decomposition methods increased as respiration rate increased. Bandpass techniques do not reflect changes in observed respiration rate. There were no significant differences in mean percentage error in the RR method to methods using the full ECG waveform. Thus, it is possible to obtain statistically similar results using heart rate information alone compared to the full ECG signal. The RR method was statistically similar to the observed rate across all test activities except fast treadmill (6.9 km/h) and stair climbing. The techniques are robust to noise introduced from body position changes (e.g., transitions). For the activities tested in the study, there were no significant differences obtained in preprocessing the ECG signal for baseline correction, which reduces computational overhead. ACKNOWLEDGMENT

The authors wish to thank the test subjects who participated in the validation studies. REFERENCES [1] A. Sarela, J. Salminen, E. Koskinen, O. Kirkeby, I. Korhonen, and D. Walters, “A home-based care model for outpatient cardiac rehabilitation based on mobile technologies,” presented at the Fourth Int. Conf. Pervasive Comput. Technol. Healthcare, London, U.K., 2009. [2] A. Kadish, A. Buxton, H. Kennedy, B. Knight, J. Mason, C. Schuger, C. Tracy, W. Winters, A. Boone, M. Elnicki, J. Hirshfeld, B. Lorell, G. Rodgers, C. Tracy, and H. Weitz, “ACC/AHA clinical competence statement on electrocardiography and ambulatory electrocardiography: A report of the American College of Cardiology/American Heart Association/American College of Physicians-American society of internal medicine task force on clinical competence (ACC/AHA committee to develop a clinical competence statement on electrocardiography and ambulatory electrocardiography),” Circulation, vol. 104, pp. 3169–3178, 2001. [3] C. Ludwig, “Beitrage zur kenntnis des einflusses der respiratons bewegungen auf den blutumlauf im aortensystem,” Arch. Anat. Physiol., vol. 13, pp. 242–257, 1847. [4] R. Bailon, L. Sornmo, and P. Laguna, “ECG-derived respiratory frequency estimation,” in Advanced Methods and Tools for ECG Analysis, G. Clifford, F. Azuaje, and P. McSharry, Eds. Norwood, MA: Artech House, 2006, pp. 215–244. [5] R. Bailon, L. Sornmo, and P. Laguna, “A robust method for ECG-based estimation of the respiratory frequency during stress testing,” IEEE Trans. Biomed. Eng., vol. 53, no. 7, pp. 1273–1285, Jul. 2006. [6] P. Grossman, F. H. Wilhelm, and M. Spoerle, “Respiratory sinus arrhythmia, cardiac vagal control, and daily activity,” Amer. J. Physiol. Heart Circ. Physiol., vol. 287, pp. H728–H734, 2004. [7] Z. Khaled and G. Farges, “First approach for respiratory monitoring by amplitude demodulation of the electrocardiogram,” in Proc. 14th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 1992, vol. 6, pp. 2535–2536. [8] J. Felblinger and C. Boesch, “Amplitude demodulation of the electrocardiogram signal (ECG) for respiration monitoring and compensation during MR examinations,” Magn. Reson. Med., vol. 38, pp. 129–136, 1997. [9] D. Dobrev and I. Daskalov, “Two-electrode telemetric instrument for infant heart rate and apnea monitoring,” Med. Eng. Phys., vol. 20, pp. 729– 734, 1998. [10] C. Mason and L. Tarassenko, “Quantitative assessment of respiratory derivation algorithms,” in Proc. 23rd Annu. Int Conf. IEEE Eng. Med. Biol. Soc., 2001, vol. 2, pp. 1998–2001.

[11] G. Moody, R. Mark, A. Zoccola, and S. Mantero, “Derivation of respiratory signals from multi-lead ECGs,” Comput. Cardiol., vol. 12, pp. 113– 116, 1985. [12] F. D. Furman, Z. Shinar, A. Baharav, and S. Akselrod, “Electrocardiogram derived respiration during sleep,” in Proc. Comput. Cardiol., 2005, pp. 351–354. [13] A. Travaglini, C. Lamberti, J. DeBie, and M. Ferri, “Respiratory signal derived from eight-lead ECG,” in Proc. Comput. Cardiol., 1998, pp. 65– 68. [14] B. Mazzanti, C. Lamberti, and J. DeBie, “Validation of an ECG-derived respiration monitoring method,” in Proc. Comput. Cardiol., 2003, pp. 613– 616. [15] L. Zhao, S. Reisman, and T. Findley, “Respiration derived from the electrocardiogram during heart rate variability studies,” in Proc. 16th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 1994, vol. 1, pp. 123–124. [16] W. Yi and K. Park, “Derivation of respiration from ECG measured without subject’s awareness using wavelet transform,” in Proc. 24th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2002, vol. 1, pp. 130–131. [17] F. Roche, V. Pichot, E. Sforza, I. Court-Fortune, D. Duverney, F. Costes, M. Garet, and J.-C. Barthélémy, “Predicting sleep apnoea syndrome from heart period: A time–frequency wavelet analysis,” Eur. Respir. J., vol. 22, pp. 937–942, 2003. [18] B. Raymond, R. Cayton, R. Bates, and M. Chappell, “Screening for obstructive sleep apnoea based on the electrocardiogram—The computers in cardiology challenge,” in Proc. Comput. Cardiol., 2000, pp. 267–270. [19] F. Ng, I. Garcia, P. Gomis, A. La Cruz, G. Passariello, and F. Mora, “Bayesian hierarchical model with wavelet transform coefficients of the ECG in obstructive sleep apnea screening,” in Proc. Comput. Cardiol., 2000, pp. 275–278. [20] M. Schrader, C. Zywietz, V. von Einem, B. Widiger, and G. Joseph, “Detection of sleep apnea in single channel ECGs from the PhysioNet data base,” in Proc. Comput. Cardiol., 2000, pp. 263–266. [21] G. Moody, R. Mark, A. Goldberger, and T. Penzel, “Stimulating rapid research advances via focused competition: The Computers in Cardiology Challenge 2000,” in Proc. Comput. Cardiol., 2000, pp. 207–210. [22] T. Penzel, J. McNames, A. Murray, P. de Chazal, G. Moody, and B. Raymond, “Systematic comparison of different algorithms for apnoea detection based on electrocardiogram recordings,” Med. Biol. Eng. Comput., vol. 40, pp. 402–407, 2002. [23] M. Jarvis and P. Mitra, “Apnea patients characterized by 0.02 Hz peak in the multitaper spectrogram of electrocardiogram signals,” in Proc. Comput. Cardiol., 2000, pp. 769–772. [24] P. de Chazal, C. Heneghan, E. Sheridan, R. Reilly, P. Nolan, and M. O’Malley, “Automated processing of the single-lead electrocardiogram for the detection of obstructive sleep apnoea,” IEEE Trans. Biomed. Eng., vol. 50, no. 6, pp. 686–696, Jun. 2003. [25] J. McNames and A. Fraser, “Obstructive sleep apnea classification based on spectrogram patterns in the electrocardiogram,” in Proc. Comput. Cardiol., 2000, pp. 749–752. [26] Z. Shinar, A. Baharav, and S. Akselrod, “Obstructive sleep apnea detection based on electrocardiogram analysis,” in Proc. Comput. Cardiol., 2000, pp. 757–760. [27] P. Stein, S. Duntley, P. Domitrovich P, P. Nishith, and R. Carney, “A simple method to identify sleep apnea using Holter recordings,” J. Cardiovasc. Electrophysiol., vol. 14, pp. 467–473, 2003. [28] F. Roche, S. Celle, and V. Pichot, “Analysis of the interbeat interval increment to detect obstructive sleep apnoea/hypopnoea,” Eur. Respir. J., vol. 29, pp. 1206–1211, 2007. [29] R. Farré, J. Rigau, and J. Montserrat, “Relevance of linearizing nasal prongs for assessing hypopneas and flow limitation during sleep,” Amer. J. Respir. Crit. Care Med., vol. 163, pp. 494–497, 2001. [30] N. Latman and R. Lanier, “Expressions of accuracy in the evaluation of biomedical instrumentation,” Biomed. Instrum. Technol., vol. 32, pp. 282– 288, 1998. [31] G. Clifford and M. Oefinger, “ECG acquisition, storage, transmission, and representation,” in Advanced Methods and Tools for ECG Analysis, G. Clifford, F. Azuaje, and P. McSharry, Eds. Norwood, MA: Artech House, 2006, pp. 32–33. [32] L. Klingbeil, T. Wark, and N. Bidargaddi, “Efficient transfer of human motion data over a wireless delay tolerant network,” in Proc. 3rd Int. Conf. Intell. Sens., Sens. Netw. Inf., 2007, pp. 583–588.

Authors’ photographs and biographies not available at the time of publication.
