IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 2, FEBRUARY 2006
343
[10] D. Wu, M. Zhang, J. C. Liu, and W. Bauman, “On the Adaptive Detection of Blood Vessels in Retinal Images,” Dept. Comput. Sci., Texas A&M Univ., College Station, TX, Tech. Rep. 2005-3-3, 2005. [11] D. Wu, “Segmentation, registration, and selective watermarking of retinal images,” Ph.D. dissertation, Dept. Comput. Sci., Texas A&M Univ., College Station, TX, 2005. [12] F. Grubbs, “Procedures for detecting outlying observations in samples,” Technometrics, vol. 11, no. 1, pp. 1–21, Feb. 1969. [13] X. Jiang and D. Mojon, “Adaptive local thresholding by verification-based multithreshold probing with application to vessel detection in retinal images,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 1, pp. 131–137, Jan. 2003.
Comparison of Atrial Signal Extraction Algorithms in 12-Lead ECGs With Atrial Fibrillation Philip Langley*, José Joaquín Rieta, Martin Stridh, José Millet, Leif Sörnmo, and Alan Murray
Fig. 3. Detection results of one normal image (0077) and one abnormal image (0001).
little calibration. In contrast, the two approaches that are being compared need eight and ten sets of parameter values, respectively. REFERENCES [1] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,” IEEE Trans. Med. Imag., vol. 19, no. 3, pp. 203–210, Mar. 2000. [2] A. Can, H. Shen, J. N. Turner, H. L. Tanenbaum, and B. Roysam, “Rapid automated tracing and feature extraction from retinal fundus images using direct exploratory algorithms,” IEEE Trans. Inf. Technol. Biomed., vol. 3, no. 2, pp. 125–138, Jun. 1999. [3] L. Zhou, M. Rzeszotarski, L. Singerman, and J. Chokreff, “The detection and quantification of retinopathy using digital angiograms,” IEEE Trans. Med. Imag., vol. 13, no. 4, pp. 619–626, Dec. 1994. [4] O. Chutatape, L. Zhang, and S. M. Krishnan, “Retinal blood vessel detection and tracking by matched Gaussian filter and Kalman filters,” in Proc. 20th Int, Conf. IEEE Engineering in Medicine and Biology Society, vol. 6, Hong Kong, China, 1998, pp. 3144–3149. [5] L. Gang, O. Chutatape, and S. Krishnan, “Detection and measurement of retinal vessels in fundus images using amplitude modified second-order Gaussian filter,” IEEE Trans. Biomed. Eng., vol. 49, no. 2, pp. 168–172, Feb. 2002. [6] A. Hoover and M. Goldbaum, “Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels,” IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 951–958, Aug. 2003. [7] J. A. Stark, “Adaptive image contrast enhancement using generalizations of histogram equalization,” IEEE Trans. Image Process., vol. 9, no. 5, pp. 889–896, May 2000. [8] J. G. Daugman, “Uncertainty relation for resolution in space, spatial frequency and orientation optimized by two-dimensional visual cortical filters,” J. Opt. Soc. Am., vol. 2, pp. 1160–1169, Jul. 1985. [9] Z.-Q. Liu, J. Cai, and R. Buse, Handwriting Recognition: Soft Computing and Probabilistic Approaches. Berlin, Germany: Springer-Verlag, 2003.
Abstract—Analysis of atrial rhythm is important in the treatment and management of patients with atrial fibrillation. Several algorithms exist for extracting the atrial signal from the electrocardiogram (ECG) in atrial fibrillation, but there are few reports on how well these techniques are able to recover the atrial signal. We assessed and compared three algorithms for extracting the atrial signal from the 12-lead ECG. The 12-lead ECGs of 30 patients in atrial fibrillation were analyzed. Atrial activity was extracted by three algorithms, Spatiotemporal QRST cancellation (STC), principal component analysis (PCA), and independent component analysis (ICA). The amplitude and frequency characteristics of the extracted atrial signals were compared between algorithms and against reference data. Mean (standard deviation) amplitude of QRST segments of V1 was 0.99 (0.54) mV, compared to 0.18 (0.11) mV (STC), 0.19 (0.13) mV (PCA), and 0.29 (0.22) mV (ICA). Hence, for all algorithms there were significant reductions in the amplitude of the ventricular activity compared with that in V1. Reference atrial signal amplitude in V1 was 0.18 (0.11) mV, compared to 0.17 (0.10) mV (STC), 0.12 (0.09) mV (PCA), and 0.18 (0.13) mV (ICA) in the extracted atrial signals. PCA tended to attenuate the atrial signal in these segments. There were no significant differences for any of the algorithms when comparing the amplitude of the reference atrial signal with that of the extracted atrial signals in segments in which ventricular activity had been removed. There were no significant differences between algorithms in the frequency characteristics of the extracted atrial signals. There were discrepancies in amplitude and frequency characteristics of the atrial signal in only a few cases resulting from notable residual ventricular activity for PCA and ICA algorithms.
Manuscript received October 7, 2004; revised May 8, 2005. This work was supported in part by the Spanish Council under Grant TIC2002-00957 and Grant IIARC0/2004/249 Generalitat Valenciana. The work of P. Langley was supported in part by the UK Engineering and Physical Sciences Research Council under Grant GR/R37890/01. Asterisk indicates corresponding author. *P. Langley is with the Cardiovascular Physics and Engineering Research Group, Medical Physics Department, Freeman Hospital, University of Newcastle upon Tyne, Newcastle upon Tyne NE7 7DN, U.K. (e-mail:
[email protected]). J. J. Rieta and J. Millet are with the Bioengineering Electronic and Telemedicine Research Group, Electronic Engineering Department, Polytechnic University of Valencia. EPSG, Carretera Nazaret Olivia s/n, 46730, Gandía, Valencia, Spain. M. Stridh and L. Sörnmo are with the Signal Processing Group, Department of Electroscience, Lund University, SE 221 00 Lund, Sweden. A. Murray is with the Cardiovascular Physics and Engineering Research Group, Medical Physics Department, Freeman Hospital, University of Newcastle upon Tyne, Newcastle upon Tyne NE7 7DN, U.K. Digital Object Identifier 10.1109/TBME.2005.862567
0018-9294/$20.00 © 2006 IEEE
344
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 2, FEBRUARY 2006
In conclusion, the extracted atrial signals from these algorithms exhibit very similar amplitude and frequency characteristics. Users of these algorithms should be observant of residual ventricular activities which can affect the analysis of the fibrillatory waveform in clinical practice. Index Terms—Atrial fibrillation, atrial signal, comparative study, independent component analysis, principal component analysis, spatiotemporal QRST cancellation.
I. INTRODUCTION A. Atrial Fibrillation Atrial fibrillation is a major cause of morbidity and mortality in the elderly population. Risk of stroke is increased fivefold and atrial fibrillation is a common arrhythmia, affecting between 2% and 10% of those over 50 years of age [1]. The mechanisms of atrial fibrillation are unclear but increased amounts of fibrous tissue, enlarged atria and rapid electrical firing in the pulmonary veins are contributory factors in some patients [2], [3]. Atrial fibrillation leads to remodeling of the electrophysiological properties of the atria and the arrhythmia is perpetuated by shortening of atrial refractoriness [4]. This leads to increased frequency of atrial activations due to shortening of the atrial cellular cycle length. Atrial activation frequency measured invasively at specific sites within the atria, or noninvasively on the body surface are highly correlated [5], [6]. It is accepted that noninvasive measurement of fibrillatory waves is useful for the treatment and management of patients [7]. The advantage that body surface measurements can be obtained repeatedly and over longer duration than invasive measurements has been exploited to reveal the time course of atrial fibrillation. Circadian variation of atrial fibrillation frequency from body surface recordings has been demonstrated suggesting influence of the autonomic nervous system on atrial activation [8], [9]. There are distinct differences in variability of the atrial fibrillation frequency in chronic and paroxysmal subclasses. The fibrillatory frequency is known to increase in relation to the length of time spent in fibrillation, and so reflects the process of atrial remodeling [10]. Fibrillatory frequency is highly predictive of spontaneous termination of paroxysmal atrial fibrillation [11]. Recently body surface atrial wave analysis has been used to assess the effect of drug treatment on atrial fibrillation [12], [13]. Fibrillatory frequency may be predictive of the potential success of electrical cardioversion [14], [15] and ablation therapy [16]. These studies demonstrate that the effects of treatment can be monitored on individual patients and raises the possibility of using characteristics of atrial waveform to guide therapeutic decisions. The clinical application of atrial fibrillation waveform analysis will be the subject of continued research. B. Atrial Signal Extraction One of the major problems of analysing the body surface atrial signal is that the atrial components of the ECG are small in comparison with the ventricular components and in much of the ECG the atrial signal is completely masked by the ventricular activities. Analysis in the frequency domain is confounded by a similar problem since there is overlap in the spectra of atrial and ventricular activities. Analysis of ECG segments free of QRS and T waves offers a simple solution [17], but not suitable when a continuous analysis is required or in patients with high ventricular rates, in which case the atrial signal can be completely obscured. Consequently algorithms to extract the atrial activities from the ECG have been developed. Specific examples are spatiotemporal QRST cancellation (STC), principal component analysis (PCA) and independent component analysis (ICA) [12], [18]–[21]. The aim of this international collaborative project was to assess and compare these atrial signal extraction algorithms. With increasing clinical application of atrial fibrillation waveform analysis it is essential
that the differences in these algorithms are assessed to allow meaningful cross study comparisons. II. METHODS A. Data, Preprocessing, and Algorithms The three research groups involved in the study, Lund, Newcastle and Valencia contributed a total of 30 12-lead ECGs of patients in atrial fibrillation to provide a common data set. ECGs were recorded to computer at a sampling rate of either 500 or 1000 Hz with amplitude resolution of less than 5 V. Those recorded at 500 Hz were re-sampled at a rate of 1000 Hz using low-pass interpolation to achieve uniformity in time resolution. Sixty-second–durations of ECG, free of ventricular ectopic activity were analyzed. Electrical supply noise was suppressed using a 50 Hz notch filter. Baseline variation was estimated in each lead using a third-order low-pass Butterworth forward/backward filter with cut-off frequency of 0.4 Hz. This was subtracted from the lead to provide ECGs free of baseline variation. Data sharing was facilitated by internet access to a server hosted by the Valencia group. Each group applied to the common data set their respective algorithm; STC by Lund, PCA by Newcastle and ICA by Valencia. B. Reference Atrial Signal The segments of V1 leads containing no ventricular activity, designated noQRST segments, were the reference data for this study since in these segments the body surface atrial signal were observed free of obscuring ventricular activity. V1 was chosen because of the dominance of the atrial signal in this lead in most patients [1]. Segments of the lead containing ventricular activity were designated QRST segments. C. Derivation of Atrial Signal in Lead V1 for PCA and ICA Inverse Transformation: Unlike STC, the atrial signals derived by PCA and ICA did not represent the atrial signal from a specific ECG lead but rather the global atrial signal derived from all leads. This prevented direct comparison of the PCA and ICA atrial signals with those derived by STC and the reference data. A further processing step was, therefore, necessary to facilitate these comparisons. The inverse transform of the PCA and ICA algorithms allows the extracted atrial activities to be projected back into a specific lead or leads. The inverse transform is described by
x^(j ) =
k
A01s(i) i;j
i=k
(k
2 1 : 12)
(1)
where x ^(j ) is the reconstructed j th ECG lead. The aim was to recreate only the atrial signal in Lead V1, omitting the ventricular activity. The transformed signals s(i) included in the inverse transform were selected so that the derived atrial signal was the one best representative of the atrial signal in V1. It should be noted that this processing step is only necessary to allow direct comparisons between the atrial signals for each of the algorithms and reference data, and would not otherwise be necessary for PCA and ICA algorithms. D. Evaluating the Algorithms 1) Amplitude: The evaluation was based primarily on the amplitude characteristics of the extracted atrial signals and address the following two main questions. a) How well do the algorithms suppress ventricular activity? This was quantified by comparing the amplitudes of V1 in QRST segments with those of corresponding segments in the extracted atrial signals. b) How well does the extracted atrial signal compare with the reference atrial signal? This was achieved by two comparisons. Firstly,
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 2, FEBRUARY 2006
345
TABLE I COMPARISON OF ECG QRST AND NOQRST SEGMENTS WITH CORRESPONDING SEGMENTS OF EXTRACTED ATRIAL SIGNALS. AMPLITUDE IS MEAN (STANDARD DEVIATION). P VALUES INDICATE SIGNIFICANCE OF DIFFERENCES BETWEEN ECG AND EACH ALGORITHM (MANN-WHITNEY TEST)
Fig. 1. Amplitudes of ECG lead V1 and extracted atrial signals in QRST (circle) and noQRST (cross) segments. The ECG noQRST segments (cross) are the reference data.
comparing the reference atrial signal with noQRST segments of the extracted atrial signal and secondly, comparing the reference atrial signal with QRST segments of the extracted signals. Amplitude of each segment was the peak-to-peak amplitude and the mean value was calculated across all segments for each ECG. 2) Frequency: Additionally, because of the importance of these algorithms in assessment of atrial fibrillation frequency, we compared the dominant atrial frequency of the extracted atrial signals. The dominant atrial frequency could not be extracted from the reference atrial signal because it was discontinuous, so intra-algorithm comparisons were made for the dominant frequency characteristics of the atrial signals. Atrial frequency was derived from the power spectrum calculated by periodogram using fast Fourier transform of the extracted atrial signals without windowing. E. Statistical Analysis
Fig. 2. QRST versus noQRST amplitude for each algorithm ( —STC, —PCA, 5—ICA).
B. Frequency
Parametric or nonparametric statistics were used according to the distribution of the variables. The statistical tests used to analyze the data are stated with the results. Statistical significance was assumed for p < 0:05. NS indicates not significant. III. RESULTS A. Amplitude Fig. 1 shows the amplitudes from QRST and noQRST segments of V1 and extracted atrial signals for all patients. Amplitude values are presented in Table I which also indicates the statistical differences for each algorithm with respect to ECG for both QRST and noQRST segments. Clearly all algorithms significantly reduce the ventricular activity as indicated by the significant differences with respect to V1 in QRST segments. Across the algorithms there were significant differences (p = 0:007, Kruskal-Wallis) in noQRST segments of the atrial signal compared to V1. Statistical analysis of the individual algorithms showed the amplitudes of the PCA atrial signals were significantly smaller than the reference signals. There were no significant differences for the other algorithms. There were no significant differences for all algorithms when comparing the reference atrial signal with QRST segments of the extracted atrial signals (p = NS, Kruskal-Wallis). The relationship between QRST and noQRST segments for each algorithm are illustrated in Fig. 2. Overall there was good agreement between the amplitudes of each segment, particularly for STC. Both PCA and ICA algorithms gave a few cases where the residual ventricular activity was relatively large compared to the corresponding atrial activity.
The patient group had a mean (standard deviation) dominant fibrillation frequency of 6.3 (1.0) Hz (STC), 6.1 (1.0) Hz (PCA), and 6.3 (1.0) Hz (ICA) (p = NS, one way analysis of variance), range 3.3 to 9.1 Hz. Fig. 3 shows the differences in dominant fibrillation frequency between pairs of algorithms. In 90% of cases the agreement was within 1 Hz. Two extreme differences were due to a poorly defined peak in the spectrum of the PCA atrial signal, and large residual ventricular components. Poorly defined peaks may arise because the fibrillatory signal can exhibit a relatively broad power spectrum indicating large variability in the frequency of the fibrillatory signal. Overall, the mean (standard deviation) differences in dominant fibrillatory frequency were 0.15 (1.02) Hz (p = PNS, one sample sign test) (STC—PCA), 0.04 (0.81) Hz (p = NS) (STC—ICA), and 0.11 (0.94) Hz (p = NS) (ICA—PCA). IV. DISCUSSION Noninvasive assessment of the atrial fibrillatory wave and particularly the fibrillation frequency is gaining acceptance as a tool for characterising the arrhythmia in individual patients and for assessing the impact of different treatment strategies [12]–[15]. Analysis of the atrial fibrillatory wave directly from the ECG is very limited because the atrial signal is largely obscured by the ventricular activity. Recently, new algorithms for the extraction of the atrial fibrillation signal from the ECG have become available [12], [19]–[21]. To ensure meaningful cross study comparisons, where different algorithms have been used, it is essential that the differences in the resulting atrial signals are fully explored. We have evaluated three such algorithms which represent two fundamentally different approaches: STC makes use of the property that the atrial and ventricular activities are uncoupled in time to each
346
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 2, FEBRUARY 2006
Fig. 3. Differences between algorithms in dominant atrial frequency between pairs of algorithms.
other, whereas the algorithms based on PCA or ICA exploit the property that the atrial and ventricular activities originate from different and uncoupled bioelectric sources. Also, STC derives the atrial signal from a specific ECG lead, whereas PCA and ICA derive a global atrial signal from all 12 leads. Because the reference signals used in this analysis were derived directly from V1, the analysis favors STC in this respect. Nonetheless, we have established a method of direct comparison between algorithms in specific leads which was not done previously. One strength of the present study was that each of the investigated algorithms was implemented at the center at which it was developed. Our evaluation has focused on the main characteristics of amplitude and frequency of the atrial fibrillation signal. Amplitude was important because it allowed us to quantify the effect of the extraction algorithms on the reference atrial signal. Frequency was important because the dominant atrial frequency is the key characteristic of the fibrillatory waveform used in clinical application. Overall, the results show that STC is the least distortive of the algorithms. It is encouraging to see that the relatively new algorithms, PCA and ICA, show good agreement with reference data. These algorithms offer the possibility of providing additional information because they contain information from all leads. Users of these algorithms need to be observant of large residual ventricular activities which can affect subsequent analysis. V. CONCLUSION There is a growing need for noninvasive tools which allow treatments of atrial fibrillation to be evaluated in terms of their effect on electrophysiological properties of the atria. Therefore, algorithms for extracting the body surface atrial signal in the presence of ventricular activity are very important. We can conclude that the extracted atrial signals from these algorithms exhibit very similar amplitude and frequency characteristics, but users should be observant of residual ventricular activities which can affect the analysis of the fibrillatory waveform in clinical practice. REFERENCES [1] V. Fuster, L. E. Ryden, R. W. Asinger, D. S. Cannom, H. J. Crijns, R. L. Frye, J. L. Halperin, G. N. Kay, W. W. Klein, S. Lévy, R. L. McNamara, E. L. Prystowsky, L. S. Wann, and D. G. Wyse, “ACC/AHA/ESC guidelines for the management of patients with atrial fibrillation: A report of the American college of cardiology/American Heart Association task force on practice guidelines and the european society of cardiology committee for practice guidelines and policy conferences (Committee to develop guidelines for the management of patients with atrial fibrillation),” in Eur. Heart J., vol. 22, 2001, pp. 1852–1923.
[2] A. Boldt, U. Wetzel, J. Lauschke, J. Weigl, J. Gummert, G. Hindricks, H. Kottkamp, and S. Dhein, “Fibrosis in left atrial tissue of patients with atrial fibrillation with and without underlying mitral valve disease,” Heart, vol. 90, pp. 400–405, 2004. [3] M. Haissaguerre, P. Jais, and D. C. Shah, “Spontaneous initiation of atrial fibrillation by ectopic beats originating in the pulmonary veins,” N. Engl. J. Med., vol. 339, pp. 659–666, 1998. [4] M. C. E. F. Wijffels, C. J. H. J. Kirchhof, R. Dorland, and M. A. Allessie, “Atrial fibrillation begets atrial fibrillation. A study in awake chronically instrumented goats,” Circulation, vol. 92, pp. 1954–1968, 1995. [5] J. E. Slocum and K. M. Ropella, “Correspondence between the frequency domain characteristics of simultaneous surface and intra-atrial recordings of atrial fibrillation,” in Comput. Cardiol. Los Alamitos, CA: IEEE Computer Society, 1994, pp. 781–784. [6] M. Holm, S. Pehrsson, M. Ingemansson, L. Sörnmo, R. Johansson, L. Sandhall, M. Sunemark, B. Smideberg, C. Olsson, and S. B. Olsson, “Non-invasive assessment of atrial refractoriness during atrial fibrillation in man—introducing, validating and illustrating a new ECG method,” Cardiovasc. Res., vol. 38, pp. 69–81, 1998. [7] S. A. Chen and C. T. Tai, “Is analysis of fibrillatory waves useful for treatment of atrial fibrillation?,” J. Cardiovasc. Electrophysiol., vol. 15, pp. 918–919, 2004. [8] A. Bollmann, K. Sonne, H. D. Esperer, I. Toepffer, and H. U. Klein, “Circadian variations in atrial fibrillatory frequency in persistent human atrial fibrillation,” Pacing Clin. Electrophysiol., vol. 23, pp. 1867–1871, 2000. [9] C. J. Meurling, J. E. Waktare, F. Holmqvist, A. Hedman, A. J. Camm, S. B. Olsson, and M. Malik, “Diurnal variations of the dominant cycle length of chronic atrial fibrillation,” Europace, vol. 1, pp. H401–H406, 1999. [10] C. J. Meurling, M. P. Ingemansson, R. A. Roijer, J. Carlson, C. J. Lindholm, B. Smideberg, L. Sörnmo, M. Stridh, and S. B. Olsson, “Attenuation of electrical remodeling in chronic atrial fibrillation following oral treatment with verapamil,” Am. J. Physiol. Heart Circ. Physiol., vol. 280, pp. 234–241, 2001. [11] S. Petrutiu, A. V. Sahakian, J. Ng, and S. Swiryn, “Analysis of the surface electrocardiogram to predict termination of atrial fibrillation: The 2004 Computers in Cardiology/PhysioNet challenge,” Comput. Cardiol., vol. 31, pp. 105–108, 2004. [12] D. Raine, P. Langley, A. Murray, A. Dunuwille, and J. P. Bourke, “Surface atrial frequency analysis in patients with atrial fibrillation: A tool for evaluating the effects of intervention,” J. Cardiovasc. Electrophysiol., vol. 15, pp. 1021–1026, 2004. [13] D. Husser, M. Stridh, L. Sörnmo, P. Platonov, S. B. Olsson, and A. Bollman, “Analysis of the surface electrocardiogram for monitoring and predicting antiarrhythmic drug effects in atrial fibrillation,” Cardiovasc. Drugs Ther., vol. 18, pp. 377–386, 2004. [14] A. Bollmann, M. Mende, A. Neugebauer, and D. Pfeiffer, “Atrial fibrillatory frequency predicts atrial defibrillation threshold and early arrhythmia recurrence in patients undergoing internal cardioversion of persistent atrial fibrillation,” Pacing Clin. Electrophysiol., vol. 25, pp. 1179–1184, 2002. [15] C. T. Tai, S. A. Chen, A. S. Liu, W. C. Yu, Y. A. Ding, M. S. Chang, and T. Kao, “Spectral analysis of chronic atrial fibrillation and its relation to minimal defibrillation energy,” Pacing Clin. Electrophysiol., vol. 25, pp. 1747–1751, 2002. [16] M. Haissaguerre, P. Sanders, M. Hocini, L. Hsu, D. C. Shah, C. Scavée, Y. Takahashi, M. Rotter, J. Pasquié, S. Garrigue, J. Clémenty, and P. Jais, “Changes in atrial fibrillation cycle length and inducibility during catheter ablation and their relation to outcome,” Circulation, vol. 109, pp. 3007–3013, 2004. [17] D. S. Rosenbaum and R. J. Cohen, “Frequency-based measures of atrial fibrillation in man,” in Proc. IEEE Engineering in Medicine and Biology Soc., 1990, pp. 582–583. [18] M. Stridh and L. Sörnmo, “Spatiotemporal QRST cancellation techniques for analysis of atrial fibrillation,” IEEE Trans. Biomed. Eng., vol. 48, no. 1, pp. 105–111, Jan. 2001. [19] P. Langley, J. P. Bourke, and A. Murray, “Frequency analysis of atrial fibrillation,” Comput. Cardiol., vol. 27, pp. 65–68, 2000. [20] J. J. Rieta, V. Zarzoso, J. Millet, R. Garcia, and R. Ruiz, “Atrial activity extraction based on blind source separation as an alternative QRST cancellation for atrial fibrillation analysis,” Comput. Cardiol., vol. 27, pp. 69–72, 2000. [21] J. J. Rieta, F. Castells, C. Sanchez, V. Zarzoso, and J. Millet, “Atrial activity extraction for atrial fibrillation analysis using blind source separation,” IEEE Trans. Biomed. Eng., vol. 51, no. 7, pp. 1176–1186, Jul. 2004.
