- Email: [email protected]
The effect of QRS cancellation on atrial fibrillatory wave signal characteristics in the surface electrocardiogram
Journal of Electrocardiology Vol. 36 No. 3 2003
The Effect of QRS Cancellation on Atrial Fibrillatory Wave Signal Characteristics in the Surface Electrocardiogram
Qin Xi, MS, Alan V. Sahakian, PhD, and Steven Swiryn, MD
Abstract: QRS cancellation methods have been used to analyze atrial activity in the electrocardiogram for such rhythms as atrioventricular dissociated ventricular tachycardia and atrial fibrillation. However, how well the cancellation methods work has never been evaluated by some gold standard. In this study of patients undergoing radiofrequency ablation of the atrioventricular junction, the contribution of imperfect cancellation was evaluated by comparing the “pure” atrial fibrillation (the gold standard) during a brief ventricular asystole to data obtained by a cancellation method during pacing just before and after the asystole. The results were compared by linear regression. The peak frequencies were 4.8-7.3 (6.1 ⫾ 0.8) Hz for the “pure” and 4.8-6.8 (5.9 ⫾ 0.7) Hz for the cancelled electrocardiogram segments (R2 ⫽ 0.89) (similar results for median frequency), and the mean short-time Fourier transform peak frequencies were 4.6-7.1 (5.9 ⫾ 0.8) Hz for the “pure” and 4.7-6.8 (5.9 ⫾ 0.7) Hz for the cancelled segments (R2 ⫽ 0.96). Further comparison was accomplished using synthesized signals. Based on our study, the cancellation method is reliable for studying atrial fibrillatory wave characteristics. As reported previously, the peak frequency and most power for atrial fibrillation in humans are in the 4-9 Hz band. Key words: Atrial fibrillation; signal processing, arrhythmia detection, power spectrum.
Atrial fibrillation is the most common arrhythmia with a prevalence of approximately 0.4% to 1.0% in the general population. Prevalence increases with age and more so in the older age groups; it is associated with a high probability of coexisting
cardiovascular or systemic disease and with an increased risk of stroke and mortality, as well as impaired exercise tolerance, fatigue, and heart failure (1,2). The electrocardiogram (ECG) diagnosis of atrial fibrillation is made by the absence of discrete P waves, which are replaced by rapid (traditionally more than 350 beats/min) fibrillatory waves of varying timing and morphology. Fibrillatory waves can be characterized as to amplitude and such measures as peak or median frequency, (3–5) which in turn allows the estimation of the underlying degree of organization (6 – 8) and effects of drugs (9,10) or autonomic tone (11). To quantitatively evaluate fibrillatory waves, it is
From the Department of Biomedical Engineering, Evanston, IL, Electrical and Computer Engineering, Northwestern University, Evanston, IL; and Evanston/Northwestern Healthcare, Evanston, IL. This work was supported in part by grants from the Dr. Scholl Foundation and the O’Shaughnessy Foundation. Reprint requests: Steven Swiryn, MD, Evanston Hospital, Room 300 Burch Hall, 2650 Ridge Ave, Evanston, IL; e-mail: [email protected]. © 2003 Elsevier Inc. All rights reserved. 0022-0736/03/3603-0009$30.00/0 doi:10.1016/S0022-0736(03)00046-3
243
244 Journal of Electrocardiology Vol. 36 No. 3 July 2003 necessary to remove the contribution of the QRS complexes and T/U waves from the ECG. This is typically accomplished by some cancellation process, (3–5) which results in a remainder ECG with almost exclusively fibrillatory waves. However, cancellation is always imperfect, and the residual QRS complexes and T waves in the remainder ECG as well as any noise in the data may persist in the signal. Some improvements in cancellation methods have been reported (12). No matter how good the cancellation is, however, there remains a concern that the residual ventricular signal or some artificial distortions brought in by the cancellation method itself may alter the results. Therefore, to evaluate the effect of the cancellation procedure on characterization of f waves, we studied “pure” atrial fibrillation obtained by briefly stopping pacing in patients with complete AV block after radiofrequency ablation of the AV junction and compared the results with the adjacent remainder ECG obtained by cancellation of paced beats just before and after the “pure” segment. In addition, we synthesized ECGs using the “pure” fibrillation segments plus QRS complexes and T waves extracted from ECGs showing sinus rhythm to further evaluate cancellation by correlation.
Materials and Methods
segments were recorded long enough (longer than 6 beats) to get a reliable median beat by the template matching method for QRS cancellation. Thus, for each patient, 3 contiguous segments were recorded: ventricular pacing, ventricular asystole with “pure” fibrillation, and then ventricular pacing again. Some patients had a second asystole also available for analysis. All the segments were extracted from the physiologic recording system (sampling rate: 977 Hz) and saved on floppy disks. Lead V1 has the highest ratio of atrial fibrillation signal amplitude to QRS amplitude; therefore, it was chosen for the analysis. The dc component was eliminated before calculation. A band-pass filter with cutoff frequencies of 50 Hz and 1 Hz was applied to avoid power line interference and baseline wander. Cancellation For the ECG segments with QRS complexes, the first step was to detect fiducial points and to align all the beats in that segment to generate a median beat (3). Next, this median beat was aligned at each fiducial point and subtracted from the ECG to obtain the remainder ECG, which was the result of cancellation, composed of almost exclusively fibrillatory waves. The fiducial point detection method was modified from the algorithm presented by Pan and Tompkins (13).
Patient Selection
Power Spectrum Analysis
Patients for this study were undergoing routine radiofrequency ablation of the AV junction for ventricular rate control of atrial fibrillation. Once AV block was achieved, a ventricular paced rhythm was present. To test for the persistence of complete heart block and to evaluate the presence of escape rhythms, pacing was briefly discontinued and then restarted. During this procedure, the surface ECG was continuously digitized on a physiologic recording system (Prucka; GE Medical, Inc, Milwaukee, WI). The project was reviewed and approved by the Institutional Review Board of Evanston Northwestern Healthcare.
For both “pure” and cancelled remainder ECGs, both ends of the signals were padded with zeros to increase frequency resolution before a discrete Fourier transform was performed to obtain the power spectrum. For this study, the frequency resolution was kept ⬍0.1 Hz. For each patient, the same frequency resolution was kept for the “pure” and corresponding cancelled segments. The peak frequency (defined as the frequency with highest power in the 4-9 Hz band) was identified. Other parameters such as median frequency (defined as the frequency obtained by dividing the 4-9 Hz region power in half) and percent power in the 4-9 Hz band (defined as the ratio of power in the 4-9 Hz region to the total power in the power spectrum) (3) were also calculated. However, atrial fibrillation lacks stationarity and even a several second difference in time might result in true differences in the f waves. In addition, these three contiguous segments did not have the same length. In some cases, the power spectrum of the cancelled ECG might contain multiple peaks
Data Preprocessing The “pure” atrial fibrillation segments were delineated as starting from after the T wave of the last paced beat, and ending before the first restarted pacing spike. The adjacent surface ECGs with paced QRS complexes just before and after the “pure”
Atrial Fibrillatory Waves •
with similar amplitudes. In some sub-segment, one of the peaks was relatively dominant, while in a nearby sub-segment, another peak might be selected with only a slight change in amplitude. Thus, to study the temporal fluctuation of peak frequency, a short-time Fourier transform (STFT) method was applied. The signal was segmented by a short window (length 1 s, window moving step 0.1 s) and the peak frequency in each window was computed, the mean of which was called the mean STFT peak frequency. The standard deviation of STFT peak frequency was calculated to study the variability of peak frequency. Similarly, mean STFT median frequency and its standard deviation were defined and studied. Comparison of “Pure” and Cancelled Data The average of the peak frequencies of the remainder ECG segments just before and just after the “pure” segment was calculated for comparison with the peak frequency of the “pure” segment using linear regression. For patients with 2 “pure” segments, the 1st “pure” segment and its paced counterparts were selected for this comparison. The median frequency, the mean STFT peak frequency, and the mean STFT median frequency were compared in a similar way. Correlation Study With Synthesized Signals As stated above, atrial fibrillation is not stationary and these 3 segments are contiguous, not simulta-
Xi et al. 245
neous, so the difference in “pure” and cancelled signals may be partially caused by true signal variation instead of imperfect cancellation. To further evaluate the effects of cancellation with identical f waves for both cancelled and the “pure” data segments, we synthesized an atrial fibrillation ECG signal. To obtain adequate signal duration for cancellation, the “pure” segment was repeated as necessary. ECGs of sinus rhythm (lead V1, with the same signal length as the repeated “pure” segment) were selected as the source of QRS complexes and T waves with all the P waves removed. R-R intervals were arbitrarily varied within 10% of the mean R-R interval (10% was chosen based on previous study of Kirsh et al. (14). In addition, amplitudes of QRS complexes and T waves were also varied randomly within 10% of the mean amplitude. Finally, the series of modified QRS complexes and T waves was superimposed onto the “pure” fibrillation signal to construct a synthesized ECG of atrial fibrillation with a gold standard for the f waves. For each “pure” signal, modified QRS complexes and T waves of several different sinus rhythm patients were used for testing. Cancellation was performed on each of the synthesized fibrillation ECG signals, and comparison of the “pure” and corresponding cancelled remainder segments was accomplished by correlation in both the time and frequency domains.
Results A total of 12 patients (3 men and 9 women, ages 49 to 88 (7313) with atrial fibrillation were in-
Fig. 1. An example of three contiguous signal recordings. (A) Preceding, “pure,” and after original ECG segments, (B) the corresponding remainders (for the “pure” segment, the “remainder” shown is the filtered “pure” segment itself), and (C) power spectra of remainders.
246 Journal of Electrocardiology Vol. 36 No. 3 July 2003 cluded in this study. The duration of data segments of “pure” atrial fibrillation ranged from 1.8 to 4.5 s (2.9 ⫾ 0.9 s) and the duration of the segments with QRS complexes just before and after the “pure” segment fell in the range of 4.5 to 13.5 s (7.6 ⫾ 2.6 s). Five patients had a 2nd “pure” segment. The time interval between 2 “pure” segments ranged from 16 to 126 s (80 ⫾ 45 s). Figure 1 is an example of 3 contiguous recordings for a representative patient. The representative “pure” atrial fibrillation segment and the ECGs with paced QRS complexes just before and after this “pure” segment are in Figure 1A, with their corresponding remainder ECGs and power spectra of the remainders shown in Figs 1B and C. To compare all 3 segments, the “remainder” of the “pure” segment is shown as the filtered “pure” segment itself. Note that there are some visible QRS-T residuals in the remainders such as in the following segment near 7.3 s and 11 s. The peak frequencies for these 3 segments were quite similar for this patient.
Fig. 3. Comparison of the standard deviation of STFT peak frequency for the “pure” and the average of cancelled segments. Each patient is represented by a pair of data points with the 1st dot representing the “pure” and the 2nd dot representing the average of the cancelled data.
Comparison of “Pure” and Cancelled Data Figure 2 is the comparison of spectral parameters of the “pure” and the average of the 2 cancelled segments for each patient. The “pure” and cancelled data were strongly correlated. A plot of the peak frequencies (in Hz) for all the patients is shown in Figure 2A. The peak frequencies fell in the range of 4.8-7.3 Hz (6.1⫾ 0.8 Hz) for the “pure” segments and 4.8-6.8 Hz (5.9 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.89). The slope of the fitted line was 0.79, and the vertical-axis intercept was 1.10 Hz. A similar result for median frequency is shown in Figure 2B. The median frequencies ranged from 4.8 to 7.3 Hz (5.4 ⫾ 0.8 Hz) for the “pure” segments and 4.8-6.7 Hz (5.8 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.91). The slope of the
Fig. 2. Comparison of spectral parameters of the “pure” versus the average of the 2 cancelled segments for each patient. (A) Peak frequency, (B) median frequency, (C) mean STFT peak frequency, and (D) mean STFT median frequency comparison. For each graph, the horizontal axis represents the value for the “pure” segment, and the vertical axis represents the average value for the adjacent cancelled ECG segments. Each patient is shown by a dot, and the line is the fit of linear regression.
Fig. 4. Comparison of percent power in the 4-9 Hz band for the “pure” and the average of the 2 cancelled segments.
Atrial Fibrillatory Waves •
Fig. 5. Mean STFT peak frequency for both “pure” segments for the 5 patients having 2 “pure” segments.
fitted line was 0.78, and the vertical-axis intercept was 1.14 Hz. The result for mean STFT peak frequency is shown in Figure 2C. The mean STFT peak frequencies varied in the range of 4.6-7.1 Hz (5.9 ⫾ 0.8 Hz) for the “pure” segments and 4.7-6.8 Hz (5.9 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.96). Compared to the peak frequency study without short-term data, this gave a closer fit with a slope near 1 (0.93) and vertical-axis intercept near 0 (0.41 Hz). The result for the mean STFT median frequency is shown in Figure 2D with data similar to those in Figure 2B. To study the variability of the peak frequency
Xi et al. 247
with time, the standard deviation of STFT peak frequency for the “pure” segments and the average of the 2 cancelled segments were calculated and compared; Figure 3 shows the result. For most cases, the standard deviation of cancelled data was larger than that of the “pure” segment. Similar results were found for standard deviation of STFT median frequency (not shown). Figure 4 is the percent power in the 4-9 Hz band for the “pure” segments and the average of the 2 cancelled segments. For both “pure” and cancelled segments, this percent power was no less than 55% (62%-87% for all the “pure” and 55%-84% for all the cancelled segments). Note that for all cases but one, the “pure” segment had a slightly higher percent power compared to the remainder ECG obtained by cancellation.
Patients With 2 “Pure” Segments For 5 patients, data were available during a second brief ventricular asystole. Therefore, we compared the mean STFT peak frequency for both segments. Figure 5 is the mean STFT peak frequency for 2 different recorded “pure” segments for the 5 patients with 2 “pure” segments. Absolute change in the mean STFT peak frequency from the 1st to 2nd segments ranged from 0.0 to 0.7 Hz, with
Fig. 6. A sample result of the synthesized ECG. (A) Modified sinus rhythm QRS and T waves, (B) the repeated “pure” segment, and (C) its power spectrum. (D) The resulted synthesized ECG of atrial fibrillation. The corresponding (E) cancelled remainder with its (F) power spectrum.
248 Journal of Electrocardiology Vol. 36 No. 3 July 2003 correlation between the “pure” and the cancelled synthesized data ranged from 0.85 to 0.98. For the power spectra, correlation was in the range of 0.90-0.98. Note that for all patients, the power spectrum correlation was slightly higher than the corresponding signal correlation.
Discussion
Fig. 7. Signal and power spectrum correlations of “pure” and cancelled synthesized signals.
only 1 patient having a large difference, but the time interval between 2 segments for this patient was only 16 s. Synthesized ECG Data Figure 6 is an example of the synthesized ECG of atrial fibrillation. Modified sinus rhythm QRS complexes and T waves (Fig. 6A) were superimposed on the repeated “pure” segment (Fig. 6B), which led to the result of the synthesized ECG of atrial fibrillation (Fig. 6D). The synthesized signal segment was 10 s long for each of the original 12 patients. The cancelled remainder is shown in (Fig. 6E). The power spectra of “pure” and cancelled synthesized segments are shown in (Fig. 6C) and (Fig. 6F), respectively. For most patients, the morphology of the power spectrum of the cancelled synthesized signals is remarkably similar to that of their corresponding “pure” signals. The “pure” segment of each patient was tested with 2 different synthesized ECGs; Figure 7 is the result of 1 of the tests. For the signals themselves,
In this study, by comparing the “pure” segment and its adjacent cancelled paced ECG counterparts, cancellation of QRS complexes was shown not to change the frequency domain characteristics of atrial fibrillatory signals qualitatively, and was associated with only minor quantitative changes. The cancellation result was good for synthesized signals in both the time and frequency domains. Therefore, the use of QRS cancellation for the purpose of characterizing atrial signals is validated. However, minor quantitative differences were found between the “pure” and the adjacent cancelled ECG counterparts, including differences in the peak frequency, median frequency, standard deviation of the STFT peak frequency, and the percent power in 4-9 Hz. There are at least 2 potential explanations for this. First, imperfect cancellation can introduce small errors by leaving uncancelled ventricular signals or over-subtracting atrial activity. Because of the limited number of beats available to generate the median beat, the atrial activity cannot be completely eliminated in the median beat. Thus, when the median beat is subtracted from the original ECG, incomplete ventricular activity plus part of the atrial activity is removed from the ECG. This is especially obvious in short-term data, such as STFT peak/median frequency, shown by its standard deviation. Shortterm data for peak frequency seems to characterize atrial fibrillation in a more accurate way, but shortterm data for median frequency does not give a better fit. This may suggest that the median frequency is less suitable for atrial fibrillation characterization. Second, atrial fibrillation lacks stationarity, and changes in the atrial fibrillation signal itself over the time scale of a few seconds may have contributed to measured differences between the “pure” segments and the preceding and following cancelled counterparts. “Simultaneous” study of cancellation using synthesized fibrillation ECGs gave additional support for the cancellation method, avoiding the problem of nonstationarity. Strong correlation between the “pure” and the corresponding cancelled
Atrial Fibrillatory Waves •
synthesized ECG existed for both the remainder signals and their power spectra. Correlation of power spectra was strongest, which suggested that characterization of atrial fibrillation in the frequency domain is robust and reliable.
Xi et al. 249
for providing the data from the ablation procedures. In addition, the authors thank K. Carrol from Evanston Northwestern Healthcare for his help with Prucka system, and Jason Ng for his constructive suggestions and discussion.
Study Limitations References There are several limitations to the present study. First, the “pure” segment was obtained by briefly stopping pacing after the AV junction ablation, when atrial signals were not obscured by ventricular signals. Under this circumstance, blood pressure drops, and there may be changes in autonomic tone and chamber size. However, we did not detect any progressive change in short-term peak frequency with time once pacing was stopped. Second, atrial fibrillation is not stationary, and comparison of “pure” and adjacent cancelled ECG counterparts included both the influence of imperfect cancellation and non-stationarity of atrial fibrillation. For the synthesized signal, although without any influence of nonstationarity, it is clear that random variability within 10% of the QRS and T-waves amplitudes and beat-to-beat intervals does not reproduce the real physiological case of atrial fibrillation.
Conclusions Frequency domain features for both “pure” atrial fibrillation signals (gold standard) and adjacent remainder ECGs obtained by the cancellation method were studied, and the results from cancellation were quite comparable to those from the gold standard. Additional evidence from studies of synthesized signals in both the time and frequency domains strengthened our confidence in the cancellation method. Thus, the remainder ECG created by the cancellation method can be used as a reliable representation to characterize atrial activity. In addition, peak frequency and most power for all the patients for both “pure” and cancelled ECG segments did fall in a range of 4-9 Hz, which confirmed previous studies of atrial fibrillation (5– 7).
Acknowledgment The authors thank Drs W. Fisher, M. Kim, and K. Krishnan from Evanston Northwestern Healthcare
1. Fuster V, Ryden LE (co-chairs), et al: ACC/AHA/ESC guidelines for the management of patients with atrial fibrillation. J Am Coll Cardiol 38:1231, 2001 2. Brand FN, Abott RD, Kannel WB: Characteristics and prognosis of lone atrial fibrillation: 30-years follow up in the Framingham study. JAMA 254:3449, 1985 3. Slocum J, Byrom E, McCarthy L, et al: Computer detection of atrioventricular dissociation from the surface electrocardiogram during wide QRS tachycardias. Circulation 72:1028, 1985 4. Slocum J, Sahakian A, Swiryn S: Diagnosis of atrial fibrillation from surface electrocardiograms based on computer-detected atrial activity. J Electrocardiol 25:1, 1992 5. Shkurovich S, Sahakian A, Swiryn S: Detection of atrial activity from high-voltage leads of implantable ventricular defibrillators using a cancellation technique. IEEE Trans Biomed Eng 45:229, 1998 6. Ropella KM, Sahakian AV, Baerman JM, et al: The coherence spectrum: A quantitative discriminator of fibrillatory and non-fibrillatory cardiac rhythms. Circulation 80:112, 1989 7. Botteron GW, Smith JM: Quantitative assessment of the spatial organization of atrial fibrillation in the intact human heart. Circulation 93:513, 1996 8. Koning KTS, Smeets JLRM, Penn OC, et al: Configuration of unipolar atrial electrograms during electrically induced atrial fibrillation in humans. Circulation 95:1231, 1997 9. Ropella KM, Sahakian AV, Baerman JM, et al: Effects of Procainamide on intra-arterial electrograms during atrial fibrillation-implication for detection algorithms. Circulation 77:1047, 1988 10. Bollmann A, Kanuru NK, McTeague KK, et al: Frequency analysis of human atrial fibrillation using the surface electrocardiogram and its response to ibutilide. Am J Cardiol 81:1439, 1998 11. Sih HJ: Coherence mapping of cardiac tissue and its application to atrial fibrillation. Doctoral thesis, Evanston, IL, Northwestern University, 1994 12. Stridh M, Sornmo L: Spatiotemporal QRST cancellation techniques for analysis of atrial fibrillation. IEEE Trans Biomed Eng 48:105, 2001 13. Pan J, Tompkins WJ: A real-time QRS detection algorithm. IEEE Trans Biomed Eng 32:228, 1985 14. Kirsh JA, Sahakian AV, Baerman JM, et al: Ventricular response to atrial fibrillation: Role of atrioventricular conduction pathways. JACC 12:1265, 1988
The Effect of QRS Cancellation on Atrial Fibrillatory Wave Signal Characteristics in the Surface Electrocardiogram
Qin Xi, MS, Alan V. Sahakian, PhD, and Steven Swiryn, MD
Abstract: QRS cancellation methods have been used to analyze atrial activity in the electrocardiogram for such rhythms as atrioventricular dissociated ventricular tachycardia and atrial fibrillation. However, how well the cancellation methods work has never been evaluated by some gold standard. In this study of patients undergoing radiofrequency ablation of the atrioventricular junction, the contribution of imperfect cancellation was evaluated by comparing the “pure” atrial fibrillation (the gold standard) during a brief ventricular asystole to data obtained by a cancellation method during pacing just before and after the asystole. The results were compared by linear regression. The peak frequencies were 4.8-7.3 (6.1 ⫾ 0.8) Hz for the “pure” and 4.8-6.8 (5.9 ⫾ 0.7) Hz for the cancelled electrocardiogram segments (R2 ⫽ 0.89) (similar results for median frequency), and the mean short-time Fourier transform peak frequencies were 4.6-7.1 (5.9 ⫾ 0.8) Hz for the “pure” and 4.7-6.8 (5.9 ⫾ 0.7) Hz for the cancelled segments (R2 ⫽ 0.96). Further comparison was accomplished using synthesized signals. Based on our study, the cancellation method is reliable for studying atrial fibrillatory wave characteristics. As reported previously, the peak frequency and most power for atrial fibrillation in humans are in the 4-9 Hz band. Key words: Atrial fibrillation; signal processing, arrhythmia detection, power spectrum.
Atrial fibrillation is the most common arrhythmia with a prevalence of approximately 0.4% to 1.0% in the general population. Prevalence increases with age and more so in the older age groups; it is associated with a high probability of coexisting
cardiovascular or systemic disease and with an increased risk of stroke and mortality, as well as impaired exercise tolerance, fatigue, and heart failure (1,2). The electrocardiogram (ECG) diagnosis of atrial fibrillation is made by the absence of discrete P waves, which are replaced by rapid (traditionally more than 350 beats/min) fibrillatory waves of varying timing and morphology. Fibrillatory waves can be characterized as to amplitude and such measures as peak or median frequency, (3–5) which in turn allows the estimation of the underlying degree of organization (6 – 8) and effects of drugs (9,10) or autonomic tone (11). To quantitatively evaluate fibrillatory waves, it is
From the Department of Biomedical Engineering, Evanston, IL, Electrical and Computer Engineering, Northwestern University, Evanston, IL; and Evanston/Northwestern Healthcare, Evanston, IL. This work was supported in part by grants from the Dr. Scholl Foundation and the O’Shaughnessy Foundation. Reprint requests: Steven Swiryn, MD, Evanston Hospital, Room 300 Burch Hall, 2650 Ridge Ave, Evanston, IL; e-mail: [email protected]. © 2003 Elsevier Inc. All rights reserved. 0022-0736/03/3603-0009$30.00/0 doi:10.1016/S0022-0736(03)00046-3
243
244 Journal of Electrocardiology Vol. 36 No. 3 July 2003 necessary to remove the contribution of the QRS complexes and T/U waves from the ECG. This is typically accomplished by some cancellation process, (3–5) which results in a remainder ECG with almost exclusively fibrillatory waves. However, cancellation is always imperfect, and the residual QRS complexes and T waves in the remainder ECG as well as any noise in the data may persist in the signal. Some improvements in cancellation methods have been reported (12). No matter how good the cancellation is, however, there remains a concern that the residual ventricular signal or some artificial distortions brought in by the cancellation method itself may alter the results. Therefore, to evaluate the effect of the cancellation procedure on characterization of f waves, we studied “pure” atrial fibrillation obtained by briefly stopping pacing in patients with complete AV block after radiofrequency ablation of the AV junction and compared the results with the adjacent remainder ECG obtained by cancellation of paced beats just before and after the “pure” segment. In addition, we synthesized ECGs using the “pure” fibrillation segments plus QRS complexes and T waves extracted from ECGs showing sinus rhythm to further evaluate cancellation by correlation.
Materials and Methods
segments were recorded long enough (longer than 6 beats) to get a reliable median beat by the template matching method for QRS cancellation. Thus, for each patient, 3 contiguous segments were recorded: ventricular pacing, ventricular asystole with “pure” fibrillation, and then ventricular pacing again. Some patients had a second asystole also available for analysis. All the segments were extracted from the physiologic recording system (sampling rate: 977 Hz) and saved on floppy disks. Lead V1 has the highest ratio of atrial fibrillation signal amplitude to QRS amplitude; therefore, it was chosen for the analysis. The dc component was eliminated before calculation. A band-pass filter with cutoff frequencies of 50 Hz and 1 Hz was applied to avoid power line interference and baseline wander. Cancellation For the ECG segments with QRS complexes, the first step was to detect fiducial points and to align all the beats in that segment to generate a median beat (3). Next, this median beat was aligned at each fiducial point and subtracted from the ECG to obtain the remainder ECG, which was the result of cancellation, composed of almost exclusively fibrillatory waves. The fiducial point detection method was modified from the algorithm presented by Pan and Tompkins (13).
Patient Selection
Power Spectrum Analysis
Patients for this study were undergoing routine radiofrequency ablation of the AV junction for ventricular rate control of atrial fibrillation. Once AV block was achieved, a ventricular paced rhythm was present. To test for the persistence of complete heart block and to evaluate the presence of escape rhythms, pacing was briefly discontinued and then restarted. During this procedure, the surface ECG was continuously digitized on a physiologic recording system (Prucka; GE Medical, Inc, Milwaukee, WI). The project was reviewed and approved by the Institutional Review Board of Evanston Northwestern Healthcare.
For both “pure” and cancelled remainder ECGs, both ends of the signals were padded with zeros to increase frequency resolution before a discrete Fourier transform was performed to obtain the power spectrum. For this study, the frequency resolution was kept ⬍0.1 Hz. For each patient, the same frequency resolution was kept for the “pure” and corresponding cancelled segments. The peak frequency (defined as the frequency with highest power in the 4-9 Hz band) was identified. Other parameters such as median frequency (defined as the frequency obtained by dividing the 4-9 Hz region power in half) and percent power in the 4-9 Hz band (defined as the ratio of power in the 4-9 Hz region to the total power in the power spectrum) (3) were also calculated. However, atrial fibrillation lacks stationarity and even a several second difference in time might result in true differences in the f waves. In addition, these three contiguous segments did not have the same length. In some cases, the power spectrum of the cancelled ECG might contain multiple peaks
Data Preprocessing The “pure” atrial fibrillation segments were delineated as starting from after the T wave of the last paced beat, and ending before the first restarted pacing spike. The adjacent surface ECGs with paced QRS complexes just before and after the “pure”
Atrial Fibrillatory Waves •
with similar amplitudes. In some sub-segment, one of the peaks was relatively dominant, while in a nearby sub-segment, another peak might be selected with only a slight change in amplitude. Thus, to study the temporal fluctuation of peak frequency, a short-time Fourier transform (STFT) method was applied. The signal was segmented by a short window (length 1 s, window moving step 0.1 s) and the peak frequency in each window was computed, the mean of which was called the mean STFT peak frequency. The standard deviation of STFT peak frequency was calculated to study the variability of peak frequency. Similarly, mean STFT median frequency and its standard deviation were defined and studied. Comparison of “Pure” and Cancelled Data The average of the peak frequencies of the remainder ECG segments just before and just after the “pure” segment was calculated for comparison with the peak frequency of the “pure” segment using linear regression. For patients with 2 “pure” segments, the 1st “pure” segment and its paced counterparts were selected for this comparison. The median frequency, the mean STFT peak frequency, and the mean STFT median frequency were compared in a similar way. Correlation Study With Synthesized Signals As stated above, atrial fibrillation is not stationary and these 3 segments are contiguous, not simulta-
Xi et al. 245
neous, so the difference in “pure” and cancelled signals may be partially caused by true signal variation instead of imperfect cancellation. To further evaluate the effects of cancellation with identical f waves for both cancelled and the “pure” data segments, we synthesized an atrial fibrillation ECG signal. To obtain adequate signal duration for cancellation, the “pure” segment was repeated as necessary. ECGs of sinus rhythm (lead V1, with the same signal length as the repeated “pure” segment) were selected as the source of QRS complexes and T waves with all the P waves removed. R-R intervals were arbitrarily varied within 10% of the mean R-R interval (10% was chosen based on previous study of Kirsh et al. (14). In addition, amplitudes of QRS complexes and T waves were also varied randomly within 10% of the mean amplitude. Finally, the series of modified QRS complexes and T waves was superimposed onto the “pure” fibrillation signal to construct a synthesized ECG of atrial fibrillation with a gold standard for the f waves. For each “pure” signal, modified QRS complexes and T waves of several different sinus rhythm patients were used for testing. Cancellation was performed on each of the synthesized fibrillation ECG signals, and comparison of the “pure” and corresponding cancelled remainder segments was accomplished by correlation in both the time and frequency domains.
Results A total of 12 patients (3 men and 9 women, ages 49 to 88 (7313) with atrial fibrillation were in-
Fig. 1. An example of three contiguous signal recordings. (A) Preceding, “pure,” and after original ECG segments, (B) the corresponding remainders (for the “pure” segment, the “remainder” shown is the filtered “pure” segment itself), and (C) power spectra of remainders.
246 Journal of Electrocardiology Vol. 36 No. 3 July 2003 cluded in this study. The duration of data segments of “pure” atrial fibrillation ranged from 1.8 to 4.5 s (2.9 ⫾ 0.9 s) and the duration of the segments with QRS complexes just before and after the “pure” segment fell in the range of 4.5 to 13.5 s (7.6 ⫾ 2.6 s). Five patients had a 2nd “pure” segment. The time interval between 2 “pure” segments ranged from 16 to 126 s (80 ⫾ 45 s). Figure 1 is an example of 3 contiguous recordings for a representative patient. The representative “pure” atrial fibrillation segment and the ECGs with paced QRS complexes just before and after this “pure” segment are in Figure 1A, with their corresponding remainder ECGs and power spectra of the remainders shown in Figs 1B and C. To compare all 3 segments, the “remainder” of the “pure” segment is shown as the filtered “pure” segment itself. Note that there are some visible QRS-T residuals in the remainders such as in the following segment near 7.3 s and 11 s. The peak frequencies for these 3 segments were quite similar for this patient.
Fig. 3. Comparison of the standard deviation of STFT peak frequency for the “pure” and the average of cancelled segments. Each patient is represented by a pair of data points with the 1st dot representing the “pure” and the 2nd dot representing the average of the cancelled data.
Comparison of “Pure” and Cancelled Data Figure 2 is the comparison of spectral parameters of the “pure” and the average of the 2 cancelled segments for each patient. The “pure” and cancelled data were strongly correlated. A plot of the peak frequencies (in Hz) for all the patients is shown in Figure 2A. The peak frequencies fell in the range of 4.8-7.3 Hz (6.1⫾ 0.8 Hz) for the “pure” segments and 4.8-6.8 Hz (5.9 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.89). The slope of the fitted line was 0.79, and the vertical-axis intercept was 1.10 Hz. A similar result for median frequency is shown in Figure 2B. The median frequencies ranged from 4.8 to 7.3 Hz (5.4 ⫾ 0.8 Hz) for the “pure” segments and 4.8-6.7 Hz (5.8 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.91). The slope of the
Fig. 2. Comparison of spectral parameters of the “pure” versus the average of the 2 cancelled segments for each patient. (A) Peak frequency, (B) median frequency, (C) mean STFT peak frequency, and (D) mean STFT median frequency comparison. For each graph, the horizontal axis represents the value for the “pure” segment, and the vertical axis represents the average value for the adjacent cancelled ECG segments. Each patient is shown by a dot, and the line is the fit of linear regression.
Fig. 4. Comparison of percent power in the 4-9 Hz band for the “pure” and the average of the 2 cancelled segments.
Atrial Fibrillatory Waves •
Fig. 5. Mean STFT peak frequency for both “pure” segments for the 5 patients having 2 “pure” segments.
fitted line was 0.78, and the vertical-axis intercept was 1.14 Hz. The result for mean STFT peak frequency is shown in Figure 2C. The mean STFT peak frequencies varied in the range of 4.6-7.1 Hz (5.9 ⫾ 0.8 Hz) for the “pure” segments and 4.7-6.8 Hz (5.9 ⫾ 0.7 Hz) for the cancelled ECGs (R2 ⫽ 0.96). Compared to the peak frequency study without short-term data, this gave a closer fit with a slope near 1 (0.93) and vertical-axis intercept near 0 (0.41 Hz). The result for the mean STFT median frequency is shown in Figure 2D with data similar to those in Figure 2B. To study the variability of the peak frequency
Xi et al. 247
with time, the standard deviation of STFT peak frequency for the “pure” segments and the average of the 2 cancelled segments were calculated and compared; Figure 3 shows the result. For most cases, the standard deviation of cancelled data was larger than that of the “pure” segment. Similar results were found for standard deviation of STFT median frequency (not shown). Figure 4 is the percent power in the 4-9 Hz band for the “pure” segments and the average of the 2 cancelled segments. For both “pure” and cancelled segments, this percent power was no less than 55% (62%-87% for all the “pure” and 55%-84% for all the cancelled segments). Note that for all cases but one, the “pure” segment had a slightly higher percent power compared to the remainder ECG obtained by cancellation.
Patients With 2 “Pure” Segments For 5 patients, data were available during a second brief ventricular asystole. Therefore, we compared the mean STFT peak frequency for both segments. Figure 5 is the mean STFT peak frequency for 2 different recorded “pure” segments for the 5 patients with 2 “pure” segments. Absolute change in the mean STFT peak frequency from the 1st to 2nd segments ranged from 0.0 to 0.7 Hz, with
Fig. 6. A sample result of the synthesized ECG. (A) Modified sinus rhythm QRS and T waves, (B) the repeated “pure” segment, and (C) its power spectrum. (D) The resulted synthesized ECG of atrial fibrillation. The corresponding (E) cancelled remainder with its (F) power spectrum.
248 Journal of Electrocardiology Vol. 36 No. 3 July 2003 correlation between the “pure” and the cancelled synthesized data ranged from 0.85 to 0.98. For the power spectra, correlation was in the range of 0.90-0.98. Note that for all patients, the power spectrum correlation was slightly higher than the corresponding signal correlation.
Discussion
Fig. 7. Signal and power spectrum correlations of “pure” and cancelled synthesized signals.
only 1 patient having a large difference, but the time interval between 2 segments for this patient was only 16 s. Synthesized ECG Data Figure 6 is an example of the synthesized ECG of atrial fibrillation. Modified sinus rhythm QRS complexes and T waves (Fig. 6A) were superimposed on the repeated “pure” segment (Fig. 6B), which led to the result of the synthesized ECG of atrial fibrillation (Fig. 6D). The synthesized signal segment was 10 s long for each of the original 12 patients. The cancelled remainder is shown in (Fig. 6E). The power spectra of “pure” and cancelled synthesized segments are shown in (Fig. 6C) and (Fig. 6F), respectively. For most patients, the morphology of the power spectrum of the cancelled synthesized signals is remarkably similar to that of their corresponding “pure” signals. The “pure” segment of each patient was tested with 2 different synthesized ECGs; Figure 7 is the result of 1 of the tests. For the signals themselves,
In this study, by comparing the “pure” segment and its adjacent cancelled paced ECG counterparts, cancellation of QRS complexes was shown not to change the frequency domain characteristics of atrial fibrillatory signals qualitatively, and was associated with only minor quantitative changes. The cancellation result was good for synthesized signals in both the time and frequency domains. Therefore, the use of QRS cancellation for the purpose of characterizing atrial signals is validated. However, minor quantitative differences were found between the “pure” and the adjacent cancelled ECG counterparts, including differences in the peak frequency, median frequency, standard deviation of the STFT peak frequency, and the percent power in 4-9 Hz. There are at least 2 potential explanations for this. First, imperfect cancellation can introduce small errors by leaving uncancelled ventricular signals or over-subtracting atrial activity. Because of the limited number of beats available to generate the median beat, the atrial activity cannot be completely eliminated in the median beat. Thus, when the median beat is subtracted from the original ECG, incomplete ventricular activity plus part of the atrial activity is removed from the ECG. This is especially obvious in short-term data, such as STFT peak/median frequency, shown by its standard deviation. Shortterm data for peak frequency seems to characterize atrial fibrillation in a more accurate way, but shortterm data for median frequency does not give a better fit. This may suggest that the median frequency is less suitable for atrial fibrillation characterization. Second, atrial fibrillation lacks stationarity, and changes in the atrial fibrillation signal itself over the time scale of a few seconds may have contributed to measured differences between the “pure” segments and the preceding and following cancelled counterparts. “Simultaneous” study of cancellation using synthesized fibrillation ECGs gave additional support for the cancellation method, avoiding the problem of nonstationarity. Strong correlation between the “pure” and the corresponding cancelled
Atrial Fibrillatory Waves •
synthesized ECG existed for both the remainder signals and their power spectra. Correlation of power spectra was strongest, which suggested that characterization of atrial fibrillation in the frequency domain is robust and reliable.
Xi et al. 249
for providing the data from the ablation procedures. In addition, the authors thank K. Carrol from Evanston Northwestern Healthcare for his help with Prucka system, and Jason Ng for his constructive suggestions and discussion.
Study Limitations References There are several limitations to the present study. First, the “pure” segment was obtained by briefly stopping pacing after the AV junction ablation, when atrial signals were not obscured by ventricular signals. Under this circumstance, blood pressure drops, and there may be changes in autonomic tone and chamber size. However, we did not detect any progressive change in short-term peak frequency with time once pacing was stopped. Second, atrial fibrillation is not stationary, and comparison of “pure” and adjacent cancelled ECG counterparts included both the influence of imperfect cancellation and non-stationarity of atrial fibrillation. For the synthesized signal, although without any influence of nonstationarity, it is clear that random variability within 10% of the QRS and T-waves amplitudes and beat-to-beat intervals does not reproduce the real physiological case of atrial fibrillation.
Conclusions Frequency domain features for both “pure” atrial fibrillation signals (gold standard) and adjacent remainder ECGs obtained by the cancellation method were studied, and the results from cancellation were quite comparable to those from the gold standard. Additional evidence from studies of synthesized signals in both the time and frequency domains strengthened our confidence in the cancellation method. Thus, the remainder ECG created by the cancellation method can be used as a reliable representation to characterize atrial activity. In addition, peak frequency and most power for all the patients for both “pure” and cancelled ECG segments did fall in a range of 4-9 Hz, which confirmed previous studies of atrial fibrillation (5– 7).
Acknowledgment The authors thank Drs W. Fisher, M. Kim, and K. Krishnan from Evanston Northwestern Healthcare
1. Fuster V, Ryden LE (co-chairs), et al: ACC/AHA/ESC guidelines for the management of patients with atrial fibrillation. J Am Coll Cardiol 38:1231, 2001 2. Brand FN, Abott RD, Kannel WB: Characteristics and prognosis of lone atrial fibrillation: 30-years follow up in the Framingham study. JAMA 254:3449, 1985 3. Slocum J, Byrom E, McCarthy L, et al: Computer detection of atrioventricular dissociation from the surface electrocardiogram during wide QRS tachycardias. Circulation 72:1028, 1985 4. Slocum J, Sahakian A, Swiryn S: Diagnosis of atrial fibrillation from surface electrocardiograms based on computer-detected atrial activity. J Electrocardiol 25:1, 1992 5. Shkurovich S, Sahakian A, Swiryn S: Detection of atrial activity from high-voltage leads of implantable ventricular defibrillators using a cancellation technique. IEEE Trans Biomed Eng 45:229, 1998 6. Ropella KM, Sahakian AV, Baerman JM, et al: The coherence spectrum: A quantitative discriminator of fibrillatory and non-fibrillatory cardiac rhythms. Circulation 80:112, 1989 7. Botteron GW, Smith JM: Quantitative assessment of the spatial organization of atrial fibrillation in the intact human heart. Circulation 93:513, 1996 8. Koning KTS, Smeets JLRM, Penn OC, et al: Configuration of unipolar atrial electrograms during electrically induced atrial fibrillation in humans. Circulation 95:1231, 1997 9. Ropella KM, Sahakian AV, Baerman JM, et al: Effects of Procainamide on intra-arterial electrograms during atrial fibrillation-implication for detection algorithms. Circulation 77:1047, 1988 10. Bollmann A, Kanuru NK, McTeague KK, et al: Frequency analysis of human atrial fibrillation using the surface electrocardiogram and its response to ibutilide. Am J Cardiol 81:1439, 1998 11. Sih HJ: Coherence mapping of cardiac tissue and its application to atrial fibrillation. Doctoral thesis, Evanston, IL, Northwestern University, 1994 12. Stridh M, Sornmo L: Spatiotemporal QRST cancellation techniques for analysis of atrial fibrillation. IEEE Trans Biomed Eng 48:105, 2001 13. Pan J, Tompkins WJ: A real-time QRS detection algorithm. IEEE Trans Biomed Eng 32:228, 1985 14. Kirsh JA, Sahakian AV, Baerman JM, et al: Ventricular response to atrial fibrillation: Role of atrioventricular conduction pathways. JACC 12:1265, 1988