Improved frequency resolution for characterization of complex fractionated atrial electrograms
© Ciaccio et al.; licensee BioMed Central Ltd. 2012
Received: 13 January 2012
Accepted: 13 March 2012
Published: 3 April 2012
The dominant frequency of the Fourier power spectrum is useful to analyze complex fractionated atrial electrograms (CFAE), but spectral resolution is limited and uniform from DC to the Nyquist frequency. Herein the spectral resolution of a recently described and relatively new spectral estimation technique is compared to the Fourier radix-2 implementation.
In 10 paroxysmal and 10 persistent atrial fibrillation patients, 216 CFAE were acquired from the pulmonary vein ostia and left atrial free wall (977 Hz sampling rate, 8192 sample points, 8.4 s duration). With these parameter values, in the physiologic range of 3–10 Hz, two frequency components can theoretically be resolved at 0.24 Hz using Fourier analysis and at 0.10 Hz on average using the new technique. For testing, two closely-spaced periodic components were synthesized from two different CFAE recordings, and combined with two other CFAE recordings magnified 2×, that served as interference signals. The ability to resolve synthesized frequency components in the range 3–4 Hz, 4–5 Hz, …, 9–10 Hz was determined for 15 trials each (105 total).
With the added interference, frequency resolution averaged 0.29 ± 0.22 Hz for Fourier versus 0.16 ± 0.10 Hz for the new method (p < 0.001). The misalignment error of spectral peaks versus actual values was ±0.023 Hz for Fourier and ±0.009 Hz for the new method (p < 0.001). One or both synthesized peaks were lost in the noise floor 13/105 times using Fourier versus 4/105 times using the new method.
Within the physiologically relevant frequency range for characterization of CFAE, the new method has approximately twice the spectral resolution of Fourier analysis, there is less error in estimating frequencies, and peaks appear more readily above the noise floor. Theoretically, when interference is not present, to resolve frequency components separated by 0.10 Hz using Fourier analysis would require an 18.2 s sequence duration, versus 8.4 s with the new method.
KeywordsAtrial fibrillation Ensemble averaging Fourier analysis Spectral estimation Spectral resolution
Accurate spectral resolution is important for analysis of atrial fibrillation (AF) in the frequency domain . The information can be used, for example, to monitor periods of AF versus sinus rhythm in patients with paroxysmal arrhythmia. For understanding the mechanism by which the arrhythmia is maintained, it is also desirable to quantify any temporal variation in frequency that can occur during AF . Since frequency shifts can be minute, sufficient resolving power is essential to detect these differences over short time intervals. Yet this becomes more difficult when signal-to-noise ratio is low, as is often the case. Another impediment to the generation of an accurate spectral profile is the presence of transients in the signal that are unrelated to physiologically relevant phenomena. One way to improve continuity in the frequency domain representation of AF from one instance of time to the next is to model and update the spectral profile based on new signal information as it is acquired in atrial fibrillation signals [2, 3]. By using Gaussian functions to model the spectral peaks, after leveling the spectral baseline, spurious or transient changes to the signal frequency content can be excluded from the updates. When AF organization is reduced, the harmonic pattern is also expected to diminish , while independent spectral peaks caused for example by wavebreak can increase . When resolving power is insufficient, a broad spectral peak in the Fourier power spectrum may thus be indicative of either a temporal variation in the frequency of a single component, or the merging of multi-component independent sources . Therefore, the need for sufficient spectral resolution is essential for AF analysis. Improved resolution could also be useful to study the evolution and spontaneous termination of paroxysmal AF, as well as to estimate the organization of atrial activity .
In recent work we introduced a paradigm for spectral estimation and signal transformation of complex fractionated atrial electrograms (CFAE) based upon ensemble averaging [7–9]. This relatively new technique, like Fourier analysis, depends upon the autocorrelation function for generation of a power spectrum . For spectral estimation, Fourier analysis models the sinusoidal properties of the autocorrelation function, while the new technique averages the autocorrelation function at lags w, 2w, 3w, …, where w is the period . The spectral resolution of Fourier analysis is constant across bandwidth, while the spectral resolution of the ensemble average method depends inversely on w. The 1/w relationship results in greater spectral resolution at lower frequencies, which includes the electrophysiologic range of interest of ~ 3–10 Hz that is used for AF characterization. Herein, the new technique is compared to Fourier analysis for discerning closely-spaced frequency components in CFAE, which are of interest to characterize the arrhythmia and to identify dominant frequencies.
Clinical data acquisition
Atrial electrograms were recorded in a series of 20 patients referred to the Columbia University Medical Center cardiac electrophysiology (EP) laboratory for catheter ablation of AF. These recordings were obtained prospectively as approved by the Internal Review Board at Columbia University Medical Center, but were analyzed retrospectively after the catheter ablation procedures were completed using standard clinical protocols. Ten patients had documented clinical paroxysmal AF, and all 10 had normal sinus rhythm as their baseline rhythm in the EP laboratory. AF was induced by burst atrial pacing from the coronary sinus or right atrial lateral wall, and persisted for at least 10 minutes for those signals included in the retrospective analysis of this study. Ten other patients had longstanding persistent AF, and had been in AF without interruption for 1–3 years prior to the catheter mapping and ablation procedure. The surface electro gram signals were acquired in analog form using the GE CardioLab system (GE Healthcare, Waukesha, WI) and filtered from 30-500 Hz with a single-pole band pass filter to remove baseline drift and high frequency noise. The filtered signals were digitally sampled by the system at 0.977 KHz and stored. Although the band pass high end was slightly above the Nyquist frequency, negligible signal energy is expected to reside in this frequency range .
Only signals identified as CFAEs by two cardiac electro physiologists were included in the retrospective analysis. Candidate CFAE recordings of at least 10 seconds in duration were obtained from two sites outside the ostia of each of the four pulmonary veins (PV). Similar recordings were obtained at two sites on the endocardial surface of the left atrial free wall, one in the mid-posterior wall, and another on the anterior ridge at the base of the left atrial appendage. From each of these recordings, 8.4-second sequences (8192 sample points) were analyzed. A total of 240 such sequences were acquired during electrophysiologic analysis – 120 from paroxysmal and 120 from longstanding AF patients. Subsequently, only 216 of the recordings were determined to be CFAE, and only these were used for subsequent analysis. As in previous studies, all CFAE signals were normalized to mean zero and unity variance prior to further processing .
This frequency resolution is uniform throughout the range of DC to 977/2 = 489 Hz. The total number of data points in the resulting power spectrum will be 8192/2 = 4096. Thus the frequency resolution above the electrophysiologic range of interest, 10 Hz, to the high limit of 489 Hz, is the same as the frequency resolution within the physiologic range of 3–10 Hz.
which has units of millivolts. The power spectrum can be displayed by plotting √n·PwRMS versus frequency f. The √n term levels the spectral baseline, which would otherwise decrease by 1/√n, the noise falloff per number of summations n used for ensemble averaging. Computer code to calculate this spectrum has been presented and described elsewhere .
where RN is the resolution of the spectral estimator, SR is the sample rate, and fw and fw+1 are any two adjacent points on its frequency spectrum. The frequency resolution will be higher (i.e., the difference fw − fw+1 will be smaller) at points along the spectrum where period w is longer (i.e. at low frequencies).
Resolution to discern closely-spaced frequency components
9.37 ± 0.47
0.21 ± 0.05
0.25 ± 0.08
P = 0.307
8.49 ± 0.24
0.23 ± 0.09
0.28 ± 0.08
P = 0.105
7.51 ± 0.32
0.18 ± 0.07
0.29 ± 0.18
P = 0.081
6.47 ± 0.27
0.14 ± 0.04
0.32 ± 0.20
P = 0.003
5.45 ± 0.32
0.13 ± 0.09
0.27 ± 0.18
P < 0.001
4.37 ± 0.27
0.08 ± 0.03
0.22 ± 0.07
P < 0.001
3.51 ± 0.29
0.07 ± 0.04
0.19 ± 0.06
P < 0.001
6.41 ± 2.01 (MN)
0.16 ± 0.10 (MN)
0.29 ± 0.22 (MN)
P < 0.001
Synthesis of simulated CFAE and power spectra
signal component 1 with repeating period w
signal component 2 with repeating period w + 2
interference 1 with 2× gain
interference 2 with 2× gain
This paradigm was repeated so that a total of 15 simulated CFAE with closely-spaced frequency components in the range 3–4 Hz, 15 in the range 4–5 Hz, …, and 15 in the range 9–10 Hz were constructed. Thus the total number of simulation signals to be analyzed, of the type shown in Figure 1D and 2D, was 7 × 15 = 105.
Prior to Fourier analysis, each simulated signal was zero-padded to a length N = 65536, for a final interval of 0.015 Hz between spectral points. By zero padding, the frequency resolution of the FFT was not increased, only the point-to-point interval in the resulting spectrum, because the length of the observed signal was unchanged. Zero padding ensured that closely-spaced spectral components would be evident within the limit given by the frequency resolution constraint. The Fourier power spectrum was generated using a radix-2 implementation . Initially the simulated CFAE were preprocessed with a Han window prior to Fourier analysis. However, while Han windowing can improve periodicity by forcing the signal ends to zero, it also adds distortion in the form of amplitude modulation. The amplitude modulation can impart spectral sidebands which reduce frequency resolution. It was also observed visually that discernment of closely-spaced frequency components was improved without this filter. Therefore a rectangular window (i.e. no filter) was used for Fourier analysis. Likewise, no window preprocessing was done prior to spectral estimation with the new method.
The spectral resolution of Fourier analysis versus the new technique in the presence of interference were determined as follows. The criteria as given by Equations 11, 12 and 13 were tested for two synthesized spectral components with periods ω and ω + γ, γ = 2 through 20 spectral points. Again, the lower limit of 2 was necessitated by the fact that two frequency components occurring at adjacent spectral points (γ = 1) cannot be resolved. The minimum value of γ for which these criteria were met was defined as the spectral resolution to discern the closely-spaced simulated spectral peaks. If these criteria were not met to γ = 20, the trial was tabulated as being one in which the resolution was indeterminable due to the loss of one or more spectral peaks in the noise floor. For both methods, the estimate error was defined as the average of the absolute difference in fw - fω and fw+α – fω+γ, where fw and fω are the measured and actual frequencies of one of the peaks, and fw+α and fω+γ are the measured and actual frequencies of the other peak, respectively. For Fourier analysis, a correction was made by subtracting 0.06 Hz from each error after calculating the absolute difference of measured to actual frequency, with the minimum error after subtraction being 0.00 Hz. This allowed for the fact that for simplicity, the synthetic frequencies that were used were always 1/w, with w being an integer, so that on average for Fourier analysis, having constant increments of 0.12 Hz between spectral points, the actual frequency components could be different by as much as ±0.06 Hz. For the new technique, the spectral points occurred at 1/w for all w in range, so no correction was done.
The overall paradigm was repeated for 15 trials in each frequency range 3–4 Hz, 4–5 Hz, …, 9–10 Hz for both Fourier analysis and the new technique. The results were averaged and tabulated as mean ± standard deviation. The unpaired t-test was used to determine the significance of the difference between methods (SigmaPlot 2004 for Windows Ver. 9.01, Systat Software, Chicago, and MedCalc ver. 9.5, 2008, MedCalc Software bvba, Mariakerke, Belgium).
For a total of 105 trials, excluding those in which the synthesized components could not be resolved, the mean frequency resolution with the presence of additive interference was 0.29 ± 0.22 Hz for Fourier and 0.16 ± 0.10 Hz for the new technique (p < 0.001). Thus the resolving power of each method, in the presence of interference, was similar to their respective theoretical resolutions of 0.10 Hz and 0.24 Hz. A complete summary of the resolving power of both techniques is provided in Table 1. The mean frequency for the random trials in each frequency bin that resolved are shown in the rows of column 1 from top to bottom (9–10 Hz, 8–9 Hz, …). The expected resolution at 9.5 Hz, 8.5 Hz, …, using the new method, are noted in the second column. The actual resolving power for the new technique and for the Fourier transform are shown in the next two columns, respectively. At all frequencies, the new technique resolving power is improved over the Fourier method. The significance of the difference in mean resolving power between the methods is noted in the right-hand column. At lower frequencies up to 7 Hz there is a highly significant difference (p ≤ 0.003). At higher frequencies the difference trends toward significance.
Estimate error in the frequency of peak values
Error NT (Hz)
Error FT (Hz)
9.37 ± 0.47
0.032 ± 0.074
0.039 ± 0.070
P = 0.708
8.49 ± 0.24
0.004 ± 0.003
0.016 ± 0.015
P < 0.001
7.51 ± 0.32
0.004 ± 0.002
0.027 ± 0.043
P = 0.006
6.47 ± 0.27
0.005 ± 0.007
0.017 ± 0.023
P = 0.011
5.45 ± 0.32
0.003 ± 0.002
0.018 ± 0.023
P = 0.001
4.37 ± 0.27
0.003 ± 0.002
0.019 ± 0.019
P < 0.001
3.51 ± 0.29
0.013 ± 0.035
0.043 ± 0.053
P = 0.016
6.41 ± 2.01 (MN)
0.009 ± 0.030 (MN)
0.023 ± 0.039 (MN)
P < 0.001
Overall, of 15 trials for each frequency bin (105 trials in all), one or both frequency components did not appear above the noise floor in 13/105 trials using the Fourier method (12.4%) versus 4/105 trials using the new technique (3.8%).
In this study a comparison was made between the ability to resolve two closely-spaced frequency components in the physiologic range of interest using Fourier power spectral analysis, versus a new technique that utilizes signal averaging. The synthesized closely-spaced frequency components and two additive interferences were selected at random from a set of 216 CFAE. The values for digital sampling rate (977 Hz) and sequence length (N = 8192, 8.4 s sequences) are typical of those used for frequency analysis of CFAE obtained during clinical EP study. Tests were made in the range 3–10 Hz, the electrophysiologic range for evaluation of atrial electrical activity. From 105 tests, the mean resolving power of Fourier versus the new technique (0.29 Hz versus 0.16 Hz; p < 0.001), were higher than the theoretical values but in accord with the presence of large interferences that could act to mask the frequency components. In 13/105 trials, interference masked frequency components in the Fourier power spectrum. By comparison, this occurred in only 4/105 trials using the new technique. The error in estimating the synthesized components was ±0.023 Hz using Fourier versus ±0.009 Hz using the new technique (p < 0.001).
Signal segment not used for spectral estimation
The dynamics of internal fibrillatory activity are of great interest to establish AF mechanisms as well as for catheter ablation, and can be measured using markers of organization . It would be desirable to correlate these invasive measurements with noninvasive electrocardiogram measurements, to determine the extent to which AF can be characterized noninvasively with for example, a Holter monitor . It is also desirable to distinguish slightly different levels of organization during AF by improving temporal resolution and sensitivity . Such improvements could lead to new ways of analyzing and understanding AF, and improved AF treatment methods . The parameters used in our study for analysis of CFAE are typical for clinical investigations. The capability of the Fourier transform to resolve two independent frequency components in close proximity, 0.24 Hz, is unacceptable for describing the temporal evolution of paroxysmal AF or to differentiate this type from persistent AF. Using Fourier analysis, the dominant frequency, defined as the tallest spectral peak in the physiologic range of interest, often varies with a standard deviation of 0.2–0.3 Hz when CFAE are recorded from each of the four pulmonary vein ostia . Thus the resolving power using the Fourier spectrum is likely to be insufficient for detailed, accurate measurement. To determine whether differences in the frequencies of spectral peaks are real or artifact requires an increase in resolving power, as is possible with our new technique. Furthermore, the ability to resolve multiple characteristic frequencies in AF signals, which may be caused by multiple wavelets propagation as are commonly found in AF data, is more difficult when the electrical activity is highly unorganized . Applying the new technique to these signals can potentially increase the ability to spectrally resolve the multiple activation wavelets that can be present in the arrhythmogenic substrate.
In this study we demonstrated that the resolving power to discern two closely-spaced synthesized frequency components was significantly greater using a new spectral estimation technique as compared with Fourier analysis. Furthermore, the new technique estimates the actual frequencies of the components to a significantly greater degree of accuracy as compared with Fourier. The simulations suggest that two closely-spaced frequency components, as might be encountered by drivers of electrical activation in close spatial proximity to one another within the arrhythmogenic substrate that is causing atrial fibrillation, can be discerned by the new technique, even in combination with substantial interference levels. Such interferences may represent, for example, independent drivers at the periphery of the substrate that are inconsequential to the maintenance of AF. Thus this method may be helpful to detect and identify independent drivers in the atrial substrate that can be masked by interference from bystander regions. Such information is potentially important to develop a mechanistic understanding of the pattern of electrical activation during atrial fibrillation. Although only CFAE simulations were done in this study, these observations are probably applicable to other types of signals including ventricular electrograms  and electrocardiogram data . Furthermore, better spectral resolution can be important for improved recognition of repeating patterns in CFAE when analyzed in the frequency domain .
For simplicity, the number of trials for discerning closely-spaced frequency components was limited to 15 for each of seven frequency bins in the electrophysiologic range of interest. The types of interference signals and the types of periodic components that were used were limited to those present in a pool of 216 CFAE recordings obtained from the pulmonary vein ostia and left atrial free wall in both paroxysmal and longstanding AF patients. The use of synthesized components was necessitated by the fact that frequency values needed to be known with certainty. Since synthesized signals were tested, they will not necessarily represent the performance of Fourier spectral analysis versus the new technique on real data. Still, since the synthesized frequency components and interferences were extracted from real CFAE, we believe that they are representative of the types of signals that would be encountered in real measured data.
EJC developed the mathematical methods, conducted the data analysis, and wrote the manuscript. ABB, WW, and HG made helpful suggestions, provided the clinical data, and determined which recordings were complex fractionated atrial electrograms. All authors have read and approved the final manuscript.
We would like to thank Dr. Niels F. Otani, Mohammed M. Premjee, and Dr. Laura M. Muñoz of the Department of Biomedical Sciences at Cornell University, and Dr. Elisa Konofagou, Dr. Jean Provost, and Alexandre Costet of the Department of Biomedical Engineering at Columbia University, for very helpful discussions. No outside funding was received for this study.
- Kostka PS, Tkacz EJ: Feature extraction based on time-frequency and Independent Component Analysis for improvement of separation ability in Atrial Fibrillation detector. Conf Proc IEEE Eng Med Biol Soc 2008, 2008: 2960–2963.Google Scholar
- Corino VD, Mainardi LT, Stridh M, Sörnmo L: Improved time-frequency analysis of atrial fibrillation signals using spectral modeling. IEEE Trans Biomed Eng 2008, 55: 2723–2730.View ArticleGoogle Scholar
- Stridh M, Sornmo L, Meurling C, Olsson SB: Sequential characterization of atrial tachyarrhythmias based on ECG time-frequency analysis. IEEE Trans Biomed Eng 2004, 51: 100–114. 10.1109/TBME.2003.820331View ArticleGoogle Scholar
- Zlochiver S, Yamazaki M, Kalifa J, Berenfeld O: Rotor meandering contributes to irregularity in electrograms during atrial fibrillation. Hear Rhythm 2008, 5: 846–854. 10.1016/j.hrthm.2008.03.010View ArticleGoogle Scholar
- Huang Z, Chen Y, Pan M: Time-frequency characterization of atrial fibrillation from surface ECG based on Hilbert-Huang transform. J Med Eng Technol. 2007, 31: 381–389. 10.1080/03091900601165314View ArticleGoogle Scholar
- Vayá C, Rieta JJ: Time and frequency series combination for non-invasive regularity analysis of atrial fibrillation. Med Biol Eng Comput 2009, 47: 687–696. 10.1007/s11517-009-0495-3View ArticleGoogle Scholar
- Ciaccio EJ, Biviano AB, Whang W, Wit AL, Garan H, Coromilas J: New methods for estimating local electrical activation rate during atrial fibrillation. Hear Rhythm 2009, 6: 21–32. 10.1016/j.hrthm.2008.10.016View ArticleGoogle Scholar
- Ciaccio EJ, Biviano AB, Whang W, Wit AL, Coromilas J, Garan H: Optimized measurement of activation rate at left atrial sites with complex fractionated electrograms during atrial fibrillation. J Cardiovasc Electrophysiol 2010, 21: 133–143. 10.1111/j.1540-8167.2009.01595.xView ArticleGoogle Scholar
- Ciaccio EJ, Biviano AB, Whang W, Coromilas J, Garan H: A new transform for the analysis of complex fractionated atrial electrograms. BioMed Eng OnLine 2011, 10: 35. 10.1186/1475-925X-10-35View ArticleGoogle Scholar
- Biviano AB, Coromilas J, Ciaccio EJ, Whang W, Hickey K, Garan H: Frequency domain and time complex analyses manifest low correlation and temporal variability when calculating activation rates in atrial fibrillation patients. Pacing Clin Electrophysiol 2011, 34: 540–548. 10.1111/j.1540-8159.2010.02993.xView ArticleGoogle Scholar
- Numerical Recipes in Fortran Cambridge, UK: Cambridge University Press 1992.Google Scholar
- Alcaraz R, Hornero F, Rieta JJ: Assessment of non-invasive time and frequency atrial fibrillation organization markers with unipolar atrial electrograms. Physiol Meas 2011, 32: 99–114. 10.1088/0967-3334/32/1/007View ArticleGoogle Scholar
- Sih HJ, Zipes DP, Berbari EJ, Olgin JE: A high-temporal resolution algorithm for quantifying organization during atrial fibrillation. IEEE Trans Biomed Eng 1999, 46: 440–450. 10.1109/10.752941View ArticleGoogle Scholar
- Ng J, Kadish AH, Goldberger JJ: Technical considerations for dominant frequency analysis. J Cardiovasc Electrophysiol 2007, 18: 757–764. 10.1111/j.1540-8167.2007.00810.xView ArticleGoogle Scholar
- Nguyen MP, Schilling C, Dossel O: A new approach for frequency analysis of complex fractionated atrial electrograms. Conf Proc IEEE Eng Med Biol Soc 2009, 2009: 368–371.Google Scholar
- Ciaccio EJ, Tennyson CA, Bhagat G, Lewis SK, Green PH: Robust spectral analysis of videocapsule images acquired from celiac disease patients. BioMed Eng OnLine 2011, 10: 78. 10.1186/1475-925X-10-78View ArticleGoogle Scholar
- Ciaccio EJ, Drzewiecki GM: Tonometric arterial pulse sensor with noise cancellation. IEEE Trans Biomed Eng 2008, 55: 2388–2396.View ArticleGoogle Scholar
- Ciaccio EJ, Biviano AB, Whang W, Garan H: Identification of recurring patterns in fractionated atrial electrograms using new transform coefficients. BioMed Eng OnLine 2012, 11: 4. 10.1186/1475-925X-11-4View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.