Professional Documents
Culture Documents
AbstractThis study introduces the use of wavelet decomposition of unipolar fibrillation electrograms for the automatic detection of local activation times during complex atrial fibrillation (AF).
The purpose of this study was to evaluate this technique in patients
with structural heart disease and longstanding persistent AF. In 46
patients undergoing cardiac surgery, unipolar fibrillation electrograms were recorded from the right atrium, using a mapping array
of 244 electrodes. In 25 patients with normal sinus rhythm, AF was
induced by rapid pacing, whereas 21 patients were in persistent AF.
In patients with longstanding AF, the atrial electrograms showed
a high degree of fractionation. In each patient, 12 s of AF were
analyzed by wavelet transformation (15 scales). The finest scales
(17) were used to reconstruct a local fibrillation electrogram,
whereas with the coarse scales (915), a far-field signal was generated. With these local and far-field electrograms, the primary fibrillation potentials, due to wave propagation underneath the electrode, could be distinguished from double potentials and multiple
components generated by remote wavefronts. Wavelet transformation resulted in AF histograms with a closely Gaussian distribution
and the automatically generated activation maps showed a good
resemblance with fibrillation maps obtained by laborious manual editing. A special chaining algorithm was developed to detect
multiple components in fractionated electrograms. The degree of
fractionation showed a positive correlation with the complexity of
fibrillation, thus providing an objective quantification of the degree
of electrical dissociation of the atria. Wavelet transformation can
be a useful technique to detect the primary potentials and quantify the degree of fractionation of fibrillation electrograms. This
could enable real-time mapping of complex cases of human AF
and classification of the underlying electropathological substrate.
Index TermsArrhythmias, atrial fibrillation (AF), electrophysiology, mapping, wavelet transform.
I. INTRODUCTION
URRENTLY, atrial fibrillation (AF) affects 2.3 million
patients in the United States and nearly 6 million in Europe. It has been estimated that by the year 2050, approximately
5.6 million Americans will have AF [1]. AF is a disorder of
the elderly, with about 80% occurring in people older than 65
years. The prevalence is about twofold higher in men than in
women [2][5]. It is characterized by a very rapid and uncoordinated activity of the atria, and a completely irregular ventricular
response. During AF, multiple wavefronts wander through the
atria separated by anatomical and functional barriers. Lines of
block preferentially occur along the border of atrial bundles like
the crista terminalis and the pectinate muscles. In diseased atria
(dilatation, fibrosis), smaller bundles also become dissociated.
Clinically, AF often starts in a paroxysmal form to become more
persistent with time. Although AF is generally not considered
as a life-threatening arrhythmia, it is associated with a considerable increase in mortality [6] and has an important impact on
the quality of life [7].
To diagnose the stage of development of the underlying
pathological substrate of AF, characterization of the underlying spatiotemporal process would be useful. Measurement of
local activation times allows identification of areas with conduction delays and/or conduction block. Previous methods for
automatic identification of local activation times in extracellular
electrograms are mainly based on amplitude, slope, and interval
criteria [8], [9]. Alternatively, we recently developed a template
matching technique for the processing of fibrillation electrograms [10], [11]. As a first step, this technique was validated in
25 patients with normal sinus rhythm, in whom AF was induced
by rapid pacing [12], [13]. 244 Unipolar electrograms were
recorded simultaneously from the free wall of the right atrium
using a slightly concave mapping electrode (diameter 3.6 cm;
interelectrode distance 2.25 mm). A silver plate in the thoracic
cavity was used as an indifferent electrode. All electrograms
were bandpass filtered (0.5500 Hz), sampled with 1 kHz, and
digitized with 12-bit resolution. All electrograms were correlated with a library of 128 mathematically defined extracellular
potentials of different duration and morphology (templates).
This resulted in a correlogram in which positive peaks represented the moments that the fibrillation waves passed under
the electrode. During acutely induced AF, template matching
yielded a very high performance with an overall sensitivity and
positive predictive value of 96.6% 2.5% and 94.3% 5.4%,
respectively. [11] False-positive detection was mainly due to
far-field potentials of remote fibrillation waves and double potentials recorded at the boundaries between waves. Incidental
fractionation of fibrillation electrograms in these patients was
responsible for undersensing of fibrillatory activity. This automatic detection of local fibrillation potentials opened the possibility for real-time imaging in the spatial domain [10]. However,
it is well known that in patients with structural heart disease
and longstanding AF, an increasing number of the fibrillation
electrograms become fractionated [14][16]. Spach and Dolber
have shown that such fractionation is because of asynchronous
1389
Fig. 1. Consecutive processing steps for the automatic analysis of unipolar fibrillation electrograms. After QRS subtraction and baseline correction, an initial
set of activation marks was generated by oversensitive template matching (bars).
Then wavelet decomposition was applied, resulting in a scalogram consisting of
15 different scales. Yellow indicates the maxima and black the minima in each
scale. A near-field electrogram was constructed by exponential summation
of the finest seven scales. This near-field electrogram was used to confirm true
fibrillation potentials (red marks), remove false positive detections, and identify double potentials (short blue marks). Multiple components in fractionated
electrograms were then identified by wavelet analysis (short blue marks). As a
last step, fractionated potentials missed by template matching (false negatives)
were added.
1390
represent the physiological variation in morphology of fibrillation potentials in normal atria, due to tissue anisotropy, variation
in conduction velocity, and slight dissociation of neighboring
muscle bundles.
All fibrillation electrograms were matched with all templates
in the library in time steps of 1 ms. This resulted in a correlogram
of all successive maximal normalized correlation coefficients.
The exact moments of activation were determined by scanning
the fibrillation electrograms for the most negative slopes in a
time window of 10 ms around the positive correlation peaks.
TABLE I
CHARACTERISTICS OF THE 15 SCALES OF THE USED GAUSSIAN WAVELET
C. Wavelet Decomposition
Wavelet transformation was introduced 18 years ago by
Daubechies [21], and has been further developed by Mallat [19]
and Strang and Nguygen [18]. Nowadays, this technique is
widely used in various biomedical applications [22][26].
The wavelet we used for decomposition of fibrillation electrograms is defined by
d
2
(t),
where
(t) = et .
(1)
dt
This wavelet was stepwise expanded to 15 different time
scales to calculate the wavelet coefficients
+
t
1
f (t)
Wf (, s) =
dt or in short
s
|s|
(t) =
d
s (t)
(2)
dt
where Wf are the wavelet coefficients, f the unipolar electrogram, the time shift, and s one of the 15 scales (115).
The
power of the different scales was normalized by the factor
1/ |s|.
The total duration of the 15 expanded wavelets ranged
between 3 and 90 ms. Transformation with these wavelets produces positive peaks at the negative slopes in the fibrillation
electrograms.
A wavelet transform can equally be considered as a bandpass
filter by rewriting formula (2) into
Wf (, s) = f
d
(f s )(t)
(3)
dt
where Wf is the derivative of the electrogram f after smoothing
by a low-pass filter defined by the scaling function s . At the
finest scales, the wavelet transform is equivalent to a bandpass
filter between 250 and 500 Hz (3dB). The intermediate scale
(8) passes frequencies between 17.8 and 61.8 Hz, whereas the
coarsest scale (15) is only sensitive for low frequencies (between
8.3 and 28.8 Hz). The wavelet transform was calculated for
the full length of all recorded fibrillation electrograms. After
calculating the wavelet transform, coefficients were corrected
for filter delays equal to half the duration of the wavelet (see
Table I).
In Fig. 2, four fibrillation maps are shown together with the
fibrillation potentials recorded at the sites indicated by open
circles. In Fig. 2 (bottom), the corresponding scalograms of
the 15 wavelet transforms are plotted. The finest scales contain the higher frequencies of the electrograms and the coarse
Wf (, s) =
scales the low frequencies. In panel A, a single broad fibrillation wave propagated under the mapping electrode, producing a sharp biphasic electrogram, composed of both near- and
far-field components. Consequently, the signal generated high
amplitudes in all scales of the scalogram. At the fine scales,
the transforms consist of monophasic narrow positive peaks
1391
Fig. 3. Creation of near- and far-field electrograms. Near- and far-field electrograms are calculated by weighed summation (weight factor between 0.11
and 1.0) of the seven finest (17) and seven coarsest scales (915). The fibrillation electrogram contains a single potential (A), a short double potential (B),
and a fractionated potential (C). In the reconstructed near-field electrogram,
the rapid negative deflections of the fibrillation electrogram are enhanced. The
reconstructed far-field electrogram was typically triphasic in shape, the time
between the peak and troughs being determined by the conduction velocity of
the waves. Broad and rapidly propagating wavefronts produced large far-field
electrograms. The far-field components of the fractionated potential (C) were of
low amplitude, indicating that they were generated by smaller fibrillation waves
propagating at close distance from the recording electrode.
1392
Fig. 4. AFCL histograms of the right atrial wall in three patients with an
increasing degree of disorganization of AF. The black histograms on the left
were generated by template matching alone. On the right, the AFCL histograms
are shown after wavelet filtering. The red histograms represent the distribution
of the true fibrillation intervals, obtained after removal of the false-positive
detections resulting from template matching (gray subpopulation). The subpopulation of long AF cycles (black) was due to undersensing of fractionated
fibrillation potentials.
where rnear and rfar indicate the amplitude of the near- and farfield electrograms as defined in (4), rnear m in the minimally required near-field amplitude, and rnear m ax the value above which
fibrillation potentials are accepted, regardless of the amplitude
of their far-field electrogram. The parameter determines the
acceptance rate of low amplitudes in the near-field electrogram, depending on the corresponding amplitude in the far-field
electrogram.
The proportion of false intervals in the AF cycle length
(AFCL) histograms obtained by template matching was estimated as the difference with the best Gaussian fit of the histograms. The rational for Gaussianity of the AFCL distribution
is given by our observation that after accurate editing of the fibrillation maps, the AFCL histogram showed a close to normal
distribution. The Gaussian fit was made from the right side of the
AFCL histogram that did not contain false-positive values. The
difference between the left side of the AFCL histogram and this
Gaussian fit was used to estimate the number of false-positive
detections by template matching.
For each patient, the wavelet filtering parameters and
rnear m ax of (5) were determined as follows. Starting at a low
rnear m ax , the parameter was stepwise increased until the number of removed activations was larger than the estimated number
of false short intervals. As long as this criterion was not satisfied, the procedure was repeated at progressively higher values
of rnear m ax . In Fig. 4, the result of wavelet filtering is shown for
three patients. The black histograms on the left-hand side are the
result of template matching alone, whereas the red histograms
on the right-hand side show the distribution of AF cycles after
removal of false-positive intervals (gray) by wavelet filtering.
The number of activations removed by wavelet filtering could
be set as a percentage of the estimated amount of false-positive
1393
1394
Fig. 7. Four examples of fractionated unipolar fibrillation electrograms. Below each electrogram, the output of the automatic detection algorithm is plotted.
Red bars indicate primary fibrillation potentials. Double potentials and multiple components are indicated by short blue bars. Fractionated portions of the
electrograms are encircled. V: far-field potentials of the ventricles.
wavelet decomposition for the determination of primary fibrillation potentials is evaluated by the degree of Gaussianity of
the obtained AFCL histograms. Template matching alone gave
highly variable results. Especially in the more complex cases
of AF, this technique failed to produce a normal distribution
of fibrillation intervals (Gaussianity <70%). Wavelet decomposition improved the detection of primary fibrillation potentials
considerably. On average, wavelet filtering increased the Gaussianity of the obtained AFCL histograms from 66.5% 19.3%
to 89.1% 10.4% (p < 0.001).
We compared the sensitivity and specificity of template
matching alone, and the combination of template matching and
wavelet decomposition in the whole population of AF patients.
A dataset of carefully edited activation times was used as a
gold standard. As expected, in patients with acutely induced
AF in whom AF was still relatively organized, the added value of
wavelet analysis was limited, with a sensitivity of 96.8% 2.3%
and 90.8% 7.3%, respectively, and a specificity of 90.1%
4.3% and 97.8% 2.7%, respectively. In contrast, in patients
with longstanding persistent AF, the specificity of detection of
fibrillation potentials was largely enhanced. In half of these
patients, showing the highest complexity of AF, specificity improved from 54.6% 8.1% to 78.0% 5.2% (p < 0.001).
In the other half, exhibiting still relatively organized AF, the
specificity increased from 77.6% 5.1% to 85.3% 7.0%
(p < 0.001).
In Fig. 9 (bottom), the number of multiple components detected by wavelet transformation is plotted against the percentage of intra-atrial conduction block (interelectrode conduction
velocity of <19 cm/s). In patients with longstanding AF (closed
1395
1396
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
V. CONCLUSION
The automatic analysis of fibrillation electrograms provides
an objective evaluation of fibrillation electrograms and visualization of the complex spatiotemporal processes during AF.
Since the calculations involved in the various algorithms are
not very complex, both template matching and wavelet transformation can be easily implemented in field programmable gate
arrays. This will provide real-time movies of the fibrillatory conduction patterns during AF, and an unbiased quantification of
the nature and degree of fractionated fibrillation electrograms.
This may open the way for a systematic diagnosis of the electropathological substrate of persistent AF in humans.
[16]
[17]
[18]
[19]
REFERENCES
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
1397
framingham heart study, Circulation, vol. 98, no. 10, pp. 946952, Sep.
1998.
J. E. Ware, K. K. Snow, M. Kosinski, and B. Gandek, SF-36 Health Survey
Manual and Inter Pretation Guide. Boston, MA: New England Medical
Center, The Health Institute, 1993.
D. Scherr, D. Dalal, A. Cheema, A. Cheng, C. A. Henrikson, D. Spragg,
J. E. Marine, R. D. Berger, H. Calkins, and J. Dong, Automated detection
and characterization of complex fractionated atrial electrograms in human
left atrium during atrial fibrillation, Heart Rhythm, vol. 4, no. 8, pp. 1013
1020, Aug. 2007.
M. Porter, W. Spear, J. G. Akar, R. Helms, N. Brysiewicz, P. Santucci, and
D. J. Wilber, Prospective study of atrial fibrillation termination during
ablation guided by automated detection of fractionated electrograms,
J. Cardiovasc.. Electrophysiol., vol. 19, no. 6, pp. 613620, Jun. 2008.
R. P. Houben and M. A. Allessie, Processing of intracardiac electrograms
in atrial fibrillation. Diagnosis of electropathological substrate of AF,
IEEE Eng. Med. Biol. Mag., vol. 25, no. 6, pp. 4051, Nov. 2006.
R. P. Houben, N. M. de groot, F. W. Lindemans, and M. A. Allessie,
Automatic mapping of human atrial fibrillation by template matching,
Heart Rhythm., vol. 3, no. 10, pp. 12211228, Oct. 2006.
K. T. Konings, C. J. Kirchhof, J. R. Smeets, H. J. Wellens, O. C. Penn,
and M. A. Allessie, High-density mapping of electrically induced atrial
fibrillation in humans, Circulation, vol. 89, no. 4, pp. 16651680, Apr.
1994.
K. T. Konings, J. L. Smeets, O. C. Penn, H. J. Wellens, and M. A. Allessie,
Configuration of unipolar atrial electrograms during electrically induced
atrial fibrillation in humans, Circulation, vol. 95, no. 5, pp. 12311241,
Mar. 1997.
K. Nademanee, M. Schwab, J. Porath, and A. Abbo, How to perform
electrogram-guided atrial fibrillation ablation, Heart Rhythm., vol. 3,
no. 8, pp. 981984, Aug. 2006.
J. Lin, B. J. Scherlag, J. Zhou, Z. Lu, E. Patterson, W. M. Jackman,
R. Lazzara, and S. S. Po, Autonomic mechanism to explain complex
fractionated atrial electrograms (CFAE), J. Cardiovasc. Electrophysiol.,
vol. 18, no. 11, pp. 11971205, Nov. 2007.
H. Oral, A. Chugh, E. Good, A. Wimmer, S. Dey, N. Gadeela, S. Sankaran,
T. Crawford, J. F. Sarrazin, M. Kuhne, N. Chalfoun, D. Wells, M. Frederick, J. Fortino, S. Loucif-Moore, K. Jongnarangsin, F. Pelosi, Jr., F. Bogun,
and F. Morady, Radiofrequency catheter ablation of chronic atrial fibrillation guided by complex electrograms, Circulation, vol. 115, no. 20,
pp. 26062612, May 2007.
M. S. Spach and P. C. Dolber, Relating extracellular potentials and their
derivatives to anisotropic propagation at a microscopic level in human
cardiac muscle. Evidence for electrical uncoupling of side-to-side fiber
connections with increasing age, Circ. Res., vol. 58, no. 3, pp. 356371,
Mar. 1986.
G. Strang and T. Nguygen, Wavelets and Filterbanks. Cambridge, MA:
Wellesley-Cambridge, 1996, pp. 1490.
S. Mallat, A Wavelet Tour of Signal Processing, 2nd ed. San Diego, CA:
Academic, 1998, pp. 1577.
J. H. Ward, Hierarchical grouping to optimize an objective function,
J. Amer. Statist. Assoc., vol. 58, pp. 236244, 1963.
I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: SIAM, 1992.
J. G. Akar, T. H. Everett, L. C. Kok, J. R. Moorman, and D. E. Haines, Effect of electrical and structural remodeling on spatiotemporal organization
in acute and persistent atrial fibrillation, J. Cardiovasc. Electrophysiol.,
vol. 13, no. 10, pp. 10271034, Oct. 2002.
J. Liu and Q. Zhang, [Detection of ECG characteristic using wavelet
transform], Zhongguo Yi. Liao. Qi Xie. Za Zhi., vol. 23, no. 6, pp. 324
327, Nov. 1999.
H. Li, D. Hao, and Y. Duan, [An approach to achieve classification of
QRS waves with wavelet multiresolution analysis and fractal dimension],
Sheng Wu Yi. Xue. Gong. Cheng Xue. Za Zhi., vol. 17, no. 4, pp. 451453,
Dec. 2000.
C. D. Swerdlow, M. L. Brown, K. Lurie, J. Zhang, N. M. Wood, W. H.
Olson, and J. M. Gillberg, Discrimination of ventricular tachycardia
from supraventricular tachycardia by a downloaded wavelet-transform
morphology algorithm: A paradigm for development of implantable cardioverter defibrillator detection algorithms, J. Cardiovasc. Electrophysiol., vol. 13, no. 5, pp. 432441, May 2002.
J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna, A
wavelet-based ECG delineator: Evaluation on standard databases, IEEE
Trans. Biomed. Eng., vol. 51, no. 4, pp. 570581, Apr. 2004.
A. Van Oosterom, Solidifying the solid angle, J. Electrocardiol., vol. 35,
pp. 181192, 2002.
1398