CLINICAL RESEARCH
Europace (2011) 13, 383–388
doi:10.1093/europace/euq417
Ablation for Atrial Fibrillation
An improved algorithm for intraoperative
registration of computed tomographic
left atrial images
Hua Zhong and David Schwartzman *
Cardiovascular Institute, University of Pittsburgh, UPMC Presbyterian, B535, Pittsburgh, PA 15213-2582, USA
Received 23 August 2010; accepted after revision 13 October 2010; online publish-ahead-of-print 17 November 2010
Aims
Although anatomically accurate and highly detailed, preoperative computed tomographic (CT) images of the left
atrium (LA) are of limited utility for guiding catheter navigation, in part because of changes in LA shape which
occur between preoperative and intraoperative settings. Such changes may produce errors in spatial juxtaposition,
or ‘registration’, of the CT image to the intraoperative environment. The goal of this study was to assess a new algorithm for CT image registration.
.....................................................................................................................................................................................
Methods
In each of 10 patients, CT images were registered using LA endocardial points derived from intraoperative intracardiac echocardiography (CartoSound, Biosense). Two registration algorithms were compared: (i) CartoMerge
and results
(Biosense), a ‘rigid’ algorithm in which the CT image was not malleable; (ii) ImageMorph, a non-rigid algorithm in
which CT image was malleable and was altered so as to more precisely fit the intraoperative point locations.
There were no significant differences in LA volume or pulmonary vein antral dimensions on CT images after registration using CartoMerge vs. ImageMorph. Shape changes induced by ImageMorph were not concentrated in any one
LA region. Registration quality was significantly better using ImageMorph, as was mock circumferential ablation
accuracy.
.....................................................................................................................................................................................
Conclusion
The potential for improved CT image-guided ablation accuracy using ImageMorph calls for further study to discern
whether this algorithm yields tangible procedural and clinical benefits relative to currently available algorithms.
----------------------------------------------------------------------------------------------------------------------------------------------------------Keywords
Atrium † Catheter ablation † Computed tomography † Intracardiac echocardiography
Introduction
Preoperative imaging techniques, in particular computed tomography (CT), provide accurate and highly detailed multidimensional
images of the left atrial (LA) endocardium.1 Thus far, no intraoperative imaging technology has been able to replicate such
detail. Experience has demonstrated that for a topographically
complex substrate such as the LA, image detail influences catheter
ablation safety and efficacy.
Various algorithms have been utilized in an attempt to harness
the information contained in CT images for guidance of catheter
ablation by attempting to spatially juxtapose, or ‘register’, them
into the operative field. The most widely used of these is CartoMerge (Biosense), which attempts to register the CT image using
intraoperative spatial points obtained using electroanatomical
mapping (Carto, Biosense). After registration, the CT image is
used to guide catheter navigation.2 Although clinical experience
with CartoMerge has been mixed, randomized studies do not
suggest that the use of this technology favourably impacts clinical
outcomes.3 – 5 One reason for this disappointing result may be
inaccuracies in the registration process, which detailed analyses
have shown to be substantial.6,7
It has become increasingly clear that the LA is malleable and, at
any point in its mechanical cycle, undergoes changes in shape
between preoperative and intraoperative settings. CartoMerge
uses a rigid registration algorithm, which does not take such
changes into account. We hypothesized that an algorithm which
altered the shape of the CT image so as to better approximate
* Corresponding author. Tel: +1 412 647 2762; fax: +1 412 647 7979, Email: [email protected]
Published on behalf of the European Society of Cardiology. All rights reserved. & The Author 2010. For permissions please email: [email protected].
384
H. Zhong and D. Schwartzman
the intraoperative setting without sacrificing data content would
yield a higher quality registration and more accurate ablation. In
order to achieve this, a large number of intraoperative LA spatial
points were needed, far greater than the ,100 points typically
obtained during electroanatomical mapping procedures. For this
purpose, we utilized CartoSound (Biosense), an intraoperative
intracardiac echocardiography (ICE)-based imaging technology
which resolves LA contours into large numbers of spatial points.8
Methods
This study was approved by the Institutional Review Board of the University of Pittsburgh Medical Center.
Patients
This report is based on a non-consecutive cohort of 10 patients who
underwent LA endocardial catheter ablation intended to cure a syndrome of paroxysmal atrial fibrillation. The mean age was 59 years
(range 52 – 68 years); eight patients were men. No patients had significant valvular heart disease, prior cardiac surgery, prior catheter ablation, or diminished left ventricular ejection fraction. All patients were
in sinus rhythm during both CT image acquisition and the entirety of
the mapping/ablation procedure.
Computed tomographic model of the left
atrium
Each patient underwent CT imaging within 24 h of the ablation procedure. The details of the imaging technique have been reported previously.1 In brief, a commercial 64-detector scanner (VCT, General
Electric Healthcare) was used to acquire electrocardiographically
gated images during a single breath-hold at functional residual lung
capacity. Image data were processed using commercial software
(Advantage version 4.2, General Electric Healthcare) by experienced
radiologists who confirmed that for each patient, the image quality
was excellent. The multidimensional image resulting from compilation
of two-dimensional images, hereinafter termed the ‘CT model’, was a
reconstruction of the LA endocardial surface (contrast border) at enddiastasis. For consistency, each pulmonary vein sleeve was terminated
at 2 cm proximal to its respective carina. The LA surface was resolved
into a ‘mesh’ comprising confluent triangles using a marching cube
algorithm (Figure 1).9 The vertices of these triangles form the ‘points’
for registration of CT data with CartoSound data.
CartoSound model of the left atrium
The imaging platform was Carto XP (Biosense, version 9.6.34). The
spatial reference was placed on the skin overlying the left scapula.
The timing reference was the moment of an atrial electrogram
recorded at the coronary sinus ostium, the timing of which approximated LA end-diastasis, and with the lungs at functional residual
capacity. Catheter access to the heart was gained via femoral veins
and to the LA via atrial septal puncture.
In each patient, ICE imaging was performed using a phased-array
transducer catheter incorporating a position sensor (SoundStar, Biosense), which recorded individual 908 planar sector images as well as
their spatial location and orientation, to the Carto workspace.8
Images were acquired using an intra-LA transducer location as the
catheter was manipulated using combinations of deflection and
torque until, in the opinion of the operator, no new endocardial
regions could be visualized. Each image was optimized by adjusting frequency (5 – 10 MHz) and contrast, and the LA endocardial surface
Figure 1 Approximate posteroanterior view of LA. (Left)
Schematic demonstration of the CT model of LA (top) resolved
into triangular mesh (middle), with inset (bottom) a magnification
of the indicated region of posterior wall. The vertices of these triangles formed the ‘points’ referred to in the text. (Right) Demonstration of the CartoSound model of LA under construction. Red
lines are individual LA endocardial contours (pulmonary vein
contours have been removed for clarity). An icon shows the
location of the ICE transducer within the LA cavity, and a
single echocardiographic sector image is shown to demonstrate
typical image quality from which contours were generated.
contours were traced by hand using the CartoSound software. The
software then resolved each contour line into a series of discrete
spatial points, with an inter-point spacing of ≤3 mm (closer spacing
in regions of curvature or angulation of the line; details are Biosense
proprietary). These points comprised the ‘CartoSound model’
(Figure 1). In each patient, the comprehensiveness of the model was
assayed by estimating the proportion of the entire LA endocardial
surface which was within a 10 mm radius of any SoundStar-derived
point.8
Registration of the computed tomographic
model to the CartoSound model using
CartoMerge
As noted above, CartoMerge utilizes a rigid algorithm, in which neither
the CT nor the CartoSound model is altered from their original
shapes. From among the individual SoundStar images, ‘landmark’
points, which permitted a preliminary orientation of the CT and
CartoSound point data sets, were chosen from observed echocardiographically discrete anatomical sites (e.g. pulmonary vein carinae,
lateral ridge, mitral annulus, and fossa ovalis).7,8 All of the other
SoundStar-derived points were then added, and CartoMerge was
then allowed to finalize the registration process.
Registration of the computed tomographic
model to the CartoSound model using
ImageMorph
A separate registration was performed using a new technology which
we have developed, hereinafter termed ‘ImageMorph’.10 Unlike CartoMerge, ImageMorph did not use a rigid algorithm. Rather, the
385
Algorithm for intraoperative registration of CT LA images
Figure 2 Schematic representation of the ImageMorph
process. (Left) Computed tomographic surface, resolved into triangular mesh, portrayed as a flat pink square. Cartosound points
(green circles) locations are shown after rigid registration with
the surface. (Center) Adjustment of individual CT vertices to
overlap the nearest CartoSound points. The CT surface
appears jagged because the number of CT points far exceeds
the number of CartoSound points. (Right) Computed tomographic surface after application of the morphing process, in
which the spatial locations of the CT points for which there
were no CartoSound points have been estimated using a diffusion
smoothing algorithm.
SoundStar-derived points were used to spatially adjust the locations of
the CT vertices so as to more adequately reflect the intraoperative
setting, without sacrifice to CT data content. After rigid registration
of the CT and CartoSound models, the CT vertex (Figure 1) nearest
to each SoundStar-derived point was moved so as to be superimposed
upon that point. Given the paucity of SoundStar-derived points relative
to CT vertices (below), a diffusion smoothing algorithm was then
applied, which estimated the locations of the remaining CT vertices.11
This process is demonstrated schematically in Figure 2. Supplementary
material online, Video files demonstrate the examples of the ‘morphing’
process in time-lapse style.
Although, as noted above, prior reports have demonstrated that
changes in LA shape occur between preoperative and intraoperative
settings, these reports have not addressed the elements of this
phenomenon including, importantly, whether and how it might vary
by atrial region. We examined this issue by dividing the LA into five
roughly equal regions (Figure 3). In each patient, the average position
change of all CT vertices comprising a region (in millimetres) incurred
during the morphing process, hereinafter termed ‘morphing magnitude’, was derived for each of the five regions as well as the LA as a
whole.
Assessment of registration quality
We directly compared the registration using CartoSound with that
using ImageMorph by employing a ‘distance to nearest endocardial
surface’ (DNES) assay7: the distal electrode of a mapping/ablation
catheter (NaviStar Irrigation, Biosense) was guided into stable but nondistending endocardial contact under direct visualization with a
mechanical array ICE catheter (UltraICE, Boston Scientific) at sites
throughout the LA (range of 32 – 57 sites per patient). At each site,
a Carto point was acquired, and these points were considered ‘true’
Figure 3 Left atrial ‘polar map,’ derived from CT (endocardial
vantage from mitral annulus, which is opened widely so as to view
the entire LA surface) to illustrate some of the measurements
performed. (A) Individual colours demarcate regions of LA for
which separate ‘morphing magnitude’ values were measured
[anterior: region 1 (red); posterior: ¼region 2 (grey); inferior:
region 3 (green); septal: region 4 (orange); lateral: region 5
(blue)]. (B) Antral maximum (Dmax) and minimum (Dmin) diameters. RPVA, right pulmonary vein antrum; LPVA, left pulmonary
vein antrum; LAAA, left atrial appendage antrum.
endocardial points. After each registration, the distance to the
nearest CT surface from each true endocardial point was measured.
DNES was the average of those distances and is a measure of registration quality independent of either CartoSound or ImageMorph.8
Assessment of circumferential ablation
accuracy
We directly compared the accuracy of CT model-guided circumferential antral ablation after registration using CartoMerge with that after
registration using ImageMorph by employing a ‘circumferential ablation
inaccuracy’ (CAI) assay.7,8 In each patient, two mock circumferential
ablations were performed on each antrum: (i) guided solely by the
CT model after registration using CartoMerge and (ii) guided solely
by the CT model after registration using ImageMorph. The order of
study was random. Mock circumferential ablation of each antrum
was performed during which the NaviStar catheter operator, blinded
to all but the CT model and unaware whether the model was based
on CartoMerge registration or ImageMorph registration, placed the
ablation electrode at 8 – 10 approximately equidistant sites along
what he perceived to be each planned lesion path; a Carto point
was recorded at each site. As we have previously described, these
paths were, where possible, 1 cm distal (towards the centre of the
LA cavity) to any venoatrial junction.7,8,12 – 14
After the mock ablations, actual circumferential ablation along the
planned lesion path was performed as a series of contiguous focal
lesions guided by UltraICE without the use of the CartoSound or CT
data.12 – 14 The minimum distance of each of the mock ablation points
386
H. Zhong and D. Schwartzman
to the nearest point on the actual lesion path was measured. Circumferential ablation inaccuracy, calculated separately for each antrum, was the
average of those distances and is a measure of circumferential ablation
accuracy independent of both CartoMerge and ImageMorph.8
Analytical methods
Data are presented as mean + standard deviation, unless otherwise
stated. A comparison of continuous variables utilized an analysis of variance, a paired two-tailed Student t-test, or a Mann– Whitney rank-sum
test, as appropriate. Correlation analysis was performed using the
Pearson product moment method. For all statistical tests, a P-value
of ,0.05 was considered significant.
Results
Information content of the models
The CartoSound model comprised an average of 5292 + 1294
(range 3349– 6758) points, encompassing an average of 95 + 3%
(range 91 –100%) of the LA endocardial surface. Despite the extensive number of points comprising the CartoSound model, the CT
models were much more data dense, each resolved into an
average of 53 555 + 13 962 (range 39 459– 60 691; P , 0.001) vertices. This difference in density was associated with marked differences in image quality, highlighting the importance of the CT image
(Figure 4). The LA volumes prior to (144 + 34 cc, range 97–
199 cc) and after (134 + 23 cc, range 96 –168 cc; P ¼ 0.42) morphing were not significantly different, nor were the maximal (38 + 7 cc
prior to morphing and 36 + 8 cc after morphing; P ¼ 0.57) and
minimal (22 + 5 cc prior to morphing and 22 + 4 cc after morphing; P ¼ 0.96) left antral diameters, or the maximal (48 + 4 cc prior
to morphing and 47 + 7 cc after morphing; P ¼ 0.69) and minimal
(28 + 5 cc prior to morphing and 29 + 4 cc after morphing; P ¼
0.74) right antral diameters (Figure 3).
Assessment of registration quality
Registration quality was significantly better using ImageMorph
(DNES 1.5 + 0.3 mm, range 1.0 –1.9 mm) than using CartoMerge
(DNES 2.9 + 0.3 mm, range 2.5 – 3.7 mm; P , 0.001). Individual
patient data are shown in Table 1.
Assessment of circumferential ablation
accuracy
Figure 4 Left atrial endocardial surfaces from similar vantages
looking towards the lateral wall derived using only CartoSound
data (left, derived using CartoSound software) or only CT data
(right). Despite large numbers of points utilized to create each
image, the quality of the CT image far exceeds that of the CartoSound image. LPVA, left pulmonary vein antrum; LAAA, left atrial
appendage antrum.
In the left antrum, CAI was significantly higher after registration
using CartoMerge (6.6 + 2.4 mm, range 3.9 – 11.3 mm) than after
registration using ImageMorph (3.6 + 1.4 mm, range 1.9 –
6.1 mm; P ¼ 0.003). An example is shown in Figure 5. A similar
finding was made for the right antrum, with CAI significantly
higher using CartoMerge (7.4 + 2.3 mm, range 4.4 –11.8 mm)
than ImageMorph (4.0 + 0.9 mm, range 2.7 –5.7 mm; P , 0.001).
Individual data are shown in Table 1.
Regional assessment of morphing
magnitude
Among the entire cohort, there were no significant interregional
(Figure 3) differences in morphing magnitude, although there was
Table 1 Individual patient registration quality and mock ablation accuracy
DNES
.............................................
CAI (left antrum)
.............................................
CartoMerge
ImageMorph
CAI (right antrum)
.............................................
CartoMerge
ImageMorph
ImageMorph
CartoMerge
1
2.8
1.7
7.7
4.2
6.5
3.9
2
3
3.7
2.8
1.1
1.9
5.8
4.3
4.9
3.2
11.8
4.4
5.2
2.9
4
2.6
1.0
5.6
3.3
9.5
4.1
5
6
3.2
2.9
1.4
1.8
3.9
7.1
2.1
2.4
7.3
5.1
4.5
2.7
7
2.8
1.3
5.1
2.8
5.7
3.1
8
9
3.1
2.7
1.3
1.4
5.6
9.9
1.9
4.8
9.6
6.7
5.1
4.3
10
2.5
1.8
11.3
6.1
7.0
4.4
...............................................................................................................................................................................
PT
387
Algorithm for intraoperative registration of CT LA images
significant inter-individual variability (Table 2). We could discern no
significant correlations between global (LA as a whole) or regional
morphing magnitudes and the magnitudes of improvement in registration quality or left/right circumferential ablation accuracy
between CartoMerge and ImageMorph.
Discussion
In this report, we provide a preliminary validation of a new technique for improving LA registration quality and ablation accuracy.
Although CT-derived LA images have been shown to be anatomically accurate and highly detailed, their utility has been limited by
inaccuracies inherent in the registration process.6,7 These
Figure 5 Endocardial CT vantage towards lateral LA wall
demonstrating mock antral ablation points (black circles) and
actual ablation points (red circles). For the CartoMerge registration, the discrepancy between actual and mock lesions is
more apparent than it is for the ImageMorph registration.
inaccuracies do not permit the operator to trust the virtual
image. Rather, it becomes one of a number of imprecise and/or
inadequate data streams, including fluoroscopic, electrographic,
and ICE, which must be melded in his/her mind’s eye. We speculate that the melding process can become counterproductive,
because individual data sources are inaccurate for different
reasons. A completely trustworthy registration would render LA
ablation more akin to painting a picture, albeit on a moving
canvas, under ‘direct vision’ (CT image of endocardium). Clearly,
we have not accomplished this using ImageMorph: inaccuracy
remains and the images are still. In addition to further improvement in accuracy, future development might introduce motion,
which is achievable with the current CT technologies.15,16
We are not aware of prior reports which have compared different registration algorithms. Our data suggest significant improvements in registration quality and antral circumferential ablation
accuracy for ImageMorph relative to CartoMerge. We are also
not aware of prior attempts to examine mechanisms of shape
changes occurring between preoperative and intraoperative settings; herein, we assayed for regional propensities. Unfortunately,
their apparent absence leaves us without new insight as to why
such changes occur. The lack of a significant correlation between
regional morphing magnitude and the degree of improvement in
ablation accuracy would suggest a more complex process
beyond the capability of our analytical technique.
These data have a number of important limitations. First, we
selected a small number of patients in sinus rhythm and with
superior image quality, elements commonly not observed in realworld practice. Secondly, in order to be effective, the ImageMorph technique requires a great deal of intraoperative data
encompassing the bulk of the LA endocardial surface. We
achieved this by a combination of an atypical transducer location
(LA) and long duration of ICE image acquisition (mean 18 min)
and processing (mean 43 min) times. Each of these items may
render the technique laborious, if not impractical, for routine
clinical use. We made no attempt to establish its lower limits,
such as those which would be imposed by a less reliable
Table 2 Individual patient morphing magnitude by LA region (Figure 3)
Region
1
2
3
4
5
PT
1
2.0
2.2
2.5
2.6
2.4
2
1.5
1.4
1.8
2.4
1.3
3
4
3.1
3.5
1.6
1.5
1.8
1.8
2.1
1.7
1.7
1.6
5
4.6
2.1
2.5
3.8
3.9
6
7
2.5
4.2
3.4
3.2
1.8
2.6
2.2
3.3
2.2
1.2
8
2.7
2.9
3.8
1.7
4.6
9
10
1.3
1.6
1.8
2.6
4.2
1.3
1.2
1.3
3.3
1.8
2.7 + 1.1
2.3 + 0.7
2.4 + 0.9
2.2 + 0.8
2.4 + 1.2
...............................................................................................................................................................................
Mean + SD
388
transducer location (e.g. right heart)8 and/or fewer LA endocardial CartoSound points acquired. Thirdly, in that the comparison
between ImageMorph and CartoMerge was performed during
mock ablation; we have not demonstrated that the magnitude
of the CAI differences observed between these techniques
would have impacted significantly on the electrophysiological or
clinical endpoints sought; prospective trials would be required
to address these questions. Similarly, our observations on ablation accuracy are limited to the pulmonary vein antra; although
the regional morphing magnitude data would not suggest it, it
is conceivable that differences between CartoMerge and ImageMorph would have been less marked in other regions of the
LA. Fourthly, although the ImageMorph algorithm should be
applicable to any multidimensional image data set, herein we
made no attempt to demonstrate its utility with magnetic resonance images. Finally, we did not compare ImageMorph with the
other commercially available registration technique, Ensite
Fusion, which is used with the NavX intraoperative imaging platform (St Jude Medical).17 However, in addition to the spatial distortions inherent in impedance-based anatomical reconstruction
relative to magnetic-based reconstruction, the Fusion algorithm
seeks to morph the intraoperative image to the preoperative
image, which is the opposite of what we sought to achieve
with ImageMorph.
In summary, herein we have demonstrated the promise of a new
technique for improving preoperative LA image registration quality
and ablation accuracy. These findings provide the impetus for
further study to discern whether these improvements translate
into tangible procedural or clinical benefits.
Supplementary material
Supplementary material is available at Europace online.
Conflict of interest: D.S. receives research grants from
Medtronic.
H. Zhong and D. Schwartzman
References
1. Schwartzman D, Lacomis J, Wigginton W. Characterization of left atrium and
distal pulmonary vein morphology using multidimensional computed tomography.
J Am Coll Cardiol 2003;41:1349 –57.
2. Reddy V, Malchano Z, Holmvang G, Schmidt EJ, d’Avila A, Houghtaling C et al. Integration of cardiac magnetic resonance imaging with three-dimensional electroanatomic mapping to guide left ventricular catheter manipulation: feasibility in a porcine
model of healed myocardial infarction. J Am Coll Cardiol 2004;44:2202 –13.
3. Kistler PM, Rajappan K, Harris S, Early MJ, Richmond L, Sporton SC et al. The
impact of image integration on catheter ablation of atrial fibrillation using electroanatomic mapping: a prospective randomized study. Eur Heart J 2008;29:3029 –36.
4. Tang K, Ma J, Zhang S, Zhang JY, Wei YD, Chen YQ et al. A randomized prospective
comparison of CartoMerge and Carto XP to guide circumferential pulmonary vein isolation for the treatment of paroxysmal atrial fibrillation. Chin Med J 2008;121:508–12.
5. Caponi D, Corleto A, Scaglione M, Blandino A, Biasco L, Cristoforetti Y et al.
Ablation of atrial fibrillation: does the addition of three-dimensional magnetic resonance imaging of the left atrium to electroanatomic mapping improve the clinical
outcome? Europace 2010;12:1098 – 104.
6. Dong J, Dickfled T, Dalal D, Cheema A, Vasamreddy CR, Henrikson CA et al.
Initial experience in the use of integrated electroanatomic mapping with threedimensional MR/CT images to guide catheter ablation of atrial fibrillation.
J Cardiovasc Electrophysiol 2006;17:459 –66.
7. Zhong H, Lacomis J, Schwartzman D. On the accuracy of CartoMerge for guiding
posterior left atrial ablation in man. Heart Rhythm 2007;4:595 –602.
8. Schwartzman D, Zhong H. On the use of CartoSound for left atrial navigation.
J Cardiovasc Electrophysiol 2010;21:656 –64.
9. Lorensen WE, Cline HE. Marching cubes: a high resolution 3D surface construction algorithm. Comput Graph 1987;21:163 –9.
10. Zhong H, Kanade T, Schwartzman D. ‘Virtual Touch’: an efficient registration
method for catheter navigation in the left atrium. Med Image Comput Comput
Assist Interv 2006;9:437 – 44.
11. Desbrun M, Meyer M, Schröder P, Barr AH. Implicit fairing of irregular meshes
using diffusion and curvature flow. Comput Graph 1999;33:317 – 24.
12. Schwartzman D, Nosbisch J, Housel D. Echocardiographically guided left atrial
ablation: characterization of a new technique. Heart Rhythm 2006;3:930 – 8.
13. Schwartzman D, Williams J. Electroanatomic properties of pulmonary vein antral
regions enclosed by encircling ablation lesions. Europace 2009;11:435 –44.
14. Chandhok S, Williams JL, Schwartzman D. Anatomical analysis of recurrent conduction after circumferential ablation. J Interv Card Electrophysiol 2010;27:41 –50.
15. Lacomis JM, Goitein O, Deible C, Moran PL, Mamone G, Madan S et al. Dynamic
multidimensional imaging of the human left atrial appendage. Europace 2007;9:
1134 –40.
16. Zhong H, Kanade T, Schwartzman D. Sensor-guided ablation procedure of left
atrial endocardium. Med Image Comput Comput Assist Interv 2005;8:1 –8.
17. Richmond L, Rajappan K, Voth E, Rangavajhala V, Earley MJ, Thomas G et al. Validation of computed tomography image integration into the Ensite NavX mapping
system to perform catheter ablation of atrial fibrillation. J Cardiovasc Electrophysiol
2008;19:821 –7.