JACC: CLINICAL ELECTROPHYSIOLOGY
VOL. 3, NO. 7, 2017
ª 2017 BY THE AMERICAN COLLEGE OF CARDIOLOGY FOUNDATION
ISSN 2405-500X/$36.00
PUBLISHED BY ELSEVIER
http://dx.doi.org/10.1016/j.jacep.2017.04.008
EDITORIAL COMMENT
Computer-Assisted Mapping in Electrophysiology Are the Machines Taking Over?* Frank Bogun, MD, Mohammed Saeed, MD, PHD
T
he 12-lead electrocardiogram can be helpful
In this issue of JACC: Clinical Electrophysiology,
in localizing the exit site in scar-related
Sapp et al. (6) report on the utility of using the
ventricular
Pioneering
12-lead electrocardiogram to determine the exit site
work by Waxman and Josephson (1) and later Miller
tachycardia
(VT).
in scar-related VT by using a novel computerized
et al. (2) introduced the concept of dividing the
algorithm. Conventional 12-lead electrocardiographic
endocardium into several regions (typically 10 to
data (from 38 patients), along with 120-point body-
12) and studied whether particular 12-lead electro-
surface potential mapping data (in 18 patients of
cardiographic patterns during VT could be reliably
the 38-patient cohort) acquired during endocardial
mapped to specific regions. Several algorithms have
pacing were pooled. This resulted in surface elec-
been proposed using mapping schemes to guide
trocardiographic data from a set of 1,012 known
electrophysiologists during catheter ablation proced-
pacing sites, which were then labeled as originating
ures (1–3). In particular, in cases in which activation
from 1 of 16 left ventricular (LV) endocardial areas.
mapping may be limited by hemodynamic instability
SEE PAGE 687
during VT, the use of pace mapping to identify VT exit sites has been demonstrated to be of critical
Two complementary localization techniques are
importance (4). Yokokawa et al. (5) proposed a
presented. The first technique involves the use of a
semiautomated algorithm based on a supervised
template matching algorithm to map surface elec-
machine-learning
ma-
trocardiographic data to 1 of the 16 LV segments and
chines) trained on a set of patients and then
further to 1 of 238 triangles with a regression algo-
technique
(support-vector
validated on a different set of patients with approx-
rithm that specifies the x, y, and z coordinates of
imately 70% accuracy in localizing the VT sites of
the candidate site in 3-dimensional space on the
origin to 1 of 10 LV endocardial regions. Such
basis of a 3-dimensional generic model of the LV
advances
learning,
endocardium. Localization error was reported to be
along with the availability of automatically acquired
a mean of 12 mm using a bootstrapping method. In
pace-mapping data, open exciting possibilities for
another 20 patients, pace mapping was performed
developing real-time “intelligent” decision support
from a “limited” number of sites within a region of
systems to guide VT ablation procedures and poten-
interest
tially
approach)
in
shorten
computing
procedure
and
machine
times
and
improve
outcomes.
(guided to
by
the
train a
segment
localization
patient-specific
regression
model that allowed (on the basis of an “optimal” set of 3 surface electrocardiographic leads) improved localization of pace-mapping sites in the LV endocardium.
*Editorials published in JACC: Clinical Electrophysiology reflect the views of the authors and do not necessarily represent the views of JACC: Clinical Electrophysiology or the American College of Cardiology. From the Division of Cardiovascular Medicine, University of Michigan, Ann Arbor, Michigan. Both authors have reported that they have no relationships relevant to the contents of this paper to disclose.
The
reported
localization
error
using
patient-specific mapping coefficients was reported to be as low as 5 mm in select patients but varied from 2.6 to 20.7 mm. The segment localization approach appears to be similar to the work of Yokokawa et al. (5), who first
Bogun and Saeed
JACC: CLINICAL ELECTROPHYSIOLOGY VOL. 3, NO. 7, 2017 JULY 2017:700–2
Pace-Mapping and Computer Algorithm
demonstrated the feasibility of using a computa-
continuous 3-dimensional points rather than only
tional algorithm for localizing VT exit sites using
discrete segments of the LV endocardium, as others
pace mapping–generated QRS configurations in post-
have done. The “inverse problem” of electrocardiog-
infarct patients as training data. Unfortunately, Sapp
raphy has been pursued by many others, and typically
et al. (6) do not provide the individual error analysis
required
for the segment localization approach and instead
mapping electrocardiographic data, in addition to
provide only summary statistics (mean error locali-
information regarding heart-torso geometry and
zation of 12 mm) and report 75.9% accuracy over all
cardiac morphology from advanced cardiac morpho-
patients when the correct segment ranks among the
metric imaging analysis (computed tomography and/
top 3 of the candidate segments (Online Table 1 in
or magnetic resonance imaging) (7). It is surprising
Sapp et al. [6]). The size of this area of interest is not
that a localization error as low as 5 mm using only
indicated but is likely in the 30- to 50-cm 2 range (5).
3 electrocardiographic leads can be achieved in the
Unfortunately, the accuracy of the top template
absence of such 3-dimensional geometric data. The
match is not given. Furthermore, the investigators
investigators did not describe the body habitus,
do not describe the distribution of the pace-mapping
cardiac volumes, scar location and orientation of their
sites and their relationships to the distribution of
patient cohort. It would be helpful to understand the
known scar, as this may significantly affect the
reliability and robustness of computer algorithms in
configuration of the acquired 12-lead electrocardio-
relation to individual patient characteristics and
grams and hence the accuracy of their template
scar distribution patterns. The work presented by
matching algorithm.
Sapp et al. (6) can be seen as a preliminary proof-of-
Sapp et al. (6) report that the “personalized” regression coefficients localize endocardial points
concept
high-density
study
that
body-surface
certainly
warrants
potential
further
investigation.
from surface electrocardiographic data with surpris-
However, several caveats need to be considered:
ing accuracy in some patients. No clear-cut validation
First, it is important to emphasize that the localiza-
data are provided for all patients. Figure 3, for
tion error was based on predicting the site of pacing
example, demonstrates that from generic patient
rather than true VT exit sites. Validation data detail-
pooled data (from 38 patients), a template match with
ing the localization accuracy of target sites for given
a correlation coefficient (CC) of 0.99 could be gener-
VTs is reported for only 2 patients. The accuracy of
ated for an individual patient. It is unclear how one is
the method used is reported as the localization error
to interpret such a high template CC. In practice, it is
between the estimated 3-dimensional coordinate and
quite difficult to obtain a CC of 0.99 in a given patient
the actual point where pacing was done. There were
when pace mapping is performed to identify VT exit
20 patients used for computing the continuous
sites (and pace-mapping configurations are compared
coordinates, in which each patient had at least 10
with VT configurations in the same patient). There is
pacing sites within a confined neighborhood. This
a wide variance in the localization accuracy. For
clustering effect of pace mapping sites (1,012 points
example, Patient #25 had a localization mean error of
distributed over 238 segments) invariably leads to
20.7 mm. This is much higher than the reported
sparse sampling in some segments (some triangles
localization
may have 0 to 2 corresponding pace maps), with
error
using
the
population-based
template matching approach.
resulting under- and/or oversampled regions of the
The investigators describe a work flow for how this
generic LV endocardial model. Furthermore, scar-
method can be used for the practicing electrophysi-
related arrhythmias may likely lead to additional
ologist: VT is induced, and the computerized algo-
variability in spatial resolution of pace-mapping sites.
rithm will rank the best 3 LV endocardial segments
Second, it would be of more interest to evaluate
matching best with the induced VT configuration.
how this method performed for localizing VT exit
This is displayed as a CC. Then the operator proceeds
sites in all patients rather than calculating localiza-
to place the catheter in the proximity of these seg-
tion predictions of known endocardial mapping
ments and performs pace mapping at as many sites as
points. Not all VTs have exit sites in the endocardium.
possible (minimum of 5 sites), and local regression is
It is not clear how this algorithm performs for
performed in real time to provide the candidate exit
epicardial or intramural VTs. VTs originating from the
site. Eventually the operator will look at the actual
septum with intramural or right ventricular exit sites
matches of pace mapping rather than predictions of
may be incorrectly classified, because right ventricu-
an algorithm that may or may not be correct.
lar data were not included in the training data.
The work of Sapp et al. (6) is ambitious in that they attempted
to
localize
VT
exit
sites
to
actual
In summary, the computer algorithm described by Sapp et al. (6) helps regionalize VT exit sites in
701
702
Bogun and Saeed
JACC: CLINICAL ELECTROPHYSIOLOGY VOL. 3, NO. 7, 2017 JULY 2017:700–2
Pace-Mapping and Computer Algorithm
patients with structural heart disease by using training data from prior patients. Accuracy can be
ADDRESS FOR CORRESPONDENCE: Dr. Frank Bogun,
improved if more pace mapping is performed in the
Division of Cardiovascular Medicine, University of
region of interest. The value of this method for VT
Michigan, Cardiovascular Center, SPC 5853, 1500 East
ablation procedures will need to be determined
Medical Center Drive, Ann Arbor, Michigan 48109-
prospectively.
5853. E-mail:
[email protected].
REFERENCES 1. Waxman HL, Josephson ME. Ventricular activation during ventricular endocardial pacing: I. Electrocardiographic patterns related to the site of
human, infarct-related ventricular tachycardia. J Cardiovasc Electrophysiol 2007;18:161–8.
6. Sapp JL, Bar-Tal M, Howes AJ, et al. Real-time localization of ventricular tachycardia origin from the 12-lead electrocardiogram. J Am Coll Cardiol
pacing. Am J Cardiol 1982;50:1–10.
4. Marchlinski FE, Callans DJ, Gottlieb CD, Zado E. Linear ablation lesions for control of unmappable ventricular tachycardia in patients with ischemic and nonischemic cardiomyopathy. Circulation
EP 2017;3:687–99.
2. Miller JM, Marchlinski FE, Buxton AE, Josephson ME. Relationship between the 12-lead electrocardiogram during ventricular tachycardia and endocardial site of origin in patients with coronary artery disease. Circulation 1988;77:759–66.
2000;101:1288–96.
3. Segal OR, Chow AW, Wong T, et al. A novel
5. Yokokawa M, Liu TY, Yoshida K, et al. Automated analysis of the 12-lead electrocardiogram
algorithm for determining endocardial VT exit site from 12-lead surface ECG characteristics in
to identify the exit site of postinfarction ventricular tachycardia. Heart Rhythm 2012;9:330–4.
7. Ramanathan C, Ghanem RN, Jia P, Ryu K, Rudy Y. Noninvasive electrocardiographic imaging for cardiac electrophysiology and arrhythmia. Nat Med 2004;10:422–8.
KEY WORDS computerized algorithm, exit site, machine learning, ventricular tachycardia