Predicting Catheter Ablation Outcome in Persistent Atrial Fibrillation

via Multivariate Analyses of ECG Fibrillatory Wave Amplitude

A. R. Hidalgo-Mu˜noz

, V. Zarzoso

, M. Meo

, O. Meste

, D. G. Latcu

, I. Popescu

and N. Saoudi

Laboratoire I3S, UMR 7271, Universit´e Nice Sophia Antipolis, CNRS, Nice, France

Department of Anesthesiology, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, U.S.A.

Service de Cardiologie, Centre Hospitalier Princesse Grace, Monaco, Monaco

1 INTRODUCTION

Atrial ﬁbrillation (AF) is the most common sustained

arrhythmia encountered in clinical practice. Among

the possible strategies to face AF, catheter ablation

(CA) is a well established therapy with proven efﬁ-

cacy to maintain sinus rhythm during follow-up. The

highest rate of AF recurrence is reported to occur

during the ﬁrst 6 months following ablation (Calkins

et al., 2012). Hence, an accurate selection of patients

who can really beneﬁt from this intervention at long-

term is essential.

In this work, two multivariate approaches for se-

lecting the most relevant features linked to CA out-

come are proposed. The ﬁrst one is known as logistic

regression (LR) (Indrayan, 2012). The second one is

based on a machine learning technique, named sup-

port vector machine (SVM) (Burges, 1998). The goal

of both analyses is to distinguish the AF patients more

prone to a successful CA outcome from those who are

more sensitive to arrhythmia recurrence. More specif-

ically, patients who suffer from persistent AF are con-

sidered in this study.

There are many possibilities to tackle the analy-

sis of the ﬁbrillatory waves (f-waves) such as time or

frequency domains (Matsuo et al., 2009; Jones et al.,

2013). One of the most direct and easily interpretable

measurements obtained from the f-waves is the ﬁbril-

latory wave amplitude (Nault et al., 2009; Meo et al.,

2013), which is the focus of the present work.

2 METHODS

2.1 Database and ECG Acquisition

Sixty-two patients (52 male, age = 61.5±10.4 years)

having persistent and long-lasting persistent

AF (Calkins et al., 2012) were recruited at the

Cardiology Department, Princess Grace Hospital,

Monaco, after giving their informed consent. One

minute standard 12-lead ECG was acquired at a

sampling rate of 977 Hz at the beginning of the CA

procedure. ECG signals were ﬁltered by a 4th-order

zero-phase bandpass Chebyshev ﬁlter with cut-off

frequencies of 0.5 Hz and 30 Hz. For every lead of

each recorded ECG, the ﬁbrillatory wave amplitude

was computed as in (Meo et al., 2013). Also as in

that reference, this work focused exclusively on the 8

linearly independent leads I, II, V1-V6.

All patients underwent stepwise CA (Calkins

et al., 2012), including lasso-guided circumferential

pulmonary vein (PV) disconnection, fractionated po-

tentials, non-PV triggers, roof line and mitral isthmus

line right atrial ablation. Outcome was followed up

during at least 6 months after the CA procedure. Fi-

nally, 47 patients were free from AF (success) while

15 patients had documented AF recurrences (failure).

2.2 Statistical Analysis

Normal distribution was ﬁrst checked by the

Kolmogorov-Smirnov test. A U-Mann Whitney test

was performed to verify any statistically signiﬁcant

differences between the groups of interest. Both the

statistical analysis and the subsequent logistic regres-

sion model were performed by using the software

SPSS version 13.0.

LR consists of a probabilistic statistical model

aimed at predicting the outcome of a categorical de-

pendent variable (success and failure after CA) based

on one or more predictor variables (features). The

next equation is an example for M predictors:

LR = log



1− θ



= b

+ b

+ ... + b

(1)

where θ is the probability of belonging to the suc-

cess CA class, {b

, b

, ..., b

} are the coefﬁcients of

the regression model and {x

, ..., x

} correspond to

the numerical values of the different features. In

our case, x

corresponds to the ﬁbrillatory amplitude

R. Hidalgo-Muñoz A., Zarzoso V., Meo M., Meste O., G. Latcu D., Popescu I. and Saoudi N..

Predicting Catheter Ablation Outcome in Persistent Atrial Fibrillation via Multivariate Analyses of ECG Fibrillatory Wave Amplitude.

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

mean value for the mth lead selected by the model

from the complete set of 8 leads. A backward elimi-

nation method that discards the least signiﬁcant vari-

able at each iteration according to the Wald index was

employed (Indrayan, 2012) to keep only the most rel-

evant ECG leads.

On the other hand, SVM separates the given set of

binary labeled training data with a hyperplane that is

maximally distant (Euclidean distance) from the two

classes success and failure CA outcome). Linear dis-

criminant functions deﬁne a decision hyperplane in a

multidimensional feature space as:

g(x) = w

x+ w

= 0 (2)

where w is known as the weight vector and x is the

feature vector (set of lead amplitude values). Symbol

denotes the threshold and (·)

the transpose op-

erator. Note that elements in vector w close to zero

have little inﬂuence on g(x) and hardly contribute to

the decision. In (Guyon et al., 2002) it was suggested

a recursive feature elimination technique (SVM-RFE)

that discards iteratively the features corresponding to

the lowest values of |w

|. After feature elimination,

the classiﬁer is evaluated using a leave-one-out cross-

validation strategy. In our case, the initial feature vec-

tor x is made up of the mean ﬁbrillatory amplitude

values computed in each of the 8 leads considered

in this study. The SVM-RFE algorithm was imple-

mented using MATLAB version R2011a.

3 RESULTS

No signiﬁcant inter-class differences were found for

any of the single-lead amplitude parameters consid-

ered separately.

After applying LR by using the backward elimi-

nation method, the remaining leads were I, V1, V2

and V5. Table 1 summarizes the best results in terms

of AUC by applying univariate and multivariate ap-

proaches. The best cut-off point was θ = 0.7 yielding

83% sensitivity, 73.3% speciﬁcity and 80.6% total ac-

curacy.

The SVM-RFE algorithm ranked the leads as fol-

lows: I, V5, V2, V1, II, V4, V6 and V3 (ordered from

Table 1: Best accuracy results by the univariate contrast

with the highest AUC and the multivariate analysis after

LR backward elimination. AUC: area under the curve; IC:

AUC’s interval of conﬁdence; p: AUC’s statistical signiﬁ-

cance.

Analysis AUC IC p

Univariate (lead I) 0.678 0.50 to 0.85 0.039

Multivariate (θ) 0.854 0.75 to 0.96 <0.001

the most to the least relevant). The best result in terms

of accuracy is obtained by using only the three most

relevant leads (I, V5 and V2) reaching up to 82.3%

accuracy, 70% speciﬁcity and 84.6% sensitivity.

4 CONCLUSION

This work has shown that analyzing the ﬁbrilla-

tory amplitudes measured in multiple ECG leads im-

proves the accuracy of previous CA outcome pre-

dictors based on a single lead (Nault et al., 2009).

The multivariate amplitude analysis model reaches

an AUC of up to 0.854 with four leads. Both LR

and SVM-RFE agree in pointing to I, V1, V2 and

V5 as the most discriminant leads to determine mid-

and long-term CA outcome. Along the lines of (Meo

et al., 2013), these results support the suitability of

taking into consideration the information of several

leads that are often neglected in CA outcome predic-

tion and AF analysis (Nault et al., 2009; Matsuo et al.,

2009).

ACKNOWLEDGEMENTS

Work supported in part by the French National

Research Agency under contract ANR-2010-JCJC-

0303-01 “PERSIST” and the Centre Scientiﬁque de

Monaco. A. R. Hidalgo-Mu˜noz holds a postdoctoral

research fellowship awarded by Univ. Nice Sophia

Antipolis.

REFERENCES

Burges, C. (1998). A tutorial on support vector machines

for pattern recognition. Data Mining and Knowledge

Discovery, 2(2):121–167.

Calkins, H., Kuck, K. H., Cappato, R., Brugada, J., Camm,

A. J., et al. (2012). 2012 HRS/EHRA/ECAS expert

consensus statement on catheter and surgical ablation

of atrial ﬁbrillation: recommendations for patient se-

lection, procedural techniques, patient management

and follow-up, deﬁnitions, endpoints, and research

trial design. Europace, 14:528–606.

Guyon, I., Weston, J., Barnhill, S., and Vapnik, V. (2002).

Gene selection for cancer classiﬁcation using support

vector machines. Machine Learning, 46:389–422.

Indrayan, A. (2012). Medical Biostatistics. CRC Press.

Jones, A. R., Krummen, D. E., and Narayan, S. M. (2013).

Non-invasive identiﬁcation of stable rotors and focal

sources for human atrial ﬁbrillation: mechanistic clas-

siﬁcation of atrial ﬁbrillation from the electrocardio-

gram. Europace, 15(9):1249–1258.

Matsuo, S., Lellouche, N., Wright, M., Bevilacqua, M.,

Knecht, S., et al. (2009). Clinical predictors of ter-

mination and clinical outcome of catheter ablation for

persistent atrial ﬁbrillation. Journal of the American

College of Cardiology, 54(9):788–795.

Meo, M., Zarzoso, V., Meste, O., Lactu, D. G., and Saoudi,

N. (2013). Spatial variability of the 12-lead surface

ECG as a tool for noninvasive prediction of catheter

ablation outcome in persistent atrial ﬁbrillation. IEEE

Trans. on Biomedical Engineering, 60(1):20–27.

Nault, I., Lellouche, N., Matsuo, S., Knecht, S., Wright,

M., et al. (2009). Clinical value of ﬁbrillatory wave

amplitude on surface ECG in patients with persistent

atrial ﬁbrillation. Journal of Interventional Cardiac

Electrophysiology, 26(1):11–19.