Available online at www.sciencedirect.

com

Medical Engineering & Physics 30 (2008) 11051111

Predicting termination of atrial brillation based on the structure and quantication of the recurrence plot
Rongrong Sun, Yuanyuan Wang
Department of Electronic Engineering, Fudan University, 220 Handan Road, Shanghai 200433, China Received 9 August 2007; received in revised form 28 October 2007; accepted 26 January 2008

Abstract Predicting the spontaneous termination of the atrial brillation (AF) leads to not only better understanding of mechanisms of the arrhythmia but also the improved treatment of the sustained AF. A novel method is proposed to characterize the AF based on structure and the quantication of the recurrence plot (RP) to predict the termination of the AF. The RP of the electrocardiogram (ECG) signal is rstly obtained and eleven features are extracted to characterize its three basic patterns. Then the sequential forward search (SFS) algorithm and DaviesBouldin criterion are utilized to select the feature subset which can predict the AF termination effectively. Finally, the multilayer perceptron (MLP) neural network is applied to predict the AF termination. An AF database which includes one training set and two testing sets (A and B) of Holter ECG recordings is studied. Experiment results show that 97% of testing set A and 95% of testing set B are correctly classied. It demonstrates that this algorithm has the ability to predict the spontaneous termination of the AF effectively. 2008 IPEM. Published by Elsevier Ltd. All rights reserved.
Keywords: Atrial brillation; Termination; Prediction; Recurrence plot; Feature selection; Electrocardiogram

1. Introduction The atrial brillation (AF) is a commonly encountered cardiac disorder in clinical practice, which occurs in up to 10% of individuals older than 70 years of age [1]. The AF can be subdivided into three different forms: the paroxysmal AF, the persistent AF, and the permanent AF. Evidence suggests that the paroxysmal AF is a precursor to the development of the persistent AF which requires external electrical intervention to allow its termination [2]. The risks of the sustained AF are serious which include strokes and myocardial infarctions caused by the formation of blood clots within stagnant volumes in the atria [1]. Thus it is important to discriminate between the paroxysmal AF and the persistent AF and predict whether the paroxysmal AF is likely to terminate spontaneously or be sustained. By identifying the termination of the AF, not only better understanding of the mechanism of the arrhythmia but also more effective therapy can be achieved. The appropriate intervention may terminate the AF
Corresponding author. Tel.: +86 21 65642756; fax: +86 21 65643526. E-mail addresses: 041021082@fudan.edu.cn, yywang@fudan.edu.cn (Y. Wang).

by predicting the AF maintenance. Vice versa, the unnecessary therapeutic intervention may be avoided and risks for patients can be minimized if the AF is predicted that it can terminate spontaneously. Several algorithms for the characterization of the AF have been proposed including cross-correlation [3,4], frequency analysis [58]. Wave-morphology analysis [9], linear prediction [10], non-linear or statistical analysis [1114] and time-frequency analysis [1416] are also proposed. The focus of this paper is to predict the spontaneous termination of the AF based on the structure and quantication of the recurrence plot (RP) of ECG signals from a non-linear point of view. The original motivation for this work was the Computers in Cardiology Challenge in 2004 [17]. The RP was initially designed to graphically display distance correlations and non-stationarity in time series [18]. No mathematical assumptions regarding the data and the generating systems constrain the construction of the recurrence plot, thus this tool is particularly suitable for the analysis of physiological signals which are often non-stationary. In this study, eleven features are extracted to identify three basic patterns of the RP, namely, the pattern along the main (45 ) diagonal, the pattern along the 135 diagonal and

1350-4533/$ see front matter 2008 IPEM. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.medengphy.2008.01.008

1106

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

the square-like pattern. The sequential forward search (SFS) algorithm and DaviesBouldin criterion are then utilized for the feature selection to predict the termination of the AF effectively. Finally, a multilayer perceptron (MLP) neural network is applied to predict the termination of the AF. Experiment results demonstrate that the new algorithm described in the study can be implemented and provide an accurate prediction for the spontaneous termination of the AF.

2. Database The 2004 Computers in Cardiology/PhysioNet Challenge database [17] is an open competition with the goal of developing automated methods to predict the spontaneous termination of the AF. It consists of 80 two-channel ECG recordings. The database contains three AF types with different termination properties: non-terminating (labeled N, meaning that the AF episode continues at least 1 h after the end of the recordings), soon-terminating (labeled S, meaning that the AF episode terminates within 1 min after the end of the recordings) and immediately terminating (labeled T, meaning that the AF episode terminates within 1 s after the end of the recordings). Each recording has 1 min length. All signals are extracted from 20 to 24 h Holter recordings from patients with paroxysmal AF and digitized at 128 Hz, with 16 bits/sample and 5 V resolution. The database has been divided into a training data set and two testing data sets (A and B). The training data set contains 30 recordings with 10 recordings for each type. The testing data set A contains 30 recordings. It is known that half recordings are N and half are T. The testing data set B contains 20 recordings with 10 S recordings and 10 T recordings. For example, Fig. 1 shows three different types of AF signals in the training data set. In this paper, all data are split into 4 s segments for the analysis, so it consists of 15 segments for each 1 min record-

ing. For each segment, it is processed using a Butterworth lter to reduce baseline wander. The band stop frequencies are set at 0.25 and 40 Hz. Following this, the QRST complexes are canceled using the averaging technique [19] to obtain atrial waveforms. The Challenge dataset has already contained the QRST complex annotation. All the complexes within a xed length of window are aligned by their R points and averaged to construct a median complex. The window length is set as 100 ms with 30 ms preceded the R point and 70 ms followed the R point. The remainder atrial activity is obtained by subtracting the median complex from every ventricular complex. In order to demonstrate that the components of the atrial brillation are not cancelled, power spectrum of the remainder signals is analyzed with the discrete Fourier transform. If the frequency with the maximum power falls between 5 and 9 Hz, it is considered that the remainder signals include the atrial activity. Results show that 99.7% of the total signal segments in the database include the atrial activity after the QRST cancellation. It can be inferred that the averaging technique achieves a high performance for the QRST cancellation.

3. Methods 3.1. State space reconstruction For a time series si (i = 1, 2, . . ., N), it is embedded to a high-dimensional space with the time-delay embedding technique [20] and a trajectory matrix is obtained X= x2 s2 . = . . . . . xM sN(d1) x1 s1 s1+ s2+ . . . sN(d2) ... ... .. . s2+(d1) . . . sN (1) where d is the embedding dimension and is the delay time. Each row vector in the matrix xj = (sj ,sj+ ,. . .,sj+(d1) ), j = 1,2,. . ., M (M = N(d 1)) represents a d-dimensional trajectory point in the reconstructed state space. For each signal in the training data set, d and are calculated using the false nearest neighbor (FNN) method and autocorrelation method respectively [21]. The mean values of them are taken as the nal embedding dimension and delay time respectively. Thus signals are embedded into 3dimensional reconstructed state space with a lag of 70 sample points. 3.2. Recurrence plot After the state space reconstruction, the RP of a signal can be obtained. The RP is a two-dimensional representation technique. It is M M arrays whose element can be s1+(d1)

Fig. 1. A segment of non-terminating (N), soon-terminating (S) and immediately terminating (T) AF signals in the training data set.

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

1107

calculated using Ri,j = ( ||xi xj ||) i, j = 1, 2, . . . , M (2)

where |||| is the norm (here Euclidean norm is adopted) and () is the Heaviside function. This means if the distance between xi and xj is less than , then Ri,j = 1 and a dot is placed at (i, j) in the RP. The scale is a cutoff distance dening a sphere centered at xi . Here is calculated as 15% of the maximum distance value, computed as the 95th percentile [22]. For example, Fig. 2 shows the RP of three types of AF signals in Fig. 1. 3.3. Structures and quantication of recurrence plot One of major reasons that the RP is popular lies in the fact that its structures are visually appealing and fascinating [22]. The recurrence plot exhibits characteristic patterns which are caused by the typical dynamical behavior. As shown in Fig. 2, the RP of non-terminating AF is characterized by large black rectangles, whereas the RP of soon-terminating and immediately terminating-AF show only small rectangles. The RP of soon-terminating AF distributes more homogeneously than the RP of immediately terminating AF, the points in the RP of immediately terminating AF mainly concentrate along the main diagonal. The measures based on diagonal lines, vertical and horizontal structures make it possible to identify and quantify the transitions between periodic, laminar and chaotic states before AF termination, therefore facilitating the possible prediction of AF termination. Thus in this study, three patterns of RP are considered, namely, the pattern along the main diagonal, the pattern along the 135 diagonal and the square-like pattern. Eleven features are totally extracted from the RP to characterize these three distinct patterns. The Recurrence Rate (REC) is the percentage of recurrent points, and is calculated as follows, REC = 1 M2
M

Ri,j
i,j=1

(3)

It quanties the number of time instants characterized by a recurrence in the signals interaction: the more periodic the signal dynamics, the higher the REC value. 3.3.1. Pattern along the main diagonal This pattern corresponds to the situation that when xi and xj are close, xi+k and xj+k are also close together for a series of k values. In this study, Lmax , Lmean , Determinism (DET), Divarication (DIV), Entropy (ENTR), and Recurrence Trend (RT) are extracted to characterize this pattern. 1) Lmax measures the length of the longest diagonal segment. It reects the longest duration of a stable interaction.

Fig. 2. Recurrence plot of non-terminating (N), soon-terminating (S) and immediately terminating (T) AF signals in Fig. 1.

2) Lmean measures the average length of the diagonal segments whose lengths exceed the certain threshold lmin : Lmean =
M l=lmin lP(l) M l=lmin P(l)

(4)

1108

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

where P(l) is the number of diagonal structures whose length is l. Here lmin = 2 is used. It reects the average duration of a stable interaction. 3) Determinism (DET) measures the percentage of determinism, which is calculated as the percentage of recurrent points forming diagonal segments whose lengths exceed the certain threshold lmin , here lmin is also 2: DET =
M l=lmin lP(l) M i,j Ri,j

1) Vmax measures the length of the longest line segment which is vertical. 2) TRAP measures the average length of the line segment which is vertical and whose length exceeds a certain threshold Vmin : TRAP =
M v=Vmin vP(v) M v=Vmin P(v)

(8)

(5)

It contains information about the duration of a stable interaction: the longer the interaction, the higher the DET value. 4) Divarication (DIV) is the inverse of Lmax . 5) Entropy (ENTR) is the Shannon entropy of the distribution of the length of those diagonal segments:
M

where P(v) is the number of vertical structures whose length is v. Here Vmin = 2 is used. It reects the average length of sojourn points. 3) Vmaxho measures the length of the longest line segment which is horizontal and it reects the longest length of sojourn points. 3.4. Feature selection and DaviesBouldin criterion For each AF signal segment, a column vector of eleven features is calculated based on the structure and quantication of the RP. However, such a high-dimensional space is not suitable for the classication in term of the computational efciency. The feature selection is one category of dimensionality reduction methods and addresses the dimensionality reduction problem by determining a subset of features which are most essential for the classication. The sequential forward search (SFS) algorithm is the one that requires the least computation among the suboptimal feature search methods. It is adopted here for its simplicity and effectiveness proved in many applications [23]. For the SFS algorithm, it is very important to nd an evaluation criterion which measures the ability of the feature subset for the prediction of the AF termination. The DaviesBouldin (DB) criterion can address the issue of class separability in the feature space directly. Its most important advantage is that there is no need to suppose that the clusters have a normal distribution in the feature space, thus the DB criterion is adopted here [24]. It is calculated as the ratio of the common variability within classes to the data variability among the class: (Dii + Djj ) DBij = (9) Dij where Dii and Djj are the dispersions of the ith and jth clusters respectively, and Dij is the distance between their mean values. Here Dii and Dij are given by, 1 Dii = Li
Li 1/2

ENTR =
l=lmin

p(l) ln p(l)

(6)

where p(l) is the probability density of the diagonal structure whose length is l and it is dened as P(l)/sum(P(l)). It quanties the distribution of the diagonal line lengths. The more different the length of the diagonals, the more complex the deterministic structure of the RP and the larger the ENTR. 6) Recurrence Trend (RT) measures the variation rate of the REC away from the main diagonal which can reect the drift and non-stationarity in a time series. Firstly, the bottom right trigonal area of the RP is divided into K parts by K 1 equally spaced lines which parallel to the 45 diagonal, then the recurrence rate of the kth part RECk is obtained and REC represents the mean of the sequence {RECk , k = 1,2,. . .,K}. RT is the linear regression slope of the sequence {RECk , k = 1,2,. . .,K} about {k = 1,2,. . .,K} and it is calculated as follows: RT =
K k=1 (k

(K + 1)/2)(RECk REC)
K k=1 (k

(K + 1)/2)2

(7)

3.3.2. Pattern along the 135 diagonal This pattern corresponds to the situation that when xi and xj are close, xi+k and xjk are also close together for a series of k values. In this study, Lmax135 is extracted to characterize the pattern. It measures the length of the longest line segment which parallels to the 135 diagonal and it reects the longest duration of a stable interaction in the opposite direction. 3.3.3. Square-like pattern For a point xi , there maybe a sequence of point xj+1 ,. . .,xj+k which are all close to xi and they form the square-like pattern which is composed of vertical and horizontal lines. Vmax , TRAP, and Vmaxho are extracted to characterize the pattern here.

||fn mi ||
n=1

fn cluster i

(10) (11)

Dij = ||mi mj ||

where Li is the number of members in cluster i, fn is the nth feature vector of cluster i and mi is the mean feature vector of cluster i. In this paper, our goal is to predict the spontaneous termination of the AF and our interest lies in the classication of

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

1109

N, S and T AF. Thus the DB index to discriminate N and T AF is obtained through determining the case of separation for N and T AF. It is the same to discriminate S and T AF. They are calculated by Eq. (9). It is shown that the lower value of the DB index, the higher degree of separability between the AF with different termination properties. Based on the DB criterion, the goodness of the selected feature subset is evaluated in the SFS algorithm and the feature subset which has the lowest DB index can be obtained. 3.5. Multilayer perceptron neural network With the feature subset obtained in the SFS algorithm, two feed forward neural networksmultilayer perceptron (MLP) with three layers are used to predict the AF termination. One is used to discriminate N and T AF and the other is used to discriminate S and T AF. They are trained with the error back propagation. In the hidden and output layer, the transformation function is the sigmoid function. They are trained by the corresponding data in the Challenge database. After completing the training procedure of the MLP, the weights of the MLP are frozen and ready for the use in the testing mode. For an unknown segment, its feature subset forms the input to the MLP and it is classied to a certain type of the AF by the MLP. Since each recording is composed of 15 4-s segments, for each recording 15 decision results are obtained. The number of segments in a recording that are classied as different AFs is counted as l1 and l2 respectively. Then l1 and l2 are compared and the recording is decided belonging to the class from which the larger of l1 and l2 comes.

4. Results A two-sample KolmogorovSmirnov goodness-of-t hypothesis test is performed to determine whether there are any differences in these eleven features between any two kinds of AF signals. The probability value p < 0.05 is considered statistically signicant. Table 1 shows the mean and standard deviation of eleven features of the N, S and T AF recordings in the training data set, and the p value is also shown.

These eleven features all have signicant difference between N and T AF as implied by the asymptotic p value in Table 1. It can also be found that the proposed eleven features have less signicant difference between S and T AF than N and T AF for the training set. To discriminate N and T AF, the feature REC has the lowest p and DB value among eleven features, thus it is selected as the rst feature to identify N and T AF in the SFS algorithm. Finally, all features are included by the SFS algorithm and their order is REC, RT, DIV, TRAP, ENTR, DET, Lmean , Vmax , Vmaxho , Lmax135 and Lmax . Fig. 3(a) illustrates the DB index to discriminate N and T AF versus the feature number. It shows that the minimum DB index is obtained for the set of the rst six features REC, RT, DIV, TRAP, ENTR and DET. It implies that this feature subset incorporates most of the discrimination power of the entire eleven features for N and T AF. In the same way, to discriminate S and T AF, all features are included by the SFS algorithm in the order of REC, RT, Vmax , DIV, ENTR, DET, Lmean , TRAP, Lmax135 , Vmax and Lmax . Fig. 3 (b) illustrates the DB index to discriminate S and T AF versus the feature number. The feature subset which has the lowest DB index to discriminate S and T AF is nally obtained as REC, RT, Vmax , DIV, ENTR. The MLP neural network based on the feature subset selected by the SFS algorithm is used to classify the AF signals in the database. Different experiments are performed during the implementation of these classiers and the number of hidden neurons is determined taking into consideration the classication accuracies. For the MLP to discriminate N and T AF, the input and output numbers are six and two respectively, and the optimum number of hidden neurons is fteen. For the MLP to discriminate S and T AF, the input and output numbers are ve and two respectively, and the optimum number of hidden neurons is ten. Results for the training data set and testing data set are presented in Tables 2 and 3 respectively. Table 2 shows that the classication accuracy for N and T AF in the training data set achieves 100% (20/20), 97% (29/30) of the testing data set A is also correctly classied into N and T AF. The classication accuracy for S and T AF in the training data set achieves 100% (20/20), 95% (19/20) of the testing data set B is also correctly classied into S and T AF as shown in Table 3.

Table 1 The statistical analysis for eleven features of N, S and T AF recordings in the training data set Feature REC Lmax Lmean DET DIV ENTR RT Lmax135 Vmax TRAP Vmaxho N (Mean S.D.) 0.064 53.42 5.04 0.80 0.026 0.521 0.0006 22.43 25.17 5.32 24.51 0.050 41.35 1.85 0.13 0.017 0.070 0.0004 9.32 10.43 2.36 10.30 S (Mean S.D.) 0.036 49.64 4.34 0.74 0.029 0.550 0.0004 19.03 19.14 4.09 18.93 0.031 49.12 0.90 0.10 0.014 0.043 0.0003 6.46 7.80 1.13 7.05 T (Mean S.D.) 0.033 46.77 4.43 0.75 0.029 0.546 0.0004 18.49 19.73 4.10 19.79 0.016 36.09 0.91 0.08 0.013 0.043 0.0002 5.71 7.11 0.93 7.12 p-value (N&S) 8.26e009 0.0027 0.0011 3.33e006 0.0027 0.0018 8.26e009 0.0002 3.33e006 3.33e006 3.33e006 p-value (N&T) 6.73e012 0.0011 0.0062 3.45e008 0.0011 0.0062 1.82e010 0.0007 1.92e005 5.10e007 1.09e005 p-value (S&T) 0.51 0.80 0.61 0.51 0.80 0.27 0.61 0.98 0.51 0.42 0.51

1110

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

5. Discussion Since the PhysioNet/Computers in Cardiology proposed the challenge in 2004, many studies are participated [6,7,8,12,13,14,16]. 97% of the testing set A and 100% of the testing set B are correctly classied based on the overall peak frequency, short term peak frequency and amplitude [6]. The major AF frequency and the mean RR interval are considered in [7] for the classication and 93% of the testing set A is correctly classied. To discriminate N and T AF, the averaged RR interval was put into relationship with the dominant atrial frequency in [8]. A linear classier was then evaluated and 90% of testing set A is correctly classied. To discriminate S and T AF, a criteria based on the dominant atrial frequency trend of each signals in its last seconds and a criteria based on the dominant atrial frequency comparison between S and T AF were jointly used. 90% of testing set B is also correctly classied. A non-linear approach based on Gaussian mixture models of reconstructed state space is proposed [12]. The prediction accuracy of the approach is 80% for testing data A and 85% for testing data B. A set of features including both linear and non-linear are extracted from ECG signals and the related RR interval in [13]. Then a feed-forward neural network was employed for the classication and 86.7% of recordings in testing set A were correctly classied. Time-frequency measures are investigated in [16]. Finally the study only restricted the prediction to the most signicant feature, brillatory frequency and 90% of the testing set A is correctly classied. Expect the time-frequency measures three complexity measures are also investigated in [14], they are named spectral entropy, sample entropy, and fractional spectral radius. But none of the three complexity measures has signicant differences between N and T AF, thus the same measure was used and the same results were obtained as the study in [16]. To make a summarization, several of most successful methods in the Challenge rstly subtract the QRST complexes from ECG signals and then analyze the residual signals which were dominated by atrial activity from the frequency point of view. Our study proposes a method to predict the spontaneous termination of the AF from a new perspective. Compared with the method mentioned above, the algorithm proposed in this paper achieves a higher or comparable accuracy for the prediction of the spontaneous terminated AF. It is the rst attempt to characterize the AF based on the structure and quantication of the RP from the non-linear point of view to predict the AF termination. The RP represents the recurrence of the state space trajectory to a certain state, which is a fundamental property of dynamical systems. Features based on the structure of the RP give some insight into the dynamics and complex patterns of activation of the atrium during AF. They have signicant difference among the AF with different termination properties, which demonstrates the potential ability of the algorithm to predict the AF termination from Holter ECG recordings effectively.

Fig. 3. (a) DB index to discriminate N and T AF vs. the feature number. (b) DB index to discriminate S and T AF vs. the feature number.

Table 2 Classication into non-terminating and immediately terminating AF for the recordings in the training data set and testing data set A Data category Actual type Experiment result N Training data set Testing data set A N T N T 10 0 15 0 T 0 10 1 14

Table 3 Classication into soon-terminating and immediately terminating AF for the recordings in the training data set and testing data set B Data category Actual type Experiment result S Training data set Testing data set B S T S T 10 0 10 1 T 0 10 0 9

R. Sun, Y. Wang / Medical Engineering & Physics 30 (2008) 11051111

1111

The feature dimensionality reduction cannot only reduce the computational complexity but also improve the classiers generalization ability since it determines a feature subset which is most essential for the classication. Experiment results demonstrate that the DaviesBouldin criterion can evaluate the goodness of the feature subset for the AF termination prediction effectively and the neural network based on the selected feature subset achieves satisfying results. The feature subset that has the most powerful ability to discriminate N, S and S, T AF is different. It demonstrates that even for the same feature subset, its ability to discriminate N, S and S, T AF is different. Moreover, some of these methods which characterize AF signals referred above, for example [11], are based on the epicardial atrial electrograms. The proposed method differs from them in that its purpose is to predict the termination of the AF using the surface ECG signals which are non-invasive and more convenient to obtain. The size of the database is small and a larger dataset will be very useful for validation and improvement of the algorithms performance.

6. Conclusions This study leads to a new perspective in predicting the termination of the AF from the non-linear point of view. It demonstrates that the structures of the RP contain the useful information which can characterize distinct levels of chaoticity among AF signals with different termination properties. It will provide a large number of clues which enable the further development of more toolkits to predict the spontaneous termination of the AF. In conclusion, the new algorithm described in the study can be implemented and provide accurate prediction for the spontaneous termination of the AF.

Acknowledgements This work was supported by the National Basic Research Program of China (No. 2006CB705707), Natural Science Foundation of China (No.30570488) and Shanghai Leading Academic Discipline Project (No. B112).

Conict of interest There are no nancial and personal relationships with other people or organizations that could inuence the paper.

References
[1] Halperin JL, Hart RG. Atrial brillation and stroke: new ideas, persisting dilemmas. Stroke 1988;19:93741.

[2] Al-Khatib SM, Wilkinson WE, Sanders LL, McCarthy EA, Pritchett EL. Observations on the transition from intermittent to permanent atrial brillation. Am Heart J 2000;140:1425. [3] Botteron GW, Smith JM. Quantitative assessment of the spatial organization of atrial brillation in the intact human heart. Circulation 1996;93:5138. [4] Botteron GW, Smith JM. A technique for measurement of the extent of spatial organization of atrial activation during atrial brillation in the intact human heart. IEEE Trans Biomed Eng 1995;42:57986. [5] Langley P, Bourke JP, Murray A. Frequency analysis of atrial brillation. Comput Cardiol 2000;27:6572. [6] Petrutiu S, Sahakian AV, Ng J, Swiryn S. Analysis of the surface electrocardiogram to predict termination of atrial Fibrillation: the 2004 Computers in Cardiology/Physionet Challenge. Comput Cardiol 2004;31:1058. [7] Hayn D, Edegger K, Scherr D, Lercher P, Rotman B, Klein W, et al. Automated prediction of spontaneous termination of atrial brillation from electrocardiograms. Comput Cardiol 2004;31:11720. [8] Cantini F, Conforti F, Varanini M, Chiarugi F, Vrouchos G. Predicting the end of an atrial brillation episode: the PhysioNet challenge. Comput Cardiol 2004;31:1214. [9] Faes L, Nollo G, Antolini R, Gaita F, Ravelli F. A method for quantifying atrial brillation organization based on wave-morphology similarity. IEEE Trans Biomed Eng 2002;49:150413. [10] Sih HJ, Zipes DP, Berbari EJ, Olgin JE. A high-temporal resolution algorithm for quantifying organization during atrial brillation. IEEE Trans Biomed Eng 1999;46:44050. [11] Kao T, Su YY, Tso HW, Lin YC, Chen SA, Tai CT. Non-linear analysis of human atrial utter and brillation using the surface electrocardiogram. Comput Cardiol 2004;31:4414. [12] Roberts FM, Povinelli RJ. A statistical feature based approach to predicting termination of atrial brillation. Comput Cardiol 2004;31:6736. [13] Mainardi LT, Matteucci M, Sassi R. On predicting the spontaneous termination of atrial brillation episodes using linear and non-linear parameters of ECG signals and RR series. Comput Cardiol 2004;31:6658. [14] Nilsson F, Stridh M, Bollmann A, Sornmo L. Predicting spontaneous termination of atrial brillation using the surface ECG. Med Eng Phys 2006;28:8028. [15] Stridh M, Sornmo L, Meurling CJ, Olsson SB. Sequential characterization of atrial tachyarrhythmias based on ECG time-frequency analysis. IEEE Trans Biomed Eng 2004;51:1104. [16] Nilsson F, Stridh M, Bollmann A, Sornmo L. Predicting spontaneous termination of atrial brillation with time-frequency information. Comput Cardiol 2004;31:65760. [17] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 2000;101:e21520. [18] Eckmann JP, Kamphorst SO, Ruelle D. Recurrence plots of dynamical systems. Europhys Lett 1987;5:9737. [19] Shkurovich S, Sahakian AV, Swiryn S. Detection of atrial activity from high-voltage leads of implantable ventricular debrillators using a cancellation technique. IEEE Trans Biomed Eng 1998;45:22934. [20] Takens F. Detecting strange attractors in turbulence. Lect Notes Math 1981;898:36681. [21] Parker TS, Chua LO. Practical numerical algorithms for chaotic systems. New York: SpringerVerlag; 1989. pp. 1132. [22] Webber CL, Zbilut JP. Dynamical assessment of physiological systems and states using recurrence plot strategies. J Appl Physiol 1994;76:96573. [23] Marill T, Green DM. On the effectiveness of receptors in recognition systems. IEEE Trans Inform Theory 1963;9:117. [24] Boostani R, Moradi MH. Evaluation of the forearm EMG signal features for the control of a prosthetic hand. Physiol Meas 2003;24:309 19.