Atrial Activity Detection through a Sparse Decomposition Technique

2008 International Conference on BioMedical Engineering and Informatics

Atrial Activity Detection Through a Sparse Decomposition Technique S. Paredes†, T. Rocha†, P. de Carvalho‡, J. Henriques‡ †Instituto Politécnico de Coimbra, Departamento de Engenharia Informática e de Sistemas, Rua Pedro Nunes, 3030-199 Coimbra. {[email protected], [email protected]} ‡CISUC, Departamento de Engenharia Informática, Universidade de Coimbra, Pólo II, 3030-290 Coimbra {[email protected], [email protected]} activity (AA) is required for the automatically detection and characterization of the AF. In fact, ventricular activity (VA) is spectrally overlapped with AA, has a larger amplitude level than the AA and there is a partial mixture of the two activities. Several methods have been proposed for atrial and ventricular separation, which can be grouped in two main categories: i) QRST cancellation approaches and ii) separation of atrial and ventricular activities. Within the group of QRST cancellation methods, i.e., that cancels the ventricular activity, it is possible to highlight the Average Beat Subtraction method [2]. This method assumes that several QRS templates can be created according to different morphologies, and subtracted from the original signal. Spatiotemporal QRST cancellation method takes into account the variations of beat amplitude, as well as the orientation of QRS average vector to make a more accurate cancellation [5]. Wavelets decomposition methods, that decompose the original signal and afterwards, based on local extreme, enable to reconstruct the signal, have been also proposed [6]. However, QRS cancellation methodologies introduce some problems, namely: QRS are highly dependent on beat variations, cancellation is never perfect and there are always residues originated by an inefficient QRS cancellation. An alternative approach to this problem is the separation of activities without an explicitly cancellation of ventricular activity. In this domain, it is possible to emphasize the importance of Principal Component Analysis method that, through the right linear transformations, permits the identification of sources of atrial and ventricular activity [7][8]. The main drawback of this type of technique is related with the requirement of several ECG leads, which hinders its application to portable systems. The approach followed in this work, Sparse Decomposition, is included in this second group of techniques. Sparse Decomposition (SD) technique enables to decompose the original signal according to a dictionary

Abstract Atrial fibrillation is the most common cardiac arrhythmia, presenting significant consequences on patient health. Automatic detection of atrial fibrillation needs, ideally, the isolated study of the atrial activity registered in the electrocardiogram. Sparse decomposition techniques make possible the decomposition of a signal into their components, thus the separation between atrial and ventricular activities. However, this technique requires the a priori construction of distinct dictionaries, usually built based on atrial and ventricular activity simulation models. This work addresses the construction of the dictionaries based on real electrocardiogram signals, where P-waves, QRS-complexes and T-waves are first identified to support the creation of the dictionaries. The effectiveness of the proposed methodology is validated with real signals, obtained from MIT-BIH Arrhythmia Database. Index Terms: Atrial Fibrillation; Sparse Decomposition; Orthogonal Matching Pursuit.

1. Introduction Atrial Fibrillation (AF) is the most common sustained cardiac arrhythmia; approximately 0.4%1.0% of the general population suffers from this illness [1]. This type of arrhythmia may appear anytime, but its prevalence tends to increase with age; up to 10% of the population older than 70 have been diagnosed with AF [1][3]. Although not life-threatening, AF may severely impact the quality of life and increase the risk of stroke, justifying its diagnosis. AF diagnosis is typically assessed by clinicians by visual inspection of the surface electrocardiogram (ECG), mainly taking into account the characteristics of P wave [2]. The previous separation of the atrial

978-0-7695-3118-2/08 $25.00 © 2008 IEEE DOI 10.1109/BMEI.2008.241

1282 358

of functions [1]. Moreover, SD only needs one ECG lead to achieve decomposition of the original signal and does not have to deal with QRS cancellation residues difficulties. The main challenges of the method are the construction of dictionaries, which are usually based on simulation models of atrial and ventricular activities, and how to avoid wrong dictionary elements selection. In this work it is proposed a new technique to build the dictionaries. The construction of dictionaries is based on real signals, being P-waves, QRS-complexes and T-waves extracted, through a segmentation process, to create the elements of the dictionaries. Once the atrial activity is isolated it is possible to perform respective frequency spectrum analysis. This information conjugated with RR interval analysis has the potential to perform an accurate detection of AF. This paper is organized as follows: in the next section, we present the sparse decomposition technique, how this approach is applicable to electrocardiogram decomposition, and the details of the proposed method to build the dictionaries. In section 3, the results from the experimental evaluation of this approach are presented and discussed. The paper concludes discussing some potentialities of the work currently underway.

D=

∪D

i

1≤ i ≤ d

(2)

i=

An ECG signal can be considered a combined composition of two activities, atrial activity (fAA) and ventricular activity (fVA), as is stated in equation (3). f ECG = fVA + f AA + r

(3)

Thus, it is necessary to construct a dictionary composed by two dictionaries, one accountable for representing atrial activity ( DAA ) and the other for representing the ventricular activity ( DVA ). Figure 1 depicts this situation, being M the number of atoms of DVA, L the number of atoms of DAA and N the atom dimension.

Figure 1 - Dictionary AA/VA [1]

2. sparse decomposition 2.1 Dictionaries

Sparse decomposition is based on the correct identification, within a given dictionary, of the set of segments (usually known by atoms) that adequately represents the signal to be decomposed. So, with the right algorithm it is possible to represent an original signal based on a set of simple atoms. Atoms of a dictionary do not have necessarily to be linearly independent, and in these conditions (not linearly independent) dictionaries are designated by overcomplete. In this situation there is no single solution for equation (1) [1], [9], [11].

According to common approaches [1], dictionaries are previously built using simulation models for both atrial and ventricular activities. Here we propose an alternative strategy, where the models are obtained from real ECG samples. At a first stage a segmentation algorithm, developed under MyHeart project, is applied [16]. This process allows the identification of P-wave, QRS complex and T-wave. Then, in a second stage, models based on the average of the selected samples waves are created. These average models are translated a fixed number of times (N times), corresponding to the length of each atom, resulting on N distinct models. In this way it is possible to decompose the original signal in to a set of successive segments of the same length of each atom. Otherwise, it would be needed to align each atom with the segment to be decomposed. However, using this strategy, some problems have been observed in the atoms detection. In fact, there is a high intersection between the two activities and a high coherence between the sub-dictionaries, since P-wave and T-wave have similar shapes. However, as it is well known, there is no ventricular activity during some period of a heart cycle, as shown in Figure 2. This fact can be used to assist the construction of dictionaries.

m

f =

∑b g

k k

+r

(1)

k =1

Variable f identifies the original signal to be decomposed, gk is an atom and bk is the respective coefficient. Error is defined by variable r. Most times dictionaries are overcomplete, enabling the determination of the best atom combination to decompose the signal. Additionally, the dictionary (D) is typically composed by two or more sub-dictionaries (Di), equation (2), which make possible to represent different features of the original signal.

1283 359

calculated it is necessary to reconstruct the signals. Each activity is reconstructed with its own dictionary. fVA ≡ DVAbVA (7) f AA ≡ DAAbAA Therefore this method obtains atrial and ventricular activities without making the QRST direct cancellation, which is one its main advantage.

3. Results 3.1 Data set In this work ECG data signals obtained from MITBIH Arrhythmia Database [15] were used for validation purposes. This database contains 23 records (the “100 series”) chosen at random from a set of over 4000 24-hour Holter tapes, and 25 records (the “200 series”) selected from the same set to include a variety of rare but clinically important phenomena that would not be well-represented by a small random sample. Each record is slightly over 30 minutes in length and each signal file contains two signals sampled at 360 Hz. Six signals with 10s duration have been selected (100, 101, 103, 201, 202 and 219) and for all signals it only the lead II has been considered. These signals have been selected because they clearly reveal the main features of a Normal Sinus Rhythm signal, (100, 101, 103) and an AF signal, (201, 202, 219). The original signals selected have been filtered with three different digital filters: 4ª order high-pass Butterworth filter fc=0.2Hz; 4ª order low-pass Butterworth filter with cutoff frequency of 60 Hz and a 5ª order notch Chebyshev filter fc=60 Hz. The selected filters were used to remove noise from the signal since they allow a fast roll-off with a controlled ripple level

Figure 2 – ECG modeling

Thus, the DVA might be deactivated in a specific interval. A QRS detection algorithm is used to identify the RR interval [10] and the QT interval at the instant n is calculated using the Bazzet formula [14], based on the current (Rn) and next R peak (Rn+1): QTn = 0.5 Rn Rn +1

(4)

Using (4) it is possible to determine the segment where DVA is deactivate, calculated according to: Seg DeactivateDVA = Qn +1 − (Qn + QT )

(5)

This strategy reduces significantly the intersection between dictionaries, as it avoids the selection of DVA atoms within the segment defined in (5).

2.2 Coefficients Once determined the atoms some approaches have been proposed in order to find the respective coefficients [4][10]. Given their properties, the Orthogonal Matching Pursuit Algorithm [12] has been used in this work. This algorithm decomposes a residue rk f through its projection on a vector of D that matches rk f at best. According with the following equation the kth term of original signal (f) decomposition is given by:

3.2 Dictionaries construction Initially, dictionaries were built based in three wave models obtained with samples extracted from MITBIH ECG 101. This signal was selected since it contains the three waves very clearly and well defined. In order to achieve the best performance, considering the capability to detect different activities and also the algorithm speed, several dictionaries were tested with different number of atoms. From several experiments, a dictionary composed by 150 atoms was selected.

K −1

f =

∑ r f,g k

γk

, gγ k + rk f

(6)

k =0

Each atom selected for the decomposition should maximize the scalar product. Moreover, rk f should be orthogonal not only towards g k f but also in relation to all the other atoms already selected for the signal decomposition. Using this approach, selected atoms will not be chosen again. Once coefficients are

 D =  D p DQRS 50×100  50×50 50×50

1284 360

 DT  50×50 

(8)

The dimension of each atom was also tested, since smaller length should allow more accurate detection but also implies reduction of speed algorithm. The length of each atom was defined as 50 points. In a second phase, models extracted from the signals to be decomposed were introduced in dictionary. As explained in section 2.1 segmentation algorithms were used to isolate the samples to build the average models of each wave. In these conditions, dictionaries were composed by 300 atoms as described in (9).  D =  D p D p Signal 50×50  50×50

50×300

DQRS

DQRS Signal

50×50

50×50

DT

50×50

DT



Signal  50×50 

ECG

AA

VA

(9)

Detection was improved and there was no clear degradation of algorithm speed. So, these dictionaries were used to decompose the original ECG signals.

Figure 4 – 1.Ecg_219; 2 . Atrial Activity; 3. Ventricular Activity (AF – Signal)

3.3 Experiments From the six signals three presents a normal sinus rhythm (100, 101, 103) and the other three atrial fibrillation events (201, 202, 219). The main goal was to assess the effectiveness of Sparse Decomposition technique in both situations. As can be seen in Figure 3, and 4, (respectively normal sinus rhythmic and atrial fibrillation signals) the two activities are correctly separated.

Good approximations to ventricular activity were obtained with all tested signals; which made it possible to confirm the presence of QRS complexes and Twaves in tested situations. Atrial activity, as expected, clearly presents the P-waves. However, this evaluation is more difficult since there is no structural reference of the signal and it has low amplitude. From results it was possible to validate the effectiveness of the SD technique when applied to ECG signals. Additional experiments to other ECG signals have confirmed the high potential to do the separation between atrial and ventricular activities. Once atrial and ventricular activities have been isolated several approaches can be applied in order to automatically detect atrial fibrillation. Power Spectrum Density of atrial activity analysis together with RR interval analysis can be an effective method to detect AF events.

ECG

AA

4. Conclusions

VA

In this paper sparse decomposition techniques have been applied to distinguish between atrial and ventricular activities in one lead electrocardiogram. The construction of the dictionaries, based on real electrocardiogram signals, was the main contribution of the work. Future work will explore isolated atrial activity conjugated with RR interval analysis to produce an accurate method to detect atrial fibrillation events.

Figure 3 – 1.Ecg_101; 2 . Atrial Activity; 3. Ventricular Activity (NSR – Signal)

1285 361

5. REFERENCES [1] Hernandez, J., “Sparse Decompositions for Ventricular and Atrial Activity Separation”, Master’s Thesis, Signal Proc. Institute, Ecole Polytechnique Fédérale de Lausanne (EPFL), August 2005. [2] Ying, H. “Algorithms in Automatic Detection of Atrial Fibrillation in ECG Signals”, Uviversität Karlssruhe, 2005. [3] Bialy, D., Lehmann, M., Shummacher, D., Steinman, R., Meissner, M. “Hospitalization for arrhythmias in the United States: importance of atrial fibrillation”, J. Am Cardiol. 19 (Suppl. A), nº 41 A, pp. 612-627, 1992. [4] Escoda O., et all, “Ventricular and Atrial Activity Estimation Through Sparse ECG Signal Decompositions”, International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Vol. II, pp. 1060-1063, 2006. [5] Stridh. M., Sornmo. L., “Spatio Temporal QRST Cancellation Techniques for Analysis of Atrial Fibrillation”, IEEE Transactions on Biomedical Engineering, Vol. 48, January, 2001. [6] Carrault. G, Hernandez. A, I., Senhadji L., “Wavelets Extrema Representation for QRS-T Cancellation and P Wave Detection”, IEEE, Comp. in Cardiology, Vol. 29, pp 37-40, 2002. [7] Sanches. C., Rieta. J., Castells. F, “Atrial Activity Extraction in Holter Registers using Adaptive Wavelet Analysis”, IEEE, Computers in Cardiology, Vol. 30, pp.569572, 2003. [8] Langley P., Bourke JP., Murray A., “Frequency Analysis of Atrial Fibrillation”, Computers in Cardiology, Vol. 27, 65-68, 2000. [9] Raine D., et all “Surface Atrial Frequency Analysis in Patients with Atrial Fibrillation: A tool for evaluating the effects of intervention”, Journal of Cardiovascular Electrophysiology, Vol. 15, Nº 9, September, 2004. [10] Pan J, Tompkins W: A real-time QRS detection algorithm. IEEE Trans Biomed Eng 1985, 32:230-236. [11] Zibulevsky M., Pearlmutter B., “Blind Source Separation by Sparse Decomposition in a Signal Dictionary”, Neural Computation, 13, pp. 863-882, 2001. [12] Rieta, J., et all, “Atrial Fibrillation, Atrial Flutter and Normal Sinus Rhythm Discrimination by Means of Blind Source Separation and Spectral Parameters Extraction”, Computers in Cardiology, Vol. 29, pp 25-28, 2002 [13] Tropp, J. “Greed is Good: Algorithms results for Sparse Approximation.”, IEEE, 2004. [14] Extramiana, F. et al. “Individual QT/RR relationship: average stability over time does not rule out an individual residual variability implication for the assessment of drug effect on the QT interval” Lariboisi re Hospital, Cardiology Dept., Paris, France; Europace 2005. [15]http://www.physionet.org/physiobank/database/mitd/, accessed in February of 2007. [16]MyHeart is a FP6 project, IST-2002-507816 http://www.extra.research.philips.com/euprojects/myhear/, accessed in February of 2007.

1286 362