Proceedings - 19th International Conference - IEEEIEMBS Oct. 30 - Nov. 2, 1997 Chicago, IL. USA
BEAT DETECTION AND CLASSIFICATION OF ECG USING SELF ORGANIZING MAPS I
Marcel0 R. Risk’, Jamil F. Sobh2,J. Philip Saul2 Fundacion Universitaria Dr. Rene G. Favaloro and Instituto Tecnologico de Buenos Aires, Solis 453 (1078), Buenos Aires, Argentina. 2 Dept. of Cardiology, Children’s Hospital and Harvard Medical School, 300 Longwood Ave, Boston, MA 02 1 15 USA. ‘E-mail:
[email protected]
Abstract - A procedure for beat detection and classification was develop using ECG recordings. This procedure can use for beat detection one or two leads, and then a portion of each detected beat is used for classifying, this task is performed by a neural network; in our work the morphology of the QRS portion of the ECG feeds a self organizing map (SOM). The SOM was previously trained with different QRS complex such as normal and ectopic beat morphologies. The beat classification is very important in the heart rate variability (HRV) analysis, because must be used only the normal beats and rejected the ectopic ones for the construction of the RR intervals beat series. I.
INTRODUCTION
The accurate detection and classification of QRS complex are very useful in the study of HRV [ 11, and in arrhythmia studies, among others. The aim of this work is to provide a simple procedure for beat detection and classification, using a preprocessing algorithm for beat detection, and then a SOM for beat classification. The preprocessing algorithm can use one or more leads, to allow lead independence in case of noise and missed lead, and is based on nonlinear transforms and derivatives [2][3]. The second procedure uses a SOM for beat classification, allowing the use of normal beats from each patient for training of the SOM, and therefore the procedure is customize to the QRS morphology of each patient.
y(nT) = ( x ( n T - T ) + x ( n T - 2 T ) (2) + x(nT - 3T))/ 3 The output of the filter is differentiated using Eq ( 3 ) equation, using a three-point central derivative:
y(nT) = ( x ( n T )- x ( n T - 2T)) / 2
(3)
The peak location is determined when the squared derivative d2 (from Eq. 3) cross a threshold value th. The th is determined like a percentage of the maximum value of d2. The fiducial point is identified within the next 50 ms after the crossing th, searching the maximum value of the vector inside this interval. The Figure 1 shows the vector and d2 signals. The th is dynamically set, updating its value from each segment of 30 seconds of the whole recording in study, using a predefined percentage of the maximum value of d2. The predefined percentage is a user defined option, typically set at 50 %, and in presence of noise can be adjusts to higher values.
l
f
i
l
11. METHODS
Preprocessing algorithm The preprocessing algorithm calculates the vector of the Fig. 1: A: vector, B: d2. ECG using all the.available leads (11, 12 , ... In) are squared and added to calculate the vector ( x ( n T ) ) , All the fiducial points determine the peak location implemented with the difference equation set, from which can be calculated the beat series of RR intervals, and therefore the time series for HRV analysis
x(nT) = Jr:(nT) +l;(nT)+...+l:(nT)
~41.
(1)
Where n is an arbitrary integer and T is the sampling period. The next step has the option of filtering, in this case. a simple FIR low pass filter may be used, implemented with the difference equation
Beat classijkation Beat classification is performed using a SOM, which analyzes the QRS morphology, determining regions in a two-dimensional array of nodes, for all the typical
89 (0-7803-4262-3/97/$10.00 (C) 1997 IEEE)
Authorized licensed use limited to: Harvard University SEAS. Downloaded on February 8, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
Proceedings - 19th International Conference - IEEE/EMBS Oct. 30 - Nov. 2, 1997 Chicago, IL. USA
morphologies, called clusters [ 5 ] [ 6 ] . In this structure of nodes, each of them is connected to the inpuj array via a variable connection weight. The nodes conform the map, and once trained the SOM, can be presented a new pattern to the SOM, and one winner node is determined; the classification of this new pattern is defined by the coordinates of the winner node in the map. The winner node is determined like the node with the minimum euclidian distance, computed for each node using the following expression:
iterations, the q coefficient is decreased and the neighborhood of the winner node is decreased too, using a gaussian function [5][6], with the following equation:
Where i a n d j are the coordinates around the node to update, belonging to the neighborhood, and a is a constant.
Implementation Where d is the euclidian distance, j is the node index, N is the number of samples of the input vector, x is the input vector, i is the index of x vector, w is the weight vector. In this work we use an input array of 50 samples, corresponding to 390 ms with a sampling rate of 128 Hz, this amount of samples is enough to enter both a typical QRS complex and a pathological QRS, such as a ventricular (ectopic) beat. One lead is selected to feed the SOM, the criteria of selection are based on availability of data over all the record, low noise and better ectopic morphologies. In the Figure 2 we can appreciate a typical normal QRS complex used for training the SOM, and in the Figure 3 we can see a QRS complex corresponding to a ventricular beat.
L
1
11
21
31
41
The preprocessing algorithm and the SOM were implemented in Borland C++, programming a class called Som. Two channels of ECG of one segment of 1 hour from 80128 record (long term data), of the HMSMIT-FFMS database was used [7]. This program run on a Pentium 120 MHz PC, taking a few seconds the training of the SOM. 111. RESULTS The results after 3000 iterations, training the SOM with 5 typical normal and two ectopic beats, conform a map showed in the Table 1 .
Table 1: Map of the SOM after the training.
51
The Table 2 shows the map after the presentation of all the beat in the 1 hour record.
I
Fig. 2: QRS complex of a typical normal beat.
Table 2: Map of the SOM after all the beats of the record. IV. DISCUSSION
Fig. 3 : QRS complex of a ventricular beat. The input array is connected to a two-dimensional map of 8 by 6 nodes, and during the training process of 3000
The accurate beat detection and classification have importance in the HRV analysis, and the morphology recognition of normal and ventricular beats allows the calculation of HR time series with data derived only from the mechanism of regulation of the cardiovascular
90
Authorized licensed use limited to: Harvard University SEAS. Downloaded on February 8, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
Proceedings - 19th International Conference - IEEE/EMBS Oct. 30 - Nov. 2, 1997 Chicago, IL. USA
system. This analysis has applications in clinical research [81. This work describes a preprocessing algorithm that combines a vector computation, a simple linear filter, and a threshold based on squared derivative of the vector. This algorithm is simple and fast, and can be used to compute large amount of data, like recordings of 24 hours of ECG (Holter studies). The SOM applied in this manuscript provides a method that uses the QRS morphologies of each patient in study, preventing changes of levels of the signals and lead independence. The combination between the preprocessing algorithm and the morphological analysis using SOM, results in an alternative to others beat detection and classification methods, and further testing is required to determine its utility in analyzing the broad range of 24 hour ECG recordings collected in clinical research studies. REFERENCES
[l] J. P. Saul, P. Albrecht, R. D. Berger and R. J. Cohen, “Analysis of long term heart rate variability: methods, l/f scaling and implications,” in Comput. Cardiol., pp. 419-422, 1988. [2] P. S. Hamilton and W. J. Tompkins, “Quantitative investigation of QRS detection rules. Using the MITIBIH arrhytmia database,” IEEE Trans. Biomed. Eng., vol. BME-33, no. 12, pp. 1157-1165, 1986. [3] M. R. Risk, J. F. Sobh, R. Barbieri, J. P. Saul, “A simple algorithm for QRS peak location: use on long term ECG recordings from the HMS-MIT-FFMS database,” in IEEE Engineering in Medicine and Biology 17th Annual Conference. 1995. R. D. Berger, S. Akselrod, D. Gordon and R. J. Cohen, “An efficient algorithm for spectral analysis of heart rate variability,” in IEEE Trans. Biomed. Eng., vol. BME-33, no. 9, pp. 900-904, 1986. T. Kohonen. “Self-organization and Associative Memory,” Springer-Verlag, Berlin. 1984. [6] R. P. Lippmann. “An introduction to computing with neural nets,” in IEEE ASSP Magazine, April 1987, pp. 4-22. [7] J. F. Sobh, M. R. Risk, R. Barbieri, J. P. Saul, “Database for ECG, arterial blood pressure, and respiration signal analysis: feature extraction, spectral estimation, and parameter quantification,” in IEEE Engineering in Medicine and Biology 17th Annual Conference. 1995. [8] T. J. Bigger, J. L. Fleiss, R. C. Steinman, L. M. Rolnitzky, R. E. Kleiger, J. N. Rottman. “Correlation’s among time and frequency domain measures of heart period variability two weeks after acute myocardial infarction,’’ in Am. J. Cardiol. 69:891-898, 1992.
91
Authorized licensed use limited to: Harvard University SEAS. Downloaded on February 8, 2010 at 08:01 from IEEE Xplore. Restrictions apply.
