Learning System for Computer-Aided ECG Analysis
Based on Support Vector Machines

Stanislaw Jankowski^a, Artur Oreziak^b

^aInstitute of Electronic Systems, Center for Complex Systems Research,
Warsaw University of Technology,Warsaw, Poland
^bChair and Department of Internal Medicine and Cardiology, Central Teaching Hospital, Warsaw, Poland

Correspondence: S Jankowski, Institute of Electronic Systems, Warsaw University of Technology, 00-665 Warsaw, ul. Nowowiejska 15/19, Poland. E-mail: sjank@ise.pw.edu.pl, phone +48 22 660 5589, fax +48 22 825 2300

1.    Introduction

We present a new tool for computer-aided analysis of ECG Holter recordings. This tool implements an idea of supervised learning from examples by using support vector machines [Cristianini, 2000; Schölkopf, 1999; Vapnik 1998]. The knowledge is incorporated by a careful selection of data set commented by specialists. Then by using statistical procedures called “learning”, a classifier is designed. The successful recognition depends strongly on the quality of learning set (selection of cases), data representation and the mathematical basis of the classifier. The system is tested at the Chair and Department of Internal Medicine and Cardiology, Central Teaching Hospital in Warsaw, Poland.

2.    Learning System for ECG Classification

The presented system is based on our previous results [Jankowski et al., 2002]. The ECG Holter recordings are filtered and segmented into single beats. We apply the wavelet analysis to detect R-points. Then a sample of beats is analyzed and labeled by cardiologists. Hence, a learning set was obtained. All data are parameterized by using the support vector approximation with Gaussian kernels. The width of Gaussian function may be tuned with respect to signal shape. The centers are support vectors. The function coefficients are equal to Lagrange multipliers. This procedure transforms the digital signal into 30-dimensional vectors of Lagrange multipliers; each vector encodes a single heartbeat. Then the classifier is trained giving rise to a certain number of support vectors. The low number of support vectors with respect to the total number of beats in the training file states the efficiency of heartbeat shape representation. For multiclass support vector machine we perform several one-to-all classifiers. The number of support vectors for classification depends on the problem complexity. We attempt to minimise the number of support vectors by experiments with various kernel functions.

The unified mathematical basis of support vector machines for approximation and classification enables to obtain perfect generalisation properties and to perform efficient numerical program. For both tasks we use sequential minimal optimisation algorithm. The program code is written in C++ language.

3.    Results

In order to estimate the system functionality we examined one case: the ECG Holter recordings of a patient having sinus rhythm with multifocal ventricular contractions. The heartbeats were assigned into 3 classes corresponding to: normal beats and 2 classes of pathological shapes from 2 focal ventricular contractions, probably LBBB – left bundle branch block and RBBB – right bundle branch block, as shown in Fig. 1. Learning set consisted of 1712 heartbeats. We obtained 3 linear classifiers using one-against-all scheme with the following numbers of support vectors: normal against all pathological beats: 46; RBBB-like morphology against all other beats: 36; LBBB-like morphology against all other beats: 20. The test results for 9690 heartbeats are listed in Table 1.

Figure 1. Several heartbeats with class labels assigned by a cardiologist.

Table 1.    Classification validation for a test dataset of 9852 heartbeats.

4.    Conclusions

The learning system for ECG computer-aided analysis is a flexible and open tool for cardiologists in order to perform the classification of all heartbeats from the Holter ECG recordings upon the shape. The cardiologists can define the goals of classification by the choice of learning sets with respect to all details of the electrocardiograms. The test results are consistent and encouraging. We hope that this system may be applied to long recordings in order to search for precursors of dangerous events and risk prediction.

Acknowledgements

This work was supported by the Dean of the Faculty of Electronics and Information Technology, Warsaw University of Technology. We thank Prof. Dr. Med. G. Opolski and Prof. J. J. Zebrowski for helpful remarks.

References

N. Cristianini, J. Shaw-Taylor: Support Vector Machines, Cambridge University Press, 2000

S. Jankowski, J. Tijink, G. Vumbaca, M. Balsi, G. Karpinski: Support vector machine for the recognition of atrial and ventricular de[polarization in Holter ECG recordings, IJBEM vol. 4, No. 2. pp. 343-344, 2002

S. Jankowski, J. Tijink, G. Vumbaca, M. Balsi, G. Karpinski: Morphological analysis of ECG Holter recordings by support vector machines, Medical Data Analysis ISMDA 2002 (eds. A. Colosimo, A. Giuliani, P. Sirabella), Lecture Notes in Computer Science 2526, Springer, Berlin 2002, pp. 134-143.

V. N. Vapnik Statistical Learning Theory, Wiley, New York, 1998.