A patient adaptable ECG beat classifier based on neural networks

Abstract

A novel supervised neural network-based algorithm is designed to reliably distinguish in electrocardiographic (ECG) records between normal and ischemic beats of the same patient. The basic idea behind this paper is to consider an ECG digital recording of two consecutive R-wave segments (RRR interval) as a noisy sample of an underlying function to be approximated by a fixed number of Radial Basis Functions (RBF). The linear expansion coefficients of the RRR interval represent the input signal of a feed-forward neural network which classifies a single beat as normal or ischemic. The system has been evaluated using several patient records taken from the European ST-T database. Experimental results show that the proposed beat classifier is very reliable, and that it may be a useful practical tool for the automatic detection of ischemic episodes.

Introduction

The electrocardiogram (ECG) is a graphic recording of the electrical activities in human heart and provides diagnostically significant information. Its shape, size and duration reflect the heart rhythm over time. The waves related to electrical impulses occurring at each beat of the heart are shown in Fig. 1. The P-wave represents the beginning of the cardiac cycle and is followed by the QRS complex, which is generally the most recognizable feature of an ECG waveform. At the end of the cardiac cycle is the T-wave. The varied sources of heart diseases provide a wide range of alterations in the shape of the electrocardiogram. For instance, inverted T waves (Fig. 2) are seen during the evolution of myocardial infarction, while ST-segment depression (Fig. 3) can be caused by ischemia. In recent years, many researches concerning automated processing of ECG signals have been conducted ([1], [4], [5], [7], [10], [11], [13], [14]). A difficult problem in computer-aided ECG analysis is related to the large variation in the morphology of ECG waveforms, not only among different patients, but even within the same patient. This makes the detection of ECG features (ST-segment, T-wave, QRS-area) a tough task.

The aim of the present work is to design an ECG beat recognition method to distinguish normal and ischemic patterns of the same patient without requiring the extraction of ECG features. The method is a supervised neural network-based algorithm, and hence its applicability requires the availability of recordings of both normal beats and annotated ischemic episodes. Once trained, the system should be capable of detecting new ischemic beats similar to those previously observed in the patient. This objective is relevant because of the observed tendency of patients discharged after an acute myocardial ischemic attack to repeat a similar ischemic episode within a short period of time (up to 30% within one month). The possibility of automatically monitoring such patients would then translate directly into reduced morbidity and mortality.

In the proposed method, the pattern under classification is defined to be the ECG signal from an RRR interval, i.e. the ECG signal from two successive heart beats, marked by three successive R peaks. We remark that RRR intervals can be easily detected and contain enough information for pattern recognition. The patterns of the RRR intervals can be represented as vectors lying in relatively high dimensional spaces. However, due to the variation of the ECG waveform, to variations in heart rate and to the presence of noise, vectors may be very different from each other and actually lie in different dimensional spaces. This makes it difficult to train a neural network for beat recognition. The proposed strategy is that of considering the ECG data from an RRR interval as a noisy sample of an underlying function. The function is approximated by means of a linear combination of a fixed number of suitable basis functions. The coefficients of the linear expansion can be easily computed by solving a linear least-squares problem, and constitute the extracted features of the RRR interval. The transformed patterns, i.e. the coefficients of the linear expansions, become the input signals of a feed-forward neural network classifier, which provides an output of zero for the normal and one for the ischemic case. The adopted technique merges all the patterns to be classified in the same lower dimensional space and reduces the influence of ECG morphology and noise.

In Section 2 we describe the approximation technique of RRR interval and the neural network ECG beat classifier. In Section 3 we present the results of the tests that were carried out in order to evaluate the designed system as beat classifier. Finally, Section 4 contains some concluding remarks.