Application Of Neural Generated Error Signal In Classification And Compression Of Ecg Beats

Abstract

This paper presents a neural network architecture connected in cascade for ECG Holter system beat classification and compression. The system works in real time, on line, and is capable of recognizing and compressing up to 40 artificial and real QRS templates. The parallelism of neural network applied in this paper increases the efficiency of computations. An error signal derived from difference between predictor and testing signal is used in the classification and compression. This error signal is processed in two ways; firstly it is encoded with Hoffrnan sequence and saved as the compressed signal that may be used later for reconstructing the original ECG signal. Secondly the error signal is converted into primitive ternary signal and used in sharp and direct classification of the ECG beats. The proposed architecture consists of the following: a neural network used to generate linear predictions for signals. Another neural network generates error signals by calculating difference between predictions taken from first neural network and a testing signal. A third neural network does the classifications utilizing the error signals instead of complex raw signal. At first, Hermite functions are used in generating artificial ECG signals. Testing is done by adding noise to the original Hermite functions used in learning. The whole process is repeated for real ECG beats taken from MIT database. The results show clear success in classification besides additional success in compression.