Hybrid multilayered perceptron network for classification of bundle branch blocks

Abstract

Electrocardiogram is an electrical representation of heart activities that provide vital information on the cardiac condition. Development of reliable intelligent systems through analysis of cardiac rhythms has been paramount for automated classification of cardiac diseases. Bundle branch block is an arrhythmia caused by defects in the conduction pathways that alters the flow and speed of electrical impulses, leading to loss of cardiac output, and in severe cases, death. This paper proposes and investigates HMLP network for classification of bundle branch block arrhythmias. Samples of normal, right bundle branch block, and left bundle branch block beats were obtained from the PTB Diagnostic ECG database. Initially, the original signal underwent a filtering process and the baseline drift were rectified using the polynomial curve fitting technique. Five morphological features were then extracted through median threshold method for a total of 150 beat samples. The features were then used for training of the single hidden layer HMLP network. The training stage employed four different learning algorithms for four hidden node implementations. Results show that the Polak-Ribiere conjugate gradient algorithm achieved the best convergence speed with 100% classification accuracy. Overall, the various HMLP network structures managed to attain 99.6% average classification accuracy.