Feature reduction and arrhythmia classification via hybrid multilayered perceptron network

Cardiac arrhythmias refer to abnormal electrical activity of the heart which results in irregular heartbeat. This paper proposes a computerized method in detecting cardiac arrhythmias using a structurally optimized hybrid multilayered perceptron (HMLP) network. ECG samples in bipolar limb lead orientations have been obtained from PTB Diagnostic ECG database for healthy, cardiomyopathy, as well as left and right bundle branch block signals. Data were initially processed for noise removal and baseline correction using finite impulse response (FIR) filters and a two-stage polynomial fitting technique. 24 morphological features were initially obtained from the sub-wave components of each lead via the median threshold method. The features are then analyzed through principal component analysis (PCA) and only 15 significant descriptors are used to optimize the network performance. A total of 1600 beat samples have been used to train, test and validate the HMLP network. Each network was trained using four types of learning algorithm. Results show that all network configurations attained more than 95% classification accuracies. In comparison, PCA-HMLP has shown better performance than the standard HMLP network.