Robust multiclass ECG arrhythmia detection using balanced trained neural network

Classification of data including Electrocardiogram (ECG) is significantly affected with imbalanced class distribution. Efforts have been made to learn from the imbalanced data to improve the performance of the classifier. This paper presents the result of ECG arrhythmia detection of normal(N) beat and five different arrhythmia beats based on balanced input to the classifier for training only. The ECG signals are taken from MIT-BIH arrhythmia database. A three layer feedforward Backpropagation neural network(FFBNN) is used for classification. The input for training the network is perfectly balanced, taking same number of patterns from each class. After training the trained network is used for classifying completely different dataset. We achieved average accuracy, sensitivity, & specificity of 99.24%, 94.90% & 99.57% respectively. The effect of balanced input at the time of training is shown.