Effecient ECG Signal Classification Using Sparsely Connected Radial Basis Function Neural Network

Precise electrocardiogram (ECG) classification to diagnose patients’ heart condition is essential in getting maximum benefits via effective treatment. Artificial neural network is known to yield good results for classification of “difficult-to-diagnose” signals in medical domain. In this work, the ECG classification is performed using sparsely connected radial basis function neural network (RBFNN). Unlike the fully-connected RBFNN architecture, this type of structure reduces the computational cost and increases the classification accuracy. Only prominent features are necessary to a certain class of ECG during classification task. The ECG signal is obtained from Boston’s Beth Israel Hospital. The processes inc ude signal pre-processing, QRS complex detection, features extraction, determination of important parameter and construction of RBF network. Wavelet decomposition method was used for feature extraction process. Five features to be used are standard deviation of R-R distance, P-wave and R-wave amplitude, QS-wave distance and minimum point of T-wave. The signal is classified into three classes; normal sinus rhythm, malignant ventricular ectopy and atrial fibrillation. The results show that sensitivity value is between 75 to 100 percent while the certainty value is between 91.7 to 100 percent. In addition, this work shows that conventional RBFNN and linear discriminant analysis (LDA) produce less accurate results in terms of overall classification accuracy.