A radial basis function neural network for the detection of abnormal intra-QRS potentials

Abnormal intra-QRS potentials (AIQP) in signal-averaged electrocardiograms have been proposed to be a potential noninvasive index for the diagnosis of the risk of ventricular arrhythmias. This study tries to develop a nonlinear neural network using radial basis functions (RBF) to approximate the normal QRS complex and then to estimate the AIQP using the approximation error, and further to quantify the estimation error of the AIQP. Different spread parameters of the Gaussian kernel function in the hidden layer have been adopted to evaluate the approximation accuracy of the RBF neural network. The study group of AIQP was constructed by adding a white noise with a root-mean-square value of 5 μV into the QRS complexes of the normal subjects to simulate the presence of AIQP. The study results illustrate that the mean root-mean-square values of the estimated AIQP in the AIQP group were 2.5 μV, 3.5 μV, 2.9 μV and 2.3 μV larger than those in the normal group using the spread parameters of 5, 10, 15 and 20, respectively. Hence the maximum accuracy of the proposed RBF neural network for the estimation of AIQP can reach 70% (3.5 μV compared to the ideal value of 5 μV).