Neural networks for the classification of heart rate variability data

Artificial neural networks are computational tools utilizing a combination of many elementary processing units. An important quality of neural networks is that when they are correctly trained, neural networks can appropriately process data that have not been used for training. This paper presents a neural network based approach to classification of supine vs. standing postures and normal vs. abnormal cases using heart rate variability data. Heart rate variability is concerned with the analysis of intervals between heartbeats. Loss of normal autonomic nervous system control of heart rate and rhythm is an important risk factor for adverse cardiovascular events. We have chosen 10 features for the network inputs. The features we have chosen are standard deviation of successive differences (SDSD), standard deviation of RR intervals (SDRR), standard deviations SD1 and SD2 obtained from Poincare plot, fractal dimension (FD), complexity measure (CM) and powers in ultra low frequency, very low frequency, low frequency and high frequency bands. Four classification algorithms have been compared viz., K-Nearest Neighbor Classifier (KNNC), Radial Basis Function (RBF) networks, Support Vector Machines (SVMs) and Back Propagation Networks or Multilayer Perceptrons (MLP) for different set of features. The obtained accuracies for the Back propagation network, RBF network units and Support vector machine with RBF kernel are higher than for K-nearest neighbor classifier. We are able to correctly classify 106 out of 116 cases corresponding to normal and abnormal subjects and in case of supine and standing subjects we are able to classify 105 out of 116 correctly.