RR Stress Test Time Series classification using Neural networks

The RR time series, obtained from the R waves of the ECG, are a representation of the heart rate. This work presents the use of an artificial neural network (ANN) to classify RR time series from an ECG stress test. Four classes of RR time series were defined: very good, good, low quality and useless. We use a preprocessing stage to split input data vectors into N_W data windows for which we compute the standard deviation of the RR interval (SD_RR) to generate the input features vector of a multilayer perceptron network architecture. We introduce a saturation value S in order to limit SD_RR values. 520 RR time series from 65 records of ECG stress test were analyzed. Experiments were performed to explore the influence of parameters S and N_W. 40 subjects records are used in training and the remaining for testing. The classification results show a matching correlation ratio above 71%, which is higher than the correlation between two human experts. The main contribution of this work constitutes the preprocessing stage proposed for a stress test RR time series schema and an acceptable performance which does not depend on parameter N_W.