Electroencephalogram-Based Pain Classification Using Artificial Neural Networks

This study investigates the variations in electroencephalogram (EEG) signals due to pain stimuli and proposes an optimal network configuration of the multilayer perceptron neural network (MLPNN) for pain state detection. EEG signals were recorded from 39 volunteers under the normal resting state and by applying external pain stimuli. Time, frequency, and wavelet domain parameters were computed and analysed. Decrease in Hjorth mobility, relative alpha power, minima of approximation coefficients (a(5)), mean and median frequency; increase in Hjorth complexity, root mean square value, relative delta power along with standard deviation, and maxima of approximation coefficients (a(5)) were observed at all the electrode positions. Several combinations of backpropagation algorithms and error functions were investigated to find the optimal configuration of MLPNN. We had classified pain state with an accuracy of 87.53%, 90.25%, 93.34%, and 90.62% in FP1, FP2, P-3, and P-4 electrode positions, respectively.