Modelling and Prediction of Pain Related Neural Firings Using Deep Learning

Abstract

We propose a deep learning approach to model and predict pain related neural firings from EEG data. In particular, we target for the first time differentiation between acute and chronic pain. Our modelling strategy followed three steps: 1) Feature extraction of EEG data using Petrosian Fractal Dimension (PFD) and Hjorth activity functions. 2) Source localization of neural firings to differentiate between acute and chronic pain. 3) Modelling and training of a deep learning model for the prediction of the related pain according to the feature extracted neural firings. Based on our results, an occipital brain activation for chronic pain and a temporal activation in the case of acute pain were recognized. Moreover, our long short-term memory (LSTM) based prediction model achieved an accuracy of 91.29% for identification of related pain. The performance of the model was evaluated using precision, recall and F1 scores. For acute pain it achieved scores of 0.90, 0.82, 0.86 and for chronic pain scores of 0.86, 0.93, 0.89 respectively. It is concluded that our approach not only shows better predictive accuracy than the results reported by previous studies, but also represents an important step towards identifying and evaluating pain when patients are incapable of self-reporting it or when the clinical observations are unobtainable or unreliable.