Decoding Method Of Pain-evoked EEG Based On Autoencoder And Stacking Model

Pain is an unpleasant subjective feeling associated with real or potential tissue damage,and self-reporting is currently the gold standard for pain assessment in clinical practice.However,the reliability of self-reporting is low,and the self-assessment of pain cannot be performed in some populations(such as patients with unconsciousness and infants),which can lead to insufficient or poor treatment of pain.Therefore,an objective pain assessment method based on physiological signals is clinically necessary.Pain induces a series of electroencephalography(EEG)components,and the magnitudes of pain-related EEG components have a strong correlation with pain intensity.Therefore,EEG-based pain assessment has attracted more and more attention in recent years.However,traditional EEG-based pain prediction methods are mainly based on domain knowledge(known pain-evoked EEG components)to extract EEG features and to build a prediction model,which has two limitations.First,pain-evoked EEG is very weak and the individual differences are large,so it is difficult to accurately extract features based on domain knowledge solely.Second,existing domain knowledge may not cover all painrelated and pain-predictive EEG components.In recent years,with the rapid development of deep learning methods,more and more data-driven EEG decoding applications based on deep neural networks have been proposed.These applications can use EEG data more effectively to achieve better performance in EEG decoding.However,there are still few studies concerning the use of deep neural networks and other advanced machine learning techniques for EEGbased pain decoding.This study is aimed to apply autoencoder and ensemble learning to develop new and accurate pain prediction models based on pain-evoked EEG.First,we propose a new algorithm(named AE-LEPNet)that combines autoencoder(AE)and convolutional neural networks(CNN)to perform unsupervised feature extraction for single-trial laser-evoked EEG potential(LEP)induced by pain,and then uses support vector machines(SVM)to build a prediction model linking extracted features and pain scores.Specifically,AE-LEPNet uses a three-layer convolutional architecture to extract the multi-scale features of the pain-evoked EEG,then integrates them into the hidden layer preset by the autoencoder.Subsequently,the features in the hidden layer are used as the pain-related features for pain prediciton.The results show that the performance of AE-LEPNet is significantly better than the traditional method(which extracts pain-related LEP components,N2 and P2 amplitudes,and uses SVM for classification).Based on a pain-evoked EEG dataset of 28 healthy subjects,the performance of AE-LEPNet for classification of low pain and high pain was accuracy 77.7%,specificity 75.0%,sensitivity 80.5%,AUC: 0.77,all of which are higher than the results of the traditional method(accuracy 75.0%,sensitivity: 77.0%,specificity: 73.2%,AUC: 0.74).Further,we propose to use an ensemble learning technique,stacking,for EEG-based pain decoding.This stacking method can effectively combine domain knowledge and data-driven methods to increase prediction accuracy.Specifically,the stacking method learns different types of EEG features through two first-class classifiers(with N2 / P2 features and with AE-LEPNet-extracted features),and builds a meta-classifier to accurately decode pain ratings.The results show that the stacking method can obtain accuracy 78.6%,sensitivity 81.0%,specificity 76.3%,and AUC 0.78,which are higher than the traditional decoding method and AE-LEPNet.In the present study,two new pain decoding methods based on the autoencoder and the stacking model were developed to predict pain ratings from single trial pain-evoked EEG signals.This research overcomes the limitation of traditional pain decoding models which are heavily dependent on domain knowledge by developing more data-driven methods and further integrating domain knowledge and data-driven methods.Therefore,the proposed new methods provide a novel and effective means for pain prediction.This study can be potentially used for the development of more objective and accurate clinical pain assessment techniques.