Rules Extraction of Relevance Vector Machine for Predicting Negative Emotions from EEG Signals

Many studies have reported that patients who are experiencing long-term negative emotions have higher risk of having health deterioration. Therefore, recognition of negative emotions from Electroencephalography (EEG) signals is crucial for monitoring patient conditions. In EEG emotion recognition, clinicians tend to need a clear explanation regarding the rules behind the EEG emotion classification process. Most of the EEG emotions classification use Support Vector Machine (SVM) causing lack of probabilistic prediction which can trigger longer computation time. To address the limitation, we applied Relevance Vector Machine (RVM) with Bayesian inference algorithm to calculate the probability of predicted output. Similar to SVM, RVM was unable to provide transparent rules behind its classification. Therefore, this study attempts to extract rules from RVM by implementing Random Forest algorithm to the relevance vectors. We extract the average energy spectrum in each frequency band as the leading feature of emotions in EEG. Through the resulted rules of RVM_RF, we found that negative emotions of EEG were determined by the average energy spectrum of delta band at fronto-central electrodes (FCZ>23.683, FC4>24.812), theta band at frontal electrode (F5>23.683), and alpha band in parietal-occipital electrode (PO8>20.212). From the evaluation on three sessions of data measurement, it shows that the proposed approach of RVM_RF can predict the negative emotions of EEG with the higher average accuracy of 85.33% and average precision rate of 0.933 compared with other rule-based methods such as RVM_CN2 Rule, RVM_C4.5 Tree, and SVM. All in all, this proposed approach has demonstrated the possibility to identify negative emotions from EEG signal using rules extraction from sparse learning method, RVM.