Electroencephalogram based emotion recognition in Parkinson’s disease using non-linear methods
MetadataShow full item record
In addition to classic motor signs and symptoms, individuals with Parkinson's disease (PD) are characterized by emotional impairments. Electroencephalogram (EEG) signals, being an activity of the central nervous system, reflect the underlying true emotional state of a person. This research focuses on analyzing different non-linear algorithms to recognize emotional states in Parkinson’s disease (PD) patients compared to healthy controls (HC) participants using EEG signals. Twenty non-demented PD patients and 20 healthy age-, gender-, and education level-matched controls viewed happiness, sadness, fear, anger, surprise, and disgust using multimodal stimulus (combination of audio and visual) while 14-channel wireless EEG was being recorded. In addition, participants were asked to report their subjective affect. The acquired EEG signals were preprocessed using thresholding method to remove eye blinks/movement artifacts. A Butterworth 6th order bandpass filter was used to extract the following EEG frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–49 Hz). To classify the emotional states and visualize the changes of emotional states over time at single-electrode level, four kinds of feature extraction methods (namely higher order spectra (HOS), non-linear analysis, fast Fourier transform and wavelet packet transform) were compared, and proposed an approach to visualize the trajectory of emotion changes with manifold learning. Three connectivity indices, including correlation, coherence, and phase synchronization index (PSI) were extracted by focusing on electrode pairs to estimate brain functional connectivity in EEG signals. New feature, namely, bispectrum based phase synchronization index (bPSI) was proposed for computing EEG functional connectivity patterns with the traditional methods. The statistical significance of all the computed features was studied using Analysis of Variance (ANOVA) test. Four different classifiers namely Fuzzy K- Nearest Neighbor (FKNN), K-Nearest Neighbor (KNN), Regression Tree (RT), and Support Vector Machine (SVM) were used to investigate the performance of the extracted features. Ten-fold cross-validation method was used for testing the reliability of the classifier results. The features extracted in all the methods were found to be statically significant (p < 0.05). The HOS based feature across ALL frequency bands (combination of five bands) performed well in recognizing emotional states of PD patients and HC participants with an averaged recognition rate of 77.43% ± 1.59% and 83.04% ± 1.87% respectively. The PD patients showed emotional impairments as demonstrated by a lower classification performance, particularly for negative emotions (sadness, fear, anger and disgust). The emotion-specific feature was mainly related to high frequency band (alpha, beta and gamma) than low frequency band (delta and theta). The trajectory of emotion changes was drawn by a manifold learning model. Also, bPSI functional connectivity index performed better with an averaged recognition rate of 51.66% ± 1.02% and 71.79% ± 1.01% for PD patients and HC respectively.