Emotion recognition using EEG phase space dynamics and Poincare intersections

• Published in: Biomedical signal processing and control Vol. 59; p. 101918
• Main Authors: Zangeneh Soroush, Morteza, Maghooli, Keivan, Setarehdan, Seyed Kamaledin, Nasrabadi, Ali Motie
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2020
• Subjects: Computational neuroscience; Electroencephalogram; Emotion classification; Phase space reconstruction; Poincare intersections; Electroencephalogram; Phase space reconstruction; Computational neuroscience; Emotion classification; Poincare intersections
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2020.101918

•We managed to classify emotions at a high classification accuracy.•A new state space called angle space was introduced for nonlinear EEG processing.•A new method based on Poincare planes was proposed to quantify AP.•Optimum Poincare planes were determined and different evaluation scenarios were considered.•The proposed method could be applied in other real-world applications. Emotions play a crucial role in our daily life. Emotion recognition has been used in numerous areas such as education, rehabilitation, etc. Simple to record and cost-effective, Electroencephalogram (EEG)-based emotion classification has been attracting a great deal of attention so far. Since our feelings are controlled by our brain which is inherently complex, it is imperative to employ nonlinear methods especially EEG phase space as it contains valuable information about EEG dynamics. In this study, EEG phase space is reconstructed and transformed into a new state space. Poincare planes are utilized to describe the proposed state space mathematically. They quantify EEG dynamics. Poincare intersections are extracted as features and then fed to the classification models including multi-layer perceptron (MLP), k-nearest neighbor (KNN) and multi-class support vector machine (MSVM). Variable and constant number of Poincare planes are considered and three different approaches are taken to determine optimum planes. A very reliable database is used and different aspects are considered to test the proposed method fairly. We employ three different evaluation scenarios including leave-one-subject-out, leave-one-trial-out and ten-fold cross validation and the recognition rates for all the scenarios are above 70% which is comparable to the previous studies. Not only is the proposed method effective in emotion recognition but it also introduces a novel approach to nonlinear signal processing which can also be employed in other applications and describe signals’ complex dynamics appropriately.