Emotional state recognition using advanced machine learning techniques on EEG data Katerina Giannakaki School of Medicine University of Crete Vasilika Vouton, 70013 Iraklio, Greece e-mail: kgiannakaki@teemail.gr Giorgos Giannakakis Institute of Computer Science Foundation for Research and Technology Hellas Vasilika Vouton, 70013 Iraklio, Greece e-mail: ggian@ics.forth.gr Christina Farmaki Institute of Computer Science Foundation for Research and Technology Hellas Vasilika Vouton, 70013 Iraklio, Greece e-mail: xfarmakh@ics.forth.gr Vangelis Sakkalis Institute of Computer Science Foundation for Research and Technology Hellas Vasilika Vouton, 70013 Iraklio, Greece e-mail: sakkalis@ics.forth.gr Abstract—This study investigates the discrimination between calm, exciting positive and exciting negative emotional states using EEG signals. Towards this direction, a publicly available dataset from eNTERFACE Workshop 2006 was used having as stimuli emotionally evocative images. At first, EEG features were extracted based on literature review. Then, a computational framework is proposed using machine learning techniques, performing feature selection and classification into two at a time emotional states. The procedure described in this paper investigates and assess the effectiveness of selection and classification techniques providing improved classification accuracy. The proposed methodology is able to obtain accuracy of 75.12% in classifying the two emotional states comparing with similar studies using the same dataset. Keywords-emotion recognition; EEG; feature selection; classification; machine learning I. INTRODUCTION Human emotional states detection and their appropriate representation are of great interest lately, leading to the development of many applications towards this. There are various methods that can be employed for emotion recognition such as imaging techniques (fMRI, PET, etc.) [1, 2], biosignals (EEG, ECG, GSR, EMG, etc.) [3, 4], videos (facial expressions, body postures, gestures, etc.) [5-7]. Among them Electroencephalogram (EEG) is considered a reliable tool that it is used widely in clinical practice mainly for neurological dysfunctions [8, 9] but also for the examination of upper cognitive functions [10] and the recognition of affective states [4]. It provides high temporal resolution and in combination with its low cost and the fact that it is semi invasive, portable and not harmful for human health makes it appropriate for extensive use for emotion analysis. Lately, methods based on machine learning receive great attention as the recent advances lead to promising results for a wide range of applications in medicine including computer-aided diagnosis. Regarding psychophysiology, machine learning techniques have been employed to disease detection [11] as well as emotion state discrimination. Many studies of affect adopt the circumplex model [12] for the representation of emotions, where emotional states can be mapped using arousal and valence axes. Valence, which extends from sadness to joy, constitutes an index of perceiving an emotion as positive or negative while arousal, which extends from calm to excitement, reflects how strongly each feeling is perceived. The purpose of this paper is the investigation, evaluation and comparison of advanced feature selection and classification methods providing the most appropriate of them to address effectively the problem of discrimination between calm, exciting positive and exciting negative emotional states. II. DATA AND EXPERIMENTAL PROCCESS In this section, the experimental procedure, the protocol followed and the public available dataset used in this study is described. A. Dataset description The data that were analyzed in this study come from the public available eNTERFACE Workshop 2006 database (Project #7: "Emotion Detection in the Loop from Brain Signals and Facial Images") [13]. The dataset consists of EEG, functional Near-Infrared Spectroscopy (fNIRS), along with some peripheral biosignals (respiration, Galvanic Skin Response, blood volume pressure) and video recordings. The devices were synchronized using a trigger mechanism and the recordings were stored in Biosemi Data Format (BDF) files. Although the combination of signals from different modalities has been reported to be efficient in emotion recognition, the acquisition of many signals may interfere with user and would make a real life application unpractical. As the research interest focuses on EEG processing, only the EEG signals were used in this analysis.