We aimed to build a model to estimate pleasant and unpleasant affective states. Affective states of pleasure or unpleasure were robustly induced in 60 participants (30 male and 30 female) by using a friendly or stressful speech task modified from the Trier Social Stress Test. A Support Vector Machine model was constructed using time-series EEG data measured from 12 electrodes on the frontal to temporal scalp sites during the speech tasks. We extracted features of EEG powers' time series in theta and alpha frequency bands in shorter segmentations (1st-order features), and fluctuations of time series of EEG power (2nd-order features). Overall, the model including the 2nd-order features showed higher accuracy of estimation of affective states, compared to the model with only 1st-order features (69.4vs. 62.6%). Since 12 participants showed large body movement which caused noise for EEG measurement, accuracy of estimation by the model was verified for all participants (N = 60) and for participants with little body movement (N = 48) separately. For participants with less body movement, right-side dominant EEG powers were showed in unpleasant affective states, consistently with previous findings. The model including the 2nd-order features showed higher accuracy in participants with less body movement, compared to accuracy in all participants (75.1% vs. 69.4%), suggesting validity of our features and model. Future work should develop more refined noise reduction techniques in analyses of EEG and examine generalizability of the model to various affective situations.
