Decoding Emotional Responses: A Comparative Study of fNIRS and EEG Neuroimaging Techniques

Abstract

This master's thesis explores the prediction of valence and arousal scores, key dimensions of emotional states, from functional Near-Infrared Spectroscopy (fNIRS) and Electroencephalography (EEG) data. The primary focus is on improving prediction accuracy by addressing the hemodynamic response function (HRF) in fNIRS and harnessing the capabilities of advanced machine learning models. The study investigates whether temporal offsets to account for the HRF, combined with selective window sizing for data analysis, enhance the prediction of emotional states. Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs) are employed to exploit the spatial and temporal features inherent in EEG and fNIRS data. This research includes two key components: (1) implementing offsets to align fNIRS data with the HRF, and (2) determining optimal window sizes to capture relevant regional brain activity. The effectiveness of CNN and LSTM models is evaluated in predicting valence and arousal scores. Results indicate that adjusting for HRF and selecting appropriate window sizes significantly improves the predictive accuracy.