Potentiality of Entropy for Semantic Concept Differentiation in EEG Signals in Alpha and Beta Waves

Semantic feature extraction is a novel field in neurotechnology research and is quickly growing in interest for its many applications, ranging from education to rehabilitation to BCIs. With few studies existing on the subject, a wide range of possibilities open for exploration. While most of the current research landscape explored the implementation of time-domain and frequency-domain features, this study proposes an approach based on entropy measures. exploring their potentiality for differentiation between concepts. This study focused on the analysis of entropy measures computed on EEG signals as, while offering less spatial resolution compared to other brain signal acquisition technologies, it best suits real-life application thanks to its portability, low cost, and current developments for this purpose. Entropy was chosen for this purpose due to its fundaments in the information theory, potentially bypassing other features limitations. For this study EEG signals have been divided based on the concept used for stimulation through different methods and paradigms (pictorial, orthographic and auditory comprehension, repeated for perception and imagination tasks) to limit the influence on the results of the modality-related processing pathways in brain activity and bring focus to the concepts. Different entropy measures, Shannon entropy, spectral entropy, sample entropy, permutation entropy, and multiscale entropy, have been calculated from the EEG signals in two bandwidths, alpha and beta, and the results across the electrodes were assessed through visual and statistical analysis. The visual analysis was performed with the help of 2d plots of each electrode’s entropy on the scalp and histograms. The statistical analysis consisted in an Analysis of Variance (ANOVA) including all three concepts (guitar, flower, and penguin) and t-tests performed on all pairs of concepts. The results showed a tendency suggesting Shannon, sample and multiscale entropies could better perform in distinguishing concepts, especially in alpha bandwidth. Sample entropy showed the best results. Multiscale entropy results could suggest how certain time windows at different time scales can hold more information than others, also highlighting how these time windows can vary across subjects and trials. This study can serve as a first step towards the creation of new methods for performing semantic feature extraction using algorithms specifically tailored on the different entropy features.