Integrated gradients for enhanced interpretation of P3b-ERP classifiers trained with EEG-superlets in traditional and virtual environments

P3b event-related potentials are neural responses occurring around 300-500 ms after the presentation of a task-relevant target stimulus, usually evoked with the active visual oddball paradigm. Usually performed in laboratories, it can be moved to virtual environments, especially for subjects who cannot reach the physical experimental setting. However, the introduction of a virtual reality headset introduces significant portions of artefacts, since it is overlapped on top of the EEG headset. This can hamper the analysis of the p3b ERP and thus limit the understanding of attentional allocation, working memory operations, and decision-making processes in the brain of a subject. The research aims to compare the quality of the EEG data collected in virtual environment against that collected in the traditional environment. After segmenting such data into epochs around each stimulus, superlets are computed for each, specific time-frequency representations, and trained with a convolutional neural network for the binary discrimination of segments collected in virtual and traditional settings. Explainable Artificial Intelligence, namely the Integrated Gradients method, is employed to facilitate the understanding of such discrimination. An equivalence two one-sided tests (TOST) is performed to verify if both the types of input EEG segments are statistically similar. This research contribute to the body of knowledge by extending the application of the Integrated Gradients XAI method in a problem within Neuroscience.