Exploring Dependence of Subject-Specific Training Strategies for EEG Based Brain-Computer Interfaces

• Published in: Proceedings of IEEE Southeastcon pp. 848 - 853
• Main Authors: de Wit, T. Warren, Menon, Vineetha, Davis, Thomas
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.03.2024
• Subjects: autonomous systems; Brain modeling; brain-computer interface; Brain-computer interfaces; context fusion; customization; decision fusion; deep learning; deployment; disaster relief; Disasters; Drones; EEG; Electroencephalography; human-robot teaming; Object detection; subject-independence; Training
• DOI: 10.1109/SoutheastCon52093.2024.10500151

Design and deployment of brain-computer interface enabled assistive systems poses many practical questions, including the decision of whether to use a pretrained model or to customize a model to each unique end-user. In this work, we apply this question to the application domain of a passive brain- controlled drone for use in disaster relief and hostage rescue situations. A six-class intent (e.g. EEG) recognition experiment is performed with 42 subjects. This pilot study explores and evaluates the effects of subject-specific ("customized") versus subject- independent ("uncustomized") modeling approaches. Based on experimental validation, we present our discussions on the observed pros and cons of each training strategy. In this study, it was noted that the uncustomized training approach had the best target detection performance with a reduction in variance. Additionally, its deployment readiness attribute make it a more relevant and feasible option for our intended use case.