Near-Brain Computation: Embedding P300-based BCIs at EEG headset level

This paper presents a first-of-a-kind BCI framework overcoming technological and environmental factors that limit the adaptability of EEG-based BCIs to everyday life contexts. Some examples of these limitations include discomfort associated with the use of EEG headsets that require conductive gel, the lack of "plug and play" solutions, and noisy environments. In this context, the proposed BCI framework aims to realize an integrated system, currently running on the STM32L4 embedded platform (oriented towards headset-level implementation), capable of: (i) analyzing data from 8 dry EEG electrodes, (ii) detecting and correcting spatially uncorrelated deflections in EEG channels caused by external disturbances, and (iii) identifying and fine-tuning hyperparameters of a Fully Connected Neural Network with a user data-driven approach. The implemented embedded system, applied for a feasibility study to a 12-choice P300 speller matrix, demonstrated matrix element recognition accuracy exceeding 80% after only 4 runs, while maintaining an information transfer rate (ITR) of approximately 16 commands per minute under non-optimal usage conditions.