Adoption of Machine Learning Techniques to Enhance Classification Performance in Reactive Brain-Computer Interfaces

Abstract

This paper proposes the adoption of an innovative algorithm to enhance the performance of highly wearable, reactive Brain-Computer Interfaces (BCIs), which exploit the Steady-State Visually Evoked Potential (SSVEP) paradigm. In particular, a combined time-domain/frequency-domain processing is performed in order to reduce the number of features of the brain signals acquired. Successively, these features are classified by means of an Artificial Neural Network (ANN) with a learnable activation function. In this way, the user intention can be translated into commands for external devices. The proposed algorithm was initially tested on a benchmark data set, composed by 35 subjects and 40 simultaneous flickering stimuli, obtaining performance comparable with the state of the art. Successively, the algorithm was also applied to a data set realized with highly wearable BCI equipment. In particular, (i) Augmented Reality (AR) smart glasses were used to generate the flickering stimuli necessary to the SSVEPs elicitation, and (ii) a single-channel EEG acquisition was conducted for each volunteer. The obtained results showed that the proposed strategy provides a significant enhancement in SSVEPs classification with respect to other state-of-the-art algorithms. This can contribute to improve reliability and usability of brain computer interfaces, thus favoring the adoption of this technology also in daily-life applications.