A Graph-Based Dynamical Characterization and Inference in Hybrid BCIs

The accurate characterization of the underlying dynamics of neural responses across various neuroimaging modalities is pivotal to enhancing the performance of hybrid brain-computer interfaces (hBCIs). While unimodal electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) neuroimaging modalities for motor imagery (MI) have been successfully decoded for BCI applications, to date, conventional feature extraction schemes have not yet provided satisfactory performance improvements. In this context, a paramount research question is whether the conventional EEG spectral features are sufficient to express the underlying MI neural dynamics fully, maximizing the synergy with fNIRS characteristics. In this study, we explore a graphbased feature extraction method that integrates the nonlinear dynamics of MI-based EEG responses and its contributions to hBCIs for amyotrophic lateral sclerosis patients (ALS) who need these systems for communication and control. Our results demonstrated that while there is a slight performance improvement using EEG nonlinear features, there is a substantial increase (~15%) in the contribution of EEG features to the total number of selected fused features indicating better synergy with fNIRS features at a multimodal level, when compared with the classical features. The extracted graph-based features can add a new informative dimension for an efficient integration with MI-based fNIRS responses.