Noninvasive Brain-Computer Interfaces Based on Sensorimotor Neurodynamics

Non-invasive brain–computer interfaces (BCIs) establish a direct channel between neural activity and external devices, enabling interaction without muscular output. This thesis investigates BCIs based on sensorimotor neurodynamics, focusing on signals recorded with EEG and MEG and on the methods that allow their transformation into reliable control commands. The thesis reviews the main stages of the BCI pipeline, from signal acquisition to preprocessing, feature extraction, and classification, highlighting how these steps shape neural data into actionable outputs. Within this framework, two complementary strategies for feature extraction are examined: event-related desynchronization (ERD), which captures spectral changes linked to motor imagery, and fractal dimension (FD), a nonlinear measure of signal complexity. In addition, artificial intelligence methods, including machine learning and deep learning, are explored for their capacity to enhance classification accuracy and adaptability. Two case studies illustrate these aspects: the first analyses ERD as a spectral marker of motor imagery, while the second applies FD to subject-independent classification. Together they show how linear features such as ERD and nonlinear measures such as FD can be combined to capture different aspects of sensorimotor activity. ERD remains a reliable and established feature, whereas FD offers promising improvements in generalization across subjects. In summary, this thesis highlights how combining sensorimotor characteristics with modern artificial intelligence techniques represents a promising direction for creating more reliable and robust non-invasive BCIs