Towards brain-computer interfaces to personalise rhythmic interventions in human motion analysis

Music therapy, through the rhythmical entrainment of physical movements, shows promise in motor rehabilitation in both healthy individuals and those with motor impairments arising from neurodegenerative or neurodevelopmental disorders. Despite progression in this field, a key challenge persists: understanding how to personalise the musical stimuli to effectively improve both gait and mood. Neurocomputational models describe how music entrains brain networks and shapes musical perception, but fail to account for behavioural characteristics such as affective states or motor outputs. Brain Computer Interfaces (BCIs) can bridge this gap by detecting and monitoring gait adaptation, revealing how brain networks interact during motor adjustments and highlighting abnormalities in cognitive function and motor control across neurological conditions. We employ Regularised Common Spatial Patterns (RCSP) and Riemannian geometry to enable robust detection of mental states from Electroencephalography (EEG) signals during gait adaptation while minimising movement artefacts. RCSP effectively improve class separability on small and noisy datasets while Riemannian geometry preserve the inherent geometric structures of covariance matrices that directly reflect functional brain connectivity. We demonstrated improved performances on two dataset splits - ‘Within’ and ‘Cross-subject’, with the integration of a hybrid architecture, fusing RCSP and Riemannian features in parallel and sequential configurations. Feature selection methods are applied to identify important spatial and temporal features, offering physiological insights into task-related neural activity, while also reducing computational demand. We propose Individual Tangent Space Alignment (ITSA), a novel pre-alignment strategy to address inter-subject and cross-domain variability in EEG-based BCI. Using leave-one-subject-out cross-validation, our ITSA-RCSP-Riemannian integration demonstrates significant performance improvements, with robust performance maintained across subjects, varying data conditions and density configurations. We explore interpretable analysis by characterising movement intention through cognitive-motor process mapping. EEG and EMG recordings are analysed alongside lower joint kinematics, visualising temporal relationships between movement-related cortical potentials (MRCPs), muscle activation and limb movements. Accurate movement intention detection that mitigates cross-subject and cross-domain variability can deepen our understanding of how musical stimuli modulates neural responses driving motor function and affective states, enhancing practical, explainable music-based BCIs.