Brain Network Dynamics from Neuronal Avalanches for Longitudinal Motor-Imagery Decoding

Brain–computer interfaces (BCIs) offer opportunities for personalized neurorehabilitation, but inter-session variability, repeated calibration, and the need for expert supervision limit their use beyond supervised settings. We propose a longitudinal BCI framework that integrates neuronal avalanches dynamics, reflecting cascades of propagation neural activity, with graph-theory and data-driven region of interest analysis to extract computationally efficient and neuro-physiological interpretable features. Source-reconstructed EEG from 20 BCI-naïve participants performing right-hand motor imagery (MI) and rest across four training sessions was represented by a directed weighted connectivity network based on neuronal avalanches. Local, mesoscopic, and global network measures revealed task-related changes predominantly in contralateral sensorimotor regions and distributed reorganization across training. We modelled here the learning-related evolution of feature distributions across previous sessions and used this trajectory to predict class-conditional distributions in an unseen fourth session, anticipating future brain states before new-session calibration. PageRank-based ROI selection achieved an AUC of 73.5 ± 7.9%, compared with 64.5 ± 10.6% using all ROIs, while reducing computational time by approximately 39% relative to performance-driven selection. The proposed framework provides interpretable, compact descriptions of avalanche propagation and exploits training-related changes for prospective decoding. Moving spatial selection and model optimization to an initial offline phase may reduce subsequent calibration demands and facilitate more autonomous BCI use.