EEG-Based Movement Decoding in Motor-Impaired Patients by Extracting and Aligning Neural Patterns with Healthy Individuals

Decoding human movement intentions from electroencephalography (EEG) signals is critical for brain-computer interface (BCI) applications in motor neurorehabilitation, active assistance, and functional augmentation. However, current BCI models face two challenges for motor-impaired patients: 1) prolonged EEG data collection from patients is difficult 2) differences in brain functional structures and motor behaviors between healthy individuals and patients limit the generalizability of models trained on healthy individuals' EEG data. To address these challenges, this study proposes a transfer learning-based model, TL-ME, to bridge the gap between healthy individuals' and patients' EEG data and improve movement decoding accuracy for patients. TL-ME integrates an attention-based feature extractor, adversarial domain discriminator, multi-source selection, and movement classifier to transfer knowledge from healthy individuals' EEG data (source domain) to patients' EEG data (target domain). Temporal and spectral visualizations are used to inspect brain activation patterns for shared motor tasks between healthy individuals and patients. Experimental results show a 10.8% improvement in upper-limb movement decoding's accuracy using TL-ME, with each module contributing to performance gains. Visualization analyses also demonstrate similar brain activation patterns across domains, validating the transferability of healthy individuals' EEG data to patient-specific models. This work introduces a novel cross-population transfer learning approach that leverages healthy individuals' EEG data to enhance neural decoding for motor-impaired patients, bridging the gap between experimental studies and real-world applications in BCI-based neurorehabilitation.