Enabling Rapid Calibration of BCI Systems that Detect Movement-Related Cortical Potentials in Children with Cerebral Palsy
Brain-computer interface neurofeedback (BCI-NFT) has shown promise for neuromotor rehabilitation, but its clinical adoption -- particularly in pediatric populations -- remains limited in part by the lengthy calibration required before each therapy session. This study developed and evaluated a deep-learning framework to reduce calibration requirements for movement-related cortical potential (MRCP)-based movement-intention detection in children with cerebral palsy (CP). Electroencephalography (EEG) was collected during repeated ankle dorsiflexion tasks across 27 sessions in four children with CP. A bidirectional long short-term memory (Bi-LSTM) network was evaluated using seven training strategies, ranging from conventional within-session calibration to cumulative cross-subject learning and transfer learning. Cross-subject cumulative learning achieved 91\% accuracy without within-session calibration, while the addition of transfer learning increased accuracy to 93\% with minimal within-session calibration. Both approaches significantly outperformed conventional calibration strategies and achieved the highest F1-scores and receiver operating characteristic (ROC) performance, demonstrating robust generalization across sessions and participants. These findings show that cumulative learning and transfer learning can substantially reduce calibration requirements while maintaining high decoding performance, supporting the development of clinically practical MRCP-based pediatric BCI systems