Improving Motor Imagery Classification for Stroke Rehabilitation Using a Stroke-Aware Capsule Attention Network

Motor imagery (MI) monitoring is crucial for targeted stroke rehabilitation. MI involves mental simulation of motor actions, engaging areas like the motor cortex. Electroencephalography (EEG) provides valuable insights due to its high temporal resolution. However, EEG variability, non-linear impacts, subtle pattern detection, and compensatory mechanisms challenge MI classification. To address these, we propose the stroke-aware capsule attention network (SACAN) for improved cross-subject classification of EEG in stroke patients. SACAN integrates advanced features, including phase-amplitude coupling, permutation entropy, phase locking value, and imaginary coherence, along with custom neural layers like squeeze-and-excitation, gated linear unit, and capsule layers. This design enhances pattern recognition and generalization. Evaluated on an open-source dataset of 50 stroke patients, SACAN achieved 94.44% accuracy, demonstrating its effectiveness in MI classification and supporting better rehabilitation planning.