Abstract Body (Do not enter title and authors here): Background: Reliable EEG biomarkers of brain-network engagement could personalize action-observation (AO) therapy after stroke.
Objective: Identify functional-connectivity (FC) metrics that most robustly decode AO stimuli.
Methods: Five right-handed adults (21-29 y) viewed 120 video trials (robot/human limb actions + controls) while 32-channel EEG was recorded. Ten central-region channels were filtered (alpha and beta bands) and FC matrices (10×10) computed using coherence (COH), imaginary coherence (iCOH), phase-locking value, partial directed coherence (PDC) and spectral Granger causality (SpcG). A graph neural network (GNN) was trained with stratified 15-fold cross validation for two tasks: (1) AO-condition decoding: six classes -- human-left, human-right, robot-left, robot-right, baseline, and landscape; (2) Action-type decoding: five upper-limb actions -- air punch, back-and-forth arm swing, lateral arm swing, overhead arm raise, and wave.
Results: iCOH achieved the highest performance across both tasks (macro-AUC 0.997 & 1.000; balanced accuracy 0.96 - 1.00). Directed metrics PDC and SpcG also performed strongly (macro-AUC ≥ 0.99). Findings persisted despite class imbalance and small sample size.
Conclusions: Volume-conduction-invariant (iCOH) and directed FC measures provide robust signatures of motor-and-cognitive-network engagement during AO. These EEG markers may inform adaptive AO therapy or BCI-guided rehabilitation post-stroke. Larger cohorts will validate clinical utility.