Stability and neurophysiological validity of graph connectivity features for non-stationary motor imagery BCIs

OBJECTIVE: Motor imagery (MI) Electroencephalography (EEG) Brain-computer interfaces (BCI) degrade under longitudinal non-stationarity, especially in amyotrophic lateral sclerosis (ALS). Functional connectivity (FC) has been proposed as an alternative feature space, but it remains unclear which FC estimators yield stable, class-informative features across sessions. Approach: Using a multi-session ALS EEG dataset, we computed a broad family of FC estimators per trial to form weighted graphs. We extracted edge weights and node strength features, and quantified (i) feature reproducibility and (ii) LH-RH separability using coefficient of variation and symmetric Kullback-Leibler divergence, respectively. We assessed neurophysiological plausibility via spatial topographies, distance-dependence controls, and evaluated selected feature sets in a strictly temporal cross-session decoding protocol against Common Spatial Patterns, Band Power and Riemannian Methods. Main Results: Coherence-based estimators, particularly magnitude-squared coherence, most consistently produced features exhibiting favourable reproducibility-separability trade-offs across subjects. Node-strength discriminability maps showed lateralised sensorimotor structure consistent with known MI physiology. In temporal generalisation, Magnitude Squared Coherence derived features achieved more consistent test performance than baseline methods for most subjects. Significance: Joint reproducibility-separability profiling provides a principled way to select FC feature spaces for longitudinal MI-BCIs and suggests coherence-based connectivity is a stronger sensor-space candidate under drift.