Résumé
Recurrence Quantification Analysis (RQA) provides nonlinear indices that characterize the temporal structure of complex signals. Despite its increasing use in EEG research, its application remains limited by heterogeneous and often insufficiently justified parameter choices, hindering reproducibility and physiological interpretability. Using a public eyes-open/eyes-closed (EO/EC) resting-state EEG dataset, we evaluated the influence of embedding dimension, time delay, recurrence radius and minimum diagonal line length. Two radius strategies were compared: a data-driven R-optimal approach and a fixed-density Recurrence Rate (RR-fixed) method. Across all electrodes and parameter sets, determinism (DET) consistently increased during EC, predominantly over occipital regions. Multichannel RQA DET feature was further assessed using an RBF-kernel support vector machine (SVM) under strictly nested leave-one-subject-out validation. Balanced accuracies exceeded chance in 16/18 parameters configuration, with no advantage for higher-dimensional embeddings relative to no embedding. To relate nonlinear dynamics to established physiological markers,DET measure was compared with spectral features. Alpha-band power topographies reproduced the canonical ECrelated posterior increase, and principal component analyses demonstrated that RQA metrics captured variance largely distinct from band-limited power. Taken together, the results show that a parsimonious configuration, no embedding, an adaptive radius, and an appropriately chosen minimal diagonal line, preserves statistical sensitivity and classification performance while remaining compatible with spectral interpretations.