Language-related functional connectivity in post-stroke aphasia: preliminary findings from a graph-theoretical and...

To have an insight into language-related functional connectivity in post-stroke aphasia (PSA) from graph theory measurements when performing an ability-matched auditory-verbal task fMRI.

Fifty-seven PSA patients were stratified into high-level (n = 22) and low-level (n = 35) groups using an ability-matched auditory-verbal fMRI paradigm. Functional connectivity was modeled via ROI-to-ROI generalized psychophysiological interactions, from which graph metrics for predefined language nodes were extracted. Network measure differences were assessed via ANCOVA, followed by binary classification with nested cross-validation. Performance (accuracy, sensitivity, specificity, AUC) and model interpretability (SHAP) were evaluated for the best-performing model.

Random Forest classification reached a significant AUC of 0.671 (p = 0.035, 95%CI [0.512, 0.816]) and an accuracy of 0.667, outperforming other models in analyzing task-embedded resting-state data. Notably, the model demonstrated high sensitivity (0.800) in identifying task levels. SHAP analysis revealed that the left temporo-occipital inferior temporal gyrus (toITG_L) and the right posterior supramarginal gyrus (pSMG_R) were the most influential predictors. High-level task was characterized by increased Local Efficiency in the bilateral pSMG and decreased Global Efficiency in the toITG_L.

Our findings suggest that the high-level group relies on a synergistic interaction between the ventral stream (toITG) and the dorsal stream (pSMG). The shift toward increased local specialization, particularly the compensatory recruitment of the right pSMG, highlights a critical neural modularity strategy for functional recovery. These results suggest the feasibility of integrating graph metrics with interpretable machine learning, offering preliminary insights that could support the development of objective tools for monitoring aphasia rehabilitation.