Post-stroke aphasia analysis using topological alterations in brain functional networks

Abstract Objective . Nearly one-third of stroke patients develop aphasia. Although the function of classical language areas (e.g. Broca’s area, Wernicke’s area) has been widely characterized, the network reorganization mechanisms behind specific language dysfunctions induced by different aphasia subtypes and the biomarkers for a rapid and objective classification remain to be clarified. Additionally, the rapid classification of aphasia subtypes continues to be a clinical challenge. Approach . To address these gaps, we developed a diagnostic framework analyzing topological changes in resting-state fMRI-derived functional brain networks. A transparent feature selection pipeline is designed through combining the topological features, the ReliefF algorithm, the elbow method, and cross-validation to alleviate the limitation of available aphasia datasets. Main results . Using a cubic SVM classifier, the proposed model achieved 88.70% ± 1.37% accuracy and a 92.92% ± 0.78% F1 score in distinguishing post-stroke aphasia patients (PWA) from non-aphasic stroke patients patients without aphasia (PWOA) on a public dataset. Further validation on an in-house dataset (13 patients with PWA and 25 normal post-stroke patients) showed similar performance (88.1% accuracy, 92.76% F1 score), demonstrating robustness. Further functional connectivity analysis revealed PWA exhibit higher global/local network efficiency, increased clustering, and shorter path lengths than PWOA. Subtype analysis for Anomic, Broca, Conduction, and Global aphasia identified distinct neural patterns via one-way ANOVA, suggesting divergent pathophysiology. Significance . The proposed framework not only improves classification accuracy but also enhances interpretability and reproducibility. Thus, it could form the basis of a new objective diagnostic approach for aphasia.