Abstract
Post-stroke aphasia involves dynamic disruptions in brain networks that are not fully captured by static analyses. This study employs dynamic amplitude of low-frequency fluctuations (dALFF) and dynamic functional connectivity (dFC) analyses combined with machine learning to comprehensively characterize these changes. Using a sliding-window approach, 22 key regions were identified based on dALFF alterations, and recurring dFC states were extracted via K-means clustering. Aphasia patients showed prolonged dwell times and fewer transitions in weak connectivity states, indicating reduced network flexibility. Network topology analysis revealed increased local and nodal efficiency alongside decreased global efficiency and modularity, suggesting a compensatory shift favoring local integration over global coordination. A support vector machine classifier, incorporating feature selection and leave-one-out cross-validation, effectively distinguished patients from controls with an AUC of 0.75, emphasizing critical dynamic temporal features associated with network dysfunction. These dynamic features correlated with language deficits and provided a more comprehensive temporal characterization than static metrics. Our findings reveal a dynamic balance between local compensation and impaired global integration in aphasia brain networks. Combining dynamic neuroimaging and machine learning offers a robust framework for capturing these alterations and supports the development of personalized rehabilitation strategies. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12880-025-02135-9.