EEG-based stroke severity classification using higher-order topological features and graph convolutional networks

Electroencephalography (EEG)-based stroke analysis has mainly relied on conventional signal and network descriptors, while higher-order brain network structures remain insufficiently characterized.

We used persistent homology to extract cycle-based topological features from EEG functional networks, capturing higher-order organization with reduced sensitivity to threshold selection. These features were integrated with conventional EEG representations and embedded into a graph convolutional network for stroke severity classification.

The proposed framework achieved 86% accuracy in discriminating mild from moderate stroke. Cycle ratio analysis further revealed that the prefrontal cortex exhibited the most prominent higher-order structures, indicating its prominent involvement in post-stroke brain network organization.

Our results suggest that higher-order topological features can enhance EEG-based stroke severity classification and offer additional insight into post-stroke brain network alterations.