BRAINet: A brain-region-aware interaction network for EEG-based diagnosis of disorders of consciousness.

Chen, Haoxiang; Zhao, Sha; Yu, Jie; Wang, Jiquan; Bai, Yumeng; Xu, Chuan; Li, Shijian; Luo, Benyan et al. · J Neural Eng · 2026

other

Where this comes from

Abstract


Reliable assessment and stratification of disorders of consciousness (DOC) is essential for patient care and clinical treatment planning. 
Electroencephalography (EEG) provides a non-invasive approach to measure neural activity and has shown promise in DOC assessment. 
However, most existing EEG-based approaches focus on binary UWS/MCS classification and often process EEG channels as a whole, without explicitly modeling anatomical brain-region organization. In this study, our goal is to distinguish among unresponsive wakefulness syndrome (UWS), minimally conscious state minus (MCS-), and minimally conscious state plus (MCS+) using resting-state EEG signals.

Approach.
We propose BRAINet, a brain-region-aware EEG framework for three-class DOC classification. BRAINet partitions EEG channels into five anatomical brain regions, extracts region-specific spatiotemporal and spectral features, models cross-region interactions using a Transformer-based attention module, and fuses the learned representations with approximate entropy features for final classification. We evaluated BRAINet on a clinical resting-state EEG dataset comprising 22 UWS, 24 MCS-, and 15 MCS+ patients using patient-wise five-fold cross-validation and comparisons with representative machine-learning and deep-learning baselines. Statistical comparisons were based on paired patient-level out-of-fold (OOF) predictions.

Main results.
BRAINet achieved the highest numerical performance among the compared methods. Across the five folds, its mean balanced accuracy was 54.07% at the epoch level and 59.89% at the subject level. Based on pooled patient-level OOF predictions, BRAINet achieved significantly higher balanced accuracy than the best-performing baseline, Conformer (two-sided paired permutation test, Holm-adjusted p=0.00513).

Significance.
These results suggest that brain-region-aware EEG modeling may provide useful information for fine-grained UWS/MCS-/MCS+ classification. BRAINet provides an interpretable framework for exploring region-specific EEG representations in DOC and may support future studies on patient stratification and prognostic assessment.