.png){kind=icon}
.png){kind=icon}
Return
Interpretable Dual-Stream EEG-MRI Fusion Uncovers Structure-Function Signatures of Stroke Motor Recovery
.png){kind=icon}
OA
AI
E
M
F
P
Y
C
G
Y
Y
DOI：10.1109/tnsre.2026.3717334.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Motor recovery prediction after stroke is hindered by the inability of single-modality imaging to capture how structural damage and functional reorganization interact. Existing multimodal approaches often treat electroencephalography (EEG) and magnetic resonance imaging (MRI) as independent signals, limiting both predictive performance and mechanistic interpretability. We introduce Dual-Stream Cross-Modal Fusion (DS-CMF), an interpretable deep learning framework that models bidirectional structure-function coupling through patient-adaptive integration of EEG and structural MRI. Applied to 26 individuals with subacute ischemic stroke performing motor imagery tasks, DS-CMF integrates structural MRI and high-density EEG recorded during key grip (KGMI), power grip (PGMI), wrist extension (WEMI), and wrist flexion (WFMI). The framework fuses task-based EEG connectivity features and structural MRI descriptors through region-aligned embedding, bidirectional cross-modal attention, and patient-adaptive gating. Feature-based attribution complements prediction to surface candidate structure-function markers. In internal cross-validation, DS-CMF achieved its strongest accuracies in PGMI (77.3%) and WEMI (73.3%), with the clearest within-task FDR-corrected improvements observed in PGMI and WEMI. Ablation analyses showed a modest overall advantage over early fusion, late fusion, and non-adaptive cross-attention variants, particularly in PGMI and WEMI, suggesting incremental value from adaptive bidirectional fusion. SHapley Additive exPlanations (SHAP) attribution identified candidate model-level features driving classification, including frontal theta and central-parietal beta connectivity patterns and structural contributions from sensorimotor and frontal regions. These attributions are consistent with known neurophysiology of motor recovery but reflect model behavior rather than validated biomarkers. Together, these results offer preliminary proof-of-concept that interpretable EEG-structural MRI fusion can capture classification-relevant structure-function patterns in subacute stroke.
Keywords:
Stroke recovery
motor assessment
EEG-MRI fusion
explainable AI
deep learning
Journal
IF:
5.2
Papers:
448
Citations:
1.6W
Organization
Cited Papers