Related Experiment Videos
Cross-Modal Masked Connectome Transformer for multimodal EEG-fMRI learning in ADHD diagnosis
Jiujiang Guo1, Monan Wang1, Xiaojing Guo2
1School of Mechanical and Power Engineering, Harbin University of Science and Technology, Harbin, 150080, China.
Background:
Attention Deficit Hyperactivity Disorder (ADHD) is associated with distributed alterations in large-scale brain systems, making reliable diagnosis from neuroimaging data challenging. EEG-derived and fMRI-derived static functional-connectivity profiles summarize patterns of electrophysiological coupling and hemodynamic covariation, respectively, but their heterogeneous signal mechanisms make direct multimodal alignment difficult.
New Method:
We propose a Cross-Modal Masked Connectome Transformer (CM-MCT) for EEG-fMRI based ADHD diagnosis. Instead of imposing direct correspondence between individual connections, CM-MCT treats each Desikan-Killiany 68-region cortical atlas (DK-68) region of interest (ROI) as a connectivity-profile token and learns to recover masked ROI-level connectome information within and across modalities. The model contains modality-specific Transformer encoders, cross-modal masked connectome objectives, and a bidirectional cross-attention fusion module for diagnostic prediction. Cross-fitted mask perturbations are used to characterize ROI-level model dependence, with explicit assessment of associations with nodal strength and perturbation magnitude.
Results:
Experiments using repeated five-by-five stratified evaluation of 210 paired HBN subjects demonstrate that the proposed method achieves 89.7% accuracy (95% confidence interval: 85.7%-93.3%) and 94.3% AUC (95% confidence interval: 91.5%-96.8%).
Comparison With Existing Methods:
Compared with the evaluated methods under matched splits, preprocessing, validation, and model-selection rules, CM-MCT shows the highest mean performance, although the AUC difference from the strongest baseline does not reach statistical significance after multiplicity correction.
Conclusions:
CM-MCT supports multimodal learning from the complementary network-level information provided by static EEG and fMRI connectivity profiles. Perturbation analysis yields modality-specific ROI sensitivity patterns that are modestly associated with the assessed factors. These heatmaps characterize model dependence without establishing causal involvement or identifying ADHD biomarkers.