Brain networks stop sharing one parcellation: IC-GNN predicts autism, Alzheimer's and depression from each person's own community organization

Disease-prediction models now treat each person's own grouping of brain regions as an explicit structure rather than a partition shared by everyone. Widely used group-level atlases give a common, comparable reference but obscure individual differences in community boundaries and membership; voxel-level individualized parcellations capture that variability but often require extensive scan data, and differing parcel definitions complicate cross-subject comparison. On ABIDE (autism), ADNI (Alzheimer's disease) and the UK Biobank (depression), IC-GNN's mean AUROC and accuracy exceeded those of the 15 compared methods, including SVM, random forest, GCN, GAT, GIN, GraphSAGE and the brain-network models BrainNetCNN, BrainGNN, Com-BrainTF and CAGT.

The individual communities set their own number and composition from the data, and are then used to filter connections and adjust each region's predictive weight. Affinity propagation is a clustering method that does not require a preset number of clusters; here it groups regions at each level by similarity of functional-connectivity profiles into discrete communities and merges them into a three-level hierarchy. On ABIDE, removing the community hierarchy, replacing individual communities with random or consensus partitions, or removing the learned edge refinement all lowered performance; individual communities outperformed both the consensus hierarchy and size-matched random hierarchies.

These individual communities are reproducible within a scan and across sessions, and carry diagnostic information on their own. Splitting ABIDE time series into halves and rebuilding communities returned each participant's own communities as the closest match for 97.2% of participants, versus 3.8% for random communities; on ADNI, with sessions a median of 14 months apart, the own-session match rate was 71.9% versus 6.3%. Using only a community-affiliation profile built against the Yeo-17 reference systems, logistic regression reached AUROC 0.707 on ABIDE and 0.725 on ADNI on held-out test participants; adding the profile to a covariate model raised AUROC by 0.079 and 0.041 respectively.

Evaluation covers three binary tasks: autism in ABIDE, Alzheimer's disease in ADNI and depression in the UK Biobank, with all scans parcellated into the same 419 regions, 400 Schaefer cortical regions and 19 noncortical regions. ABIDE and ADNI use five-fold site-grouped cross-validation, and the UK Biobank uses a fixed held-out test set of 4,289 participants. Readers should treat this as disease classification and community analysis on resting-state functional connectivity, not a clinical diagnostic tool.

AUROC on the UK Biobank is 60.62%, clearly below ABIDE and ADNI, and the authors offer no explanation. Depression labels come from participants' self-reported lifetime professional diagnosis. The 618 participants with mild cognitive impairment in ADNI were excluded from classification and used only in community analyses. The authors note that voxel-level individualized parcellation is sensitive to fMRI data quality, and their strategy sidesteps that trade-off at atlas-region granularity.