Skip to main content
Back

POSTECH's MuHL Detects Alzheimer's at 93.2% by Rethinking Brain Connections

MuHL learns adaptive multi-scale hyperedges from brain scans, beating 19 baselines on Alzheimer's and Parkinson's classification while pinpointing disease-linked brain regions.

Paper, Full Access
  • MuHL introduces adaptive multi-scale hyperedge learning for brain network analysis, dynamically grouping multiple brain regions instead of relying on pairwise edges.
  • It uses learnable spectral graph wavelets to extract features at fine-to-coarse spatial resolutions, then learns a separate hypergraph structure at each scale.
  • On ADNI (Alzheimer's, 5-way classification), MuHL hits 93.2% accuracy — up 2.4pp over the next best method across 19 baselines.
  • On PPMI (Parkinson's), it gains 6.0pp in F1-score; zero-shot transfer to unseen cohorts also outperforms all baselines.
  • The learned hyperedges identify clinically meaningful brain regions (hippocampus, amygdala, thalamus) with statistically stable rankings across random seeds.
  • Accepted at ICML 2026; paper available on arXiv; code not yet publicly released.

Standard graph neural networks see the brain as a web of one-to-one connections: region A talks to region B, region C talks to region D. But neurodegenerative diseases like Alzheimer's and Parkinson's don't work that way. They emerge from coordinated dysfunction across groups of brain regions simultaneously, a pattern that pairwise edges simply cannot capture. A new paper from POSTECH introduces MuHL (Multi-scale Hyperedge Learning), a framework that replaces fixed pairwise edges with dynamically learned hyperedges , connections that link any number of brain regions at once , and does so across multiple spatial resolutions.

Why pairwise graphs fall short

In a standard graph, each edge connects exactly two nodes. For brain networks, nodes are regions of interest (ROIs) and edges encode structural or functional connectivity. The problem is that stacking graph convolution layers to approximate group interactions causes oversmoothing , node features become indistinguishable as depth increases. Prior hypergraph methods tried to fix this, but most either used predefined hyperedges (hand-crafted rules about which regions belong together) or only learned scalar weights on top of a fixed structure. Neither approach lets the model freely discover which groups of regions matter, or at what spatial scale.

MuHL attacks both limitations at once. It learns which regions form a group and at what resolution to look, all end-to-end from data.

The core idea: wavelets meet hypergraphs

The key insight is to use Spectral Graph Wavelet Transform (SGWT) , a technique that decomposes a graph signal into components at different spatial frequencies, analogous to how audio wavelets separate low rumbles from high-pitched tones. In brain terms, a fine-scale wavelet captures local neighborhood features for each ROI, while a coarse-scale wavelet blurs features across the whole brain, making distant regions look similar.