Differentiable Cell Complexes Discover Interpretable Metabolic Connectivity Subtypes in Alzheimer's Disease

Oankar R. Patil, Keenan Hom, Daniel Drane, May Dongmei Wang

Keywords: Topological Deep Learning, Cell Complexes, Conditional Variational Autoencoder, Disease Subtyping, FDG-PET, Alzheimer's Disease

Registration Requirement: Yes

Abstract: Data-driven subtyping of Alzheimer's disease (AD) using conditional variational autoencoders (cVAEs) has identified metabolic subtypes from FDG-PET, but existing approaches provide no insight into which inter-regional metabolic relationships define each subtype. We introduce a parcellated cVAE with a Differentiable Cell Complex Module (DCM) that learns higher-order topology over atlas-parcellated brain regions, enabling simultaneous subtype discovery and interpretable connectivity mapping. Applied to 716 AD subjects from ADNI, our model identifies two severity-matched subtypes with anti-correlated connectivity (r=-0.81), distinct cognitive profiles (p<0.001), and differential CSF tau (p=0.0008, corrected), corresponding to posterior-cortical and limbic AD variants.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 132