Uncovering Latent Heterogeneity in Alzheimer's Disease via Clustering of Cortical GNN Explanations

Ian Liu, Da Ma

Keywords: Alzheimer's disease, graph neural networks, explainability, clustering, cortical surface, subtype discovery

TL;DR: We propose reframing GNN explainability as a tool for disease phenotyping by clustering cortical explanation maps to uncover latent heterogeneity in AD.

Registration Requirement: Yes

Abstract: Alzheimer's disease (AD) is clinically and biologically heterogeneous, yet many neuroimaging classifiers reduce this variability to a single disease score. We investigate whether subject-level explanation maps from a cortical graph neural network (GNN) can reveal latent heterogeneity within AD. Using baseline T1 MRI from stable cognitively normal (CN) and AD participants from ADNI, we utilize the best-performing GATv2 from our prior benchmark and create node-feature attribution maps over area, curvature, and thickness. We performed split-safe k-means clustering, with fitting and preprocessing only on training AD subjects. We compared clustering results from raw cortical features, pre-classifier GATv2 embeddings, and GNN explanation maps across $k = 2, ..., 6$ using training-set silhouette score, cluster-size constraints, and anatomical interpretability of held-out assignments. Compared to raw features and graph embeddings, node explanation maps yielded the most interpretable non-degenerate structure, separating AD subjects into two explanation phenotypes with distinct temporal versus more diffuse cortical reliance patterns.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 119