MPP-GNN: Subject-Adaptive Community Detection for fMRI-Based Alzheimer’s Disease Classification

Alzheimer’s disease research relies heavily on functional magnetic resonance imaging (fMRI) to capture how brain regions interact over time. Turning those time‑series into graphs lets researchers apply graph neural networks (GNNs) to predict whether a subject has AD. Most existing GNN pipelines force every brain graph into the same set of predefined communities, ignoring the fact that each person’s functional architecture can differ. This mismatch can blur disease‑related signals and limit model performance. The new MPP‑GNN approach tackles both the community‑size assumption and the disconnect between discovered modules and the connectivity patterns the network learns.

What You Need to Know

The authors formulate the learning problem as a bilevel optimization. At the upper level, the GNN parameters are updated to minimize classification loss on AD labels. At the lower level, a meta‑probabilistic pooling layer learns, for each subject, a soft assignment of nodes to communities that best supports the upper‑level objective. Because the pooling is probabilistic, each node can belong to multiple communities with varying strengths, and the number of communities is not fixed a‑priori—instead, it emerges from the data through a learned Dirichlet‑process‑like prior.

Hierarchical graph partitioning is applied repeatedly: after each pooling step, the coarsened graph retains the subject‑specific community structure, allowing the GNN to capture multi‑scale patterns (from local clusters to whole‑brain networks). The pooling weights are themselves generated by a small meta‑network that conditions on subject‑level statistics (e.g., global signal variance, head‑motion metrics), making the community detection adaptive across individuals without manual tuning.

Experiments were conducted on the ADNI dataset, comparing MPP‑GNN against baseline GNNs that use fixed‑k spectral clustering or uniform pooling. The adaptive method consistently improved classification accuracy (by roughly 3‑5 % absolute) and yielded community maps that aligned better with known resting‑state networks, such as the default mode and salience systems, while also revealing subject‑specific deviations that correlated with cognitive scores.

Why It Matters

Current fMRI‑based diagnostic tools often treat brain connectivity as a static template, which can obscure the heterogeneity that characterizes Alzheimer’s pathology. By letting the model discover, per subject, how functional modules are organized and how strongly each region participates in those modules, MPP‑GNN provides a more faithful representation of the underlying neurobiology. This specificity can improve early detection, especially in prodromal stages where inter‑subject variability is high.

From a methodological standpoint, the bilevel formulation clarifies how unsupervised community detection can be directly supervised by a downstream task. Rather than treating clustering as a preprocessing step, the approach integrates it into the learning loop, ensuring that the discovered partitions are those that actually help the classifier. This principle could extend beyond AD to other neurological disorders where individualized brain networks are informative.

Key Details

• Meta‑probabilistic pooling: learns a soft, subject‑specific assignment matrix Sᵢ for each fMRI graph Gᵢ, allowing overlapping communities and a flexible number of modules.
• Bilevel optimization: upper level updates GNN weights θ to minimize AD classification loss; lower level updates pooling parameters φ to maximize the expected classification performance under the current θ.
• Hierarchical coarsening: after each pooling layer, the graph is reconstructed using Sᵢ, enabling the GNN to operate at multiple spatial scales while preserving subject‑level community structure.
• Subject‑conditioned meta‑network: a small feed‑forward network takes subject‑level covariates (e.g., age, motion) and outputs the Dirichlet‑process concentration parameters that govern the prior over community counts.
• ADNI evaluation: on 1,200 subjects (≈400 AD, 800 CN), MPP‑GNN reached 84.2 % accuracy, outperforming fixed‑k GNN baselines (79.5 %‑81.0 %) and showing statistically significant improvements in AUC.
• Interpretability: the learned soft assignments visualized as community maps revealed subject‑specific shifts in the default mode network that correlated with MMSE scores (r = ‑0.32, p < 0.01).

What’s Next

Future work could extend the meta‑probabilistic pooling to longitudinal fMRI, letting the community structure evolve over time and potentially capture disease progression trajectories. Additionally, integrating multimodal data (e.g., structural MRI or PET) as conditioning inputs for the meta‑network may yield even more subject‑tailored representations, moving the field closer to personalized, biologically grounded diagnostics for Alzheimer’s disease.

📌 Source: Arxiv Ml