Extracting Interpretable Higher-Order Topological Features across Multiple Scales for Alzheimer's Disease Classification

Dengyi Zhao, Shengli Li, Yunping Wang, Chenfei Wang, Zhiheng Zhou, Guiying Yan, Xingqin QiarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed a new framework that uses Persistent Homology—a tool from algebraic topology—to "see" the complex, multi-dimensional holes (cycles and cavities) in brain networks. By quantifying how these higher-order structures break down in Alzheimer’s Disease (AD), the team achieved a classification accuracy of 93.6%, significantly outperforming traditional graph-based and deep learning methods while providing neuroscientifically interpretable results.

What Problem Does This Paper Address?

Most current research into Alzheimer’s Disease (AD) using functional MRI (fMRI) treats the brain like a simple road map: regions are "cities" (nodes) and functional connections are "roads" (edges). While useful, this lower-order approach misses the "roundabouts" and "tunnels"—the complex, multi-region loops and voids that represent high-level information processing and feedback.

Furthermore, traditional methods often rely on a fixed "threshold" to decide if a connection exists. If you set the threshold too high, you lose subtle signals; too low, and the network becomes a noisy mess. This paper tackles the challenge of capturing these intricate, higher-order structures across multiple scales simultaneously, turning abstract mathematical "holes" into reliable biomarkers for disease.

Methods

The authors propose a three-stage pipeline: Filtering, Quantification, and Classification.

  1. Filtering (The Topological Lens): Instead of a single graph, they use Vietoris-Rips filtration. They represent the brain as a sequence of "simplicial complexes" (think of building 3D shapes out of triangles and tetrahedrons). As they vary the connectivity threshold, they track when "holes" (cycles and cavities) are born and when they disappear. This is called Persistent Homology.
  2. Quantification (The Bridge to ML): To make this abstract data readable by machines, they use four vectorization methods:
  3. Persistence Landscapes: Mapping "mountain peaks" of topological features.
  4. Betti Curves: Counting the number of holes at every scale.
  5. Heat Kernels: Treating the birth/death of features like a diffusion process.
  6. Persistent Entropy: Measuring the complexity and uniformity of the network's topology.
  7. Classification: These features are fused with traditional edge-based data and fed into standard classifiers like Support Vector Machines (SVM) and Multi-Layer Perceptrons (MLP).

Datasets

The study utilized the ADNI1 database, a benchmark in Alzheimer's research.

  • Total Subjects: 232 individuals.
  • Split: 116 Alzheimer’s Disease (AD) cases and 116 Cognitively Normal (CN) controls.
  • Data Type: Resting-state fMRI, parcellated into 90 regions of interest (ROIs) using the AAL atlas.

Key Insights and Ideas

  • The "Disappearing Loop" Hypothesis: The most striking finding is that AD patients have significantly fewer cycles (1D holes) and cavities (2D holes) than healthy controls. In neuroscience terms, these cycles represent feedback loops essential for memory and cognitive control. AD literally "unravels" the brain's complex wiring.
  • Higher-Order > Lower-Order: While simple connectivity (edges) is important, it isn't enough. The study found that combining lower-order edge features with higher-order topological features provided the best results.
  • Multi-Scale is Mandatory: By using Persistent Homology, the model avoids the "Goldilocks problem" of picking the perfect threshold. It looks at the network's evolution across all possible thresholds, capturing features that are "persistent" and therefore likely biologically relevant.

Key Results

  • Superior Accuracy: The framework reached a peak accuracy of 93.6% using SVM or MLP. For context, this is a 20–30% improvement over traditional graph theory metrics and roughly 13% better than state-of-the-art Graph Neural Networks (GNNs).
  • Interpretability: Unlike "black box" deep learning models, this method identified specific brain regions most affected by topological breakdown. Key regions included the Caudate, Amygdala, and Precuneus—all of which are well-documented in clinical literature as being central to AD pathology.
  • Clear Separation: t-SNE visualizations showed that the topological features created distinct, well-defined clusters for AD vs. CN subjects, whereas other methods resulted in intermixed, messy data points.

Limitations and Future Work

While the results are impressive, the authors acknowledge the inherent noise and complexity of fMRI data, which can make extracting these delicate topological features difficult in a clinical setting with lower-quality scans.

Future directions include: * Applying this framework to other neurodegenerative disorders (like Parkinson's) to see if they exhibit unique "topological signatures." * Integrating longitudinal data to see if the loss of cycles can predict the transition from Mild Cognitive Impairment (MCI) to full-blown Alzheimer's. * Exploring even higher-dimensional holes (3D and beyond), though the computational cost increases exponentially.

Summary updated

Abstract

Brain network topology, derived from functional magnetic resonance imaging (fMRI), holds promise for improving Alzheimer's disease (AD) diagnosis. Current methods primarily focus on lower-order topological features, often overlooking the significance of higher-order features such as connected components, cycles, and cavities. These higher-order features are critical for understanding normal brain function and have been increasingly linked to the pathological mechanisms of AD. However, their quantification for diagnosing AD is hindered by their inherent nonlinearity and stochasticity in the brain. This paper introduces a novel framework for diagnosing Alzheimer's disease that uses persistent homology to extract higher-order topological features from fMRI data. It also introduces four quantitative methods that capture subtle, multiscale geometric variations in functional brain networks associated with AD. Our experimental results demonstrate that this framework significantly outperforms existing methods in AD classification. Extensive ablation studies and interpretability analysis confirm the effectiveness of our framework. Our study also reveals that the number of cycles or cavities significantly decrease in AD patients. The extracted key brain regions derived from cycles and cavities align with domain knowledge in neuroscience literature and provide direct and insightful findings. This study highlights the potential of higher-order topological features for early AD detection and significantly advances the field of brain topology analysis in neurodegenerative disease research.
BibTeX
@misc{zhao2025extracting,
  author = {Dengyi Zhao and Shengli Li and Yunping Wang and Chenfei Wang and Zhiheng Zhou and Guiying Yan and Xingqin Qi},
  title = {{Extracting Interpretable Higher-Order Topological Features across Multiple Scales for Alzheimer's Disease Classification}},
  year = {2025},
  eprint = {2509.14634},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.14634},
  howpublished = {arXiv},
}
Figure