Title: Leveraging Graph-Guided Universum Learning for Enhanced Alzheimer’s Disease Classification via Generalized Eigenvalue Proximal SVMs

Abstract:

The early and precise identification of Alzheimer’s disease (AD) is critical for implementing timely interventions and effective disease management. While Generalized Eigenvalue Proximal Support Vector Machine (GEPSVM) and its Universum-based extensions have demonstrated potential in AD classification, current approaches often treat Universum samples as isolated points, neglecting the geometric relationships among them. To address this limitation, this study introduces two novel graph-guided Universum learning frameworks: UG-GEPSVM and IUG-GEPSVM, designed to distinguish AD patients from cognitively normal (CN) individuals using structural MRI data.

In our proposed approach, subjects with mild cognitive impairment (MCI) serve as Universum data, offering intermediate informational cues between the AD and CN classes. We construct a graph over these MCI samples by leveraging Gaussian similarity, Minimum Spanning Tree connectivity, and multi-hop propagation. From this graph, we derive a Laplacian matrix that encapsulates the geometric structure of the MCI data. This Laplacian-based regularization replaces the conventional independent Universum penalty term within the learning process. Specifically, UG-GEPSVM incorporates this regularization into the generalized eigenvalue formulation, whereas IUG-GEPSVM extends the numerically stable improved GEPSVM framework using a standard eigenvalue formulation.

Experimental evaluations conducted on variants of the ADNI MRI dataset, utilizing both ICA- and PCA-based features across five distinct noise levels, demonstrate that both proposed models consistently surpass existing GEPSVM and Universum-based methods. Notably, UG-GEPSVM achieves a superior average AUC of 88.07% and exhibits robust performance stability even as noise levels increase. Statistical analyses further validate the significance of these performance improvements.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC