Beyond Thresholds: Multi-Modal Graph-Based Learning for Predictive Scoring in Preclinical Alzheimer's Disease

Nupur Thakur, Riti Paul, Yuxiang Zhou, Oana Dumitrascu, Baoxin Li

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Graph Neural Network, MR Imaging, Multi-modal Learning, Preclinical Alzheimer's Disease, Semi-supervised Learning

TL;DR: A novel semi-supervised, graph-based approach, using MR imaging and non-imaging data graph to predict the preclinical Alzheimer's Disease risk (that approximates SUVR score).

Abstract:

The regression task of predicting preclinical Alzheimer’s disease (AD) risk using imaging and other data is essential for evaluating the varying risk levels across individuals. Existing data-driven methods for identifying preclinical AD primarily focus on binary classification, which relies on thresholds derived from specific imaging datasets. This reliance on dataset-specific thresholds can hinder the generalization capability of such methods across datasets, limit the integration of other types of data, and restrict the assessment of varying risk levels among individuals. In this paper, we propose a novel, multi-modal, semi-supervised regression framework to predict amyloid positivity scores by integrating brain MR imaging and non-imaging data, such as genetic information and cognitive assessments.
We introduce an unsupervised, imaging data-driven module for prototype label generation, for the `regression by classification' learning strategy. We employ graph neural networks to model the complex relationships within diverse non-imaging data. Our learning algorithm makes optimal use of the labeled and unlabeled data, maximizing the utility of limited labeled information, while eliminating the need for rigid thresholds. Through extensive evaluation on the ADNI dataset, which includes real-world patient data, our framework demonstrates effectiveness, suggesting that it could offer a more adaptable, precise tool for preclinical AD assessment and a step forward in the application of computer vision and deep learning to neurodegenerative disease detection.

Track: 4. Clinical Informatics

Registration Id: 2HN59CRJRTG

Submission Number: 131