GRACE — Graph-based Reasoning for Alzheimer’s Classification and Explanation
Khamitova, Ainur
Khamitova, Ainur
Citations
Altmetric:
Author
Supervisor
Department
Machine Learning
Embargo End Date
2027-05-01
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Alzheimer’s disease (AD) is the most common cause of dementia and is frequently preceded by Mild Cognitive Impairment (MCI), an intermediate stage characterised by measurable cognitive decline with relative preservation of everyday function. Distinguishing MCI from cognitively normal (CN) aging is clinically important for prognosis and monitoring, yet remains difficult because early disease effects are subtle, heterogeneous, and often expressed as distributed alterations in brain structure and network function. Neuroimaging provides complementary biomarkers through structural T1-weighted MRI and functional MRI (fMRI), but effectively combining these sources of information for clinically relevant multi-class classification (CN vs. MCI vs. AD) requires models capable of capturing complex interactions across brain regions. Moreover, strong predictive performance alone is insufficient for translation: clinical decision-making demands explainability, so that model outputs can be inspected, validated, and related to established neurobiological patterns rather than treated as opaque predictions.
This thesis investigates explainable multi-class classification of cognitive status using multimodal neuroimaging from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Participants with both modalities are curated and processed in a reproducible pipeline, after which each scan is represented as a subject-specific brain graph in which nodes correspond to regions of interest (ROIs) with region-wise anatomical descriptors and edges
encode functional connectivity. Building on this representation, we propose GRACE, an explainable graph-and-language framework that learns ROI embeddings using a GATv2-based graph neural network and generates clinician-readable diagnostic rationales by conditioning a large language model on graph-derived evidence via graph-token projection and parameter-efficient fine-tuning.
In this formulation, explainability is treated as a core objective: predictions are paired with natural-language reasoning intended to summarise both global network characteristics and region-level evidence, enabling direct inspection beyond post-hoc saliency. Experimental analysis demonstrates that graph construction and node feature design substantially influence class-balanced performance, with correlation-based connectivity and richer
anatomical features improving separability of the diagnostically challenging MCI group. Evaluation of the explanation component indicates that fluent, high-quality reports do not always correspond to improved label prediction, highlighting the importance of enforcing faithfulness constraints that more tightly couple generated rationales to the evidence driving the classifier. Overall, this thesis contributes an evidence-grounded approach for explainable multimodal neuroimaging classification of CN, MCI, and AD, and identifies methodological directions for improving robustness, aligning explanations with predictive evidence, and supporting transparent clinical reasoning.
Citation
Khamitova, Ainur, "GRACE — Graph-based Reasoning for Alzheimer’s Classification and Explanation," M.S. Thesis, Machine Learning, MBZUAI, 2026.
Source
Conference
Keywords
Graph Neural Networks, Alzheimer’s disease, Mild Cognitive Impairment, Neuroimaging, Explainable AI