Connecting the Nodes: Multimodal Graph Learning for Early Differential Diagnosis and Prognosis in Dementia
Hassan, Salma Wael Fekry
Hassan, Salma Wael Fekry
Author
Supervisor
Department
Machine Learning
Embargo End Date
2026-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Alzheimer’s Disease (AD) represents one of the most pressing challenges in global healthcare, with devastating impacts affecting over 55 million people worldwide and 10 million
new cases diagnosed annually. Despite significant research efforts, the field faces three critical challenges: differential diagnosis between dementia subtypes, accurate estimation
of cognitive decline trajectories, and reliable prediction of conversion time from preclinical to full AD onset. Existing approaches often treat these problems in isolation and fail to
leverage the complex relationships inherent in multimodal medical data. This thesis addresses these challenges through three novel deep learning frameworks that advance multimodal data integration for comprehensive AD management. First, MINDSETS achieves 89.25% accuracy in differentiating AD from Vascular Dementia through the integration of MRI radiomics features, genetic data, and clinical assessments. The framework introduces a Deep Feature Generation module that enhances discriminative power and enables monitoring of treatment efficacy by tracking temporal changes in feature importance patterns, demonstrating a 6.38% average decrease in MCI probability for treated patients compared to 2.53% for controls. Second, ClinGRAD enhances AD classification accuracy to 93.15% through a clinically guided heterogeneous graph neural network that implements biologically-informed connections between brain structures and genetic factors. By modeling gene co-expression networks and diffusion-weighted imaging-derived connectivity patterns, this framework reveals underlying biological mechanisms through pathway-based clustering while maintaining interpretability through attention-based visualization of gene-structure interactions. Third, MAGNET-AD pioneers multitask spatiotemporal graph neural networks to predict cognitive decline and conversion time in preclinical AD. The framework achieves a concordance index of 0.858 for conversion time prediction and a mean square error of 1.983 for PACC score estimation through its hybrid fusion architecture that combines static genetic markers with dynamic neuroimaging patterns. MAGNET-AD introduces a temporal importance weighting mechanism that automatically identifies critical progression time points, demonstrating robust performance even with limited longitudinal data. Collectively, these frameworks establish a new paradigm for early detection, differential diagnosis, and progression monitoring in AD through innovative architectures, robust temporal modeling, and clinically interpretable outputs. The demonstrated performance highlights their practical utility in real-world clinical settings, where they can support earlier intervention and more personalized treatment strategies for patients with neurodegenerative disorders.
Citation
Salma Wael Fekry Hassan, “Connecting the Nodes: Multimodal Graph Learning for Early Differential Diagnosis and Prognosis in Dementia,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Dementia, Multimodal, Graph Neural Network, Neuroimaging, Genomics, Spatiotemporal Modelling