Deep Learning Methods for Predicting Alzheimer’s Disease Progression: Transformer and Graph Neural Network Approaches

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder affecting millions globally. Early detection of disease progression is critical for timely intervention. This thesis presents two complementary deep learning approaches for predicting AD progression using longitudinal data. The first approach employs a Transformer-based model that leverages visit history features, including cognitive test scores and imaging features. The second approach uses a history-aware graph neural network (HA-GNN) that operates on functional connectivity derived from resting-state functional MRI data. Both approaches address key challenges in longitudinal prediction, including irregular visit spacing, missing data, and class imbalance. Results demonstrate that Transformer-based models achieve superior performance on multi-class diagnosis prediction (82.4 accuracy), particularly for identifying disease converters. The HA-GNN model achieves 82.9 accuracy in binary conversion prediction, offering potential for early intervention. These findings emphasize the importance of model selection in predicting cognitive decline, with implications for clinical decision-making and patient outcomes.