Yasmine Attia, a doctoral candidate in computer science, will defend their dissertation titled “Anatomically-Informed Interpretable Multimodal Learning for Alzheimer’s Disease Diagnosis and Progression.” Their advisor, Dr. Seung-Jong Park is an Kummer Endowed Chair in the computer science department. The dissertation abstract is provided below.

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by gradual structural brain changes that begin long before clinical symptoms emerge. Detecting these changes across disease stages remains challenging, requiring models that capture subtle anatomical variations, integrate heterogeneous data, and provide clinically meaningful interpretations. Although multimodal learning offers strong potential, existing approaches often suffer from complexity, redundant computation, and limited interpretability. Our work progresses from efficient multimodal fusion to anatomically grounded and temporally aware modeling. It begins with an early–late fusion (ELF) framework that achieves strong performance with reduced complexity. Sensitivity to structural changes is improved using Jacobian-domain representations, while interpretability is introduced through Jacobian Saliency Maps (JSMs) and an efficient alternative, Sobel Kernel Angle Difference (SKAD). To better model multimodal interactions and missing data, transformer-based architectures are adopted. The Anatomical Attention Transformer (AAT) integrates MRI, PET, and clinical text using atlas-guided region tokens, embedding anatomical priors, and enabling region-level reasoning aligned with known biomarkers. Synthetic clinical narratives further support multimodal alignment without diagnostic leakage. Efficiency and deployability are improved through anatomically guided pruning of 3D Vision Transformers, reducing computation while enhancing focus on disease-relevant regions. Finally, disease progression is modeled using Forecast-Guided Latent Transition (FGLT), which captures longitudinal dynamics through conditional latent forecasting and supports progression prediction and counterfactual analysis. Together, these contributions demonstrate that combining efficient fusion, interpretability, anatomical priors, and longitudinal modeling yields accurate, robust, and clinically meaningful systems for AD analysis.