Interpretable Transformer-Based Longitudinal Modeling of Alzheimer’s Disease Diagnosis and Progression

Abstract

Alzheimer’s disease (AD) is the most common cause of dementia worldwide, accounting for 60-80% of cases. It is a progressive neurodegenerative disorder characterized by memory loss, cognitive decline and functional impairment, ultimately leading to loss of independence and death. Despite decades of research, reliable early diagnosis and prediction of disease progression remain a critical challenge, particularly at the stage of mild cognitive impairment (MCI), when clinical symptoms are subtle but risk of conversion to AD is elevated. This thesis develops and evaluates deep learning Transformer-based approaches for modeling disease status and progression using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The work focuses on two predictive tasks: (i) current-visit diagnosis (AD vs. mild cognitive impairment (MCI) and cognitively normal (CN)), and (ii) next-visit conversion prediction, from MCI to AD. Results show strong performance for current-visit diagnosis (mean ROC-AUC > 0.90 for AD vs CN), while conversion prediction remains more challenging. To enhance intepretability, Integrated Gradients were employed to attribute model predictions to individual features. This analysis consistently highlights established AD biomarkers. Subgroup analysis was also conducted by clustering patient embeddings in the learned representation space. This revealed patterns of risk: one subgroup characterized by more high-risk individuals, and the other more low-risk. Together, these findings underscore the potential of Transformer-based models for longitudinal prediction in Alzheimer’s disease.