AlzheimerODE: A Knowledge-Guided Generative Model for Alzheimer's Disease Progression

Abstract

Accurately modeling the progression of Alzheimer's disease (AD) is essential for understanding disease mechanisms and optimizing therapeutic windows, yet it is impeded by the sparse, irregular, and heterogeneous nature of clinical longitudinal data. Existing generative approaches either struggle to capture the continuous-time dynamics or neglect the biological constraints that govern disease evolution. To address these challenges, we present AlzheimerODE, a knowledge-guided generative framework that learns structured disease dynamics as a reaction–diffusion flow in continuous time. Unlike purely data-driven models, we introduce mechanistic inductive biases from Alzheimer's pathology to parameterize the disease velocity field, capturing nonlinear interactions among biomarkers and spatial propagation across brain networks. To tackle substantial inter-subject heterogeneity, we incorporate a survival-conditioned modulation of the dynamics to enable personalized trajectory generation. Empirical evaluations on longitudinal neuroimaging datasets demonstrate that our method consistently outperforms state-of-the-art baselines across trajectory accuracy, temporal alignment, spatial pattern consistency, and distributional fidelity. By unifying structured dynamical modeling with deep generative learning, AlzheimerODE provides a robust computational tool for longitudinal data augmentation and the fine-grained simulation of biologically coherent disease dynamics.