Conditional Diffusion Model using Ordinal Regression for Longitudinal Neurodegenerative Data Generation

作者：Hyuna Cho, Ziquan Wei, Seungjoo Lee, Tingting Dan, Guorong Wu, Won Hwa Kim · 发表于：Alzheimer s & Dementia · 年份：2025 · DOI：10.1002/alz70855_096933 · 被引用次数：1 · 研究领域：Dementia and Cognitive Impairment Research、Functional Brain Connectivity Studies、Advanced Neuroimaging Techniques and Applications

BACKGROUND: Neurodegenerative diseases like Alzheimer's disease (AD) progress irreversibly, making early detection critical. However, limited longitudinal data and irregular patient visits (e.g., 6 months to years apart) challenge modeling efforts. Recent generative methods do not consider the ordinal dynamics and temporal gaps in disease progression. Therefore, we propose a novel conditional generative model for synthesizing long-term brain regional measurements based on disease-relevant ordinal conditions like age and diagnostic labels. METHOD: We used three AD biomarkers from the Alzheimer's Disease Neuroimaging Initiative: cortical thickness (N = 178), Amyloid Standardized Uptake Value Ratio (SUVR) (N = 687), and Fluorodeoxyglucose (FDG) SUVR (N = 678), measured across 148 brain regions, with 2 to 10 time points per subject. Five diagnostic labels were used: Cognitively Normal (CN), Significant Memory Concern, Early Mild Cognitive Impairment, Late MCI, and AD, where the disease progresses irreversibly from CN to AD. Using age and label as independent variables, an ordinal regression model is fitted on the whole cohort to learn global patterns of disease progression. The fitted model yields multiple pseudo-samples that fill gaps between sparse observed data points within a sequence. Afterward, a conditional diffusion model sequentially estimates the difference in consecutive pseudo-samples to reconstruct the observed data points. RESULT: Table-1 shows that our method outpe...