Heterogeneous progression of AT(N) biomarkers in data-driven subtypes of Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) exhibits substantial heterogeneity in brain atrophy patterns and progression rates. Data-driven subtyping based on the Subtype and Stage Inference (SuStaIn) model has identified distinct MRI-derived subtypes, but the longitudinal dynamics of these subtypes within the AT(N) framework remain unknown. This study included 1,576 participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) with at least two visits. SuStaIn was applied to volumetric measurements of 11 brain regions to identify imaging-based subtypes. Within each subtype, we fitted a Bayesian latent time longitudinal joint mixed-effects model incorporating six biomarkers: amyloid PET (A), CSF p-tau (T), hippocampal volume, cingulate volume, thalamic volume (N), and ADAS-Cog13 (X), with APOE ε4 as a covariate. The three subtypes exhibited distinctly different longitudinal patterns. Subtype 1 showed early elevation of amyloid, followed by a sharp decline in hippocampal volume, which was tightly coupled with cognitive deterioration. Subtype 2 featured early involvement of cortical regions such as the cingulate cortex, with hippocampal changes occurring at a similar timing to Subtype 1 but accompanied by earlier cortical and thalamic involvement. Subtype 3 exhibited early rapid hippocampal atrophy that later plateaued, and demonstrated the strongest hippocampus–cognition coupling. The association of APOE ε4 with biomarker trajectories was subtype-dependent: its effect estimates were largest for amyloid and p-tau in Subtype 1, whereas in Subtype 3 the effect estimates were significant for nearly all biomarkers, with the largest estimate for cognition. This study demonstrates that AD subtypes defined by spatial atrophy patterns also differ fundamentally in temporal biomarker dynamics and APOE ε4 association, supporting subtype-stratified modeling of disease progression.