Spatiotemporal Atrophy Subtypes in Alzheimer's Disease: Neurodegenerative-Vascular Associations and Potential Prognostic Value.
OBJECTIVE To characterize spatiotemporal atrophy subtypes in biomarker-confirmed Alzheimer's disease (AD) and examine their cognitive, vascular, molecular, and longitudinal correlates. METHODS We applied Subtype and Stage Inference (SuStaIn) to baseline structural magnetic resonance imaging (MRI) from 484 amyloid-positive patients using 17 regional volumes. We compared subtype-specific cognition, cerebral small vessel disease burden, and post-baseline Mini-Mental State Examination (MMSE) trajectories; vascular risk factors and fluid biomarkers were exploratory. In a complete-case subset (45 patients; 134 observations), we compared SuStaIn with cerebrospinal fluid (CSF), plasma, hippocampal-volume, and combined-biomarker prognostic models. RESULTS A four-subtype solution was reproducible across cross-validation folds (Bhattacharyya coefficient, 0.93): Typical (55%), Parietal-predominant (18%), Limbic-predominant (14%), and Hippocampal-sparing (13%). Higher SuStaIn stages correlated with lower baseline MMSE scores (ρ = -0.56). The Parietal-predominant subtype showed earlier onset, executive and visuospatial dysfunction, and rapid MMSE decline. The Limbic-predominant subtype showed delayed-recall and naming deficits. The Hippocampal-sparing subtype showed relative delayed-recall preservation, the highest adjusted deep white matter hyperintensity burden, and the lowest adjusted perivascular space count. It also had a higher adjusted CSF t-tau/Aβ42 ratio than the Typical and Parietal-predominant subtypes. SuStaIn showed the best in-sample prognostic performance based on marginal R² (0.579; comparator range, 0.239-0.483) and AIC. CONCLUSIONS Biomarker-confirmed AD comprised internally reproducible atrophy trajectories with distinct cognitive, vascular, and molecular profiles. Spatiotemporal MRI may complement fluid biomarkers and conventional volumetry for patient stratification and prognosis. External validation is required to establish predictive generalizability.