From Biomarker Anchors to Disease-State Transitions: A Biologically Anchored Probabilistic Principal Component Analysis Framework for Alzheimer’s Disease Progression Modeling

To develop and validate a biomarker-anchored probabilistic principal component analysis (PPCA) framework for identifying biologically interpretable latent dimensions of Alzheimer’s disease (AD) and evaluating their utility for disease progression modeling. Data from 1058 participants that are amyloid-positive in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) were analyzed. Anchored PPCA used biomarker data only; cognitive, functional, diagnostic, and prognostic outcomes were withheld from latent-space construction and reserved for validation. Amyloid (A) and tau (T) factors were biologically anchored, neurodegeneration (N) was softly constrained, and a ventricular–vascular/residual (V/R) factor was empirically estimated. Robustness, reproducibility, held-out validity, clinical associations, prognostic performance, and clinical-state transitions were evaluated. Anchored PPCA identified four biologically coherent dimensions: A, T, N, and V/R. Solutions were reproducible across repeated initializations and split-half analyses, and generalized to held-out participants, particularly for A, T, and N. The model explained approximately 60% of model-implied standardized biomarker variance, with highest explained variance for A and T and lowest for V/R. Latent factors explained variance in outcomes not used for model construction: 49.7% for ADAS-Cog13, 35.1% for CDR-SB, and 27.8% for FAQ; they also aligned with diagnostic classifications. Cox-model C-indices were 0.843 for latent factors alone and 0.933 for clinical-plus-latent factors, with better fit than the clinical benchmark. T showed the strongest prognostic association, followed by A, N, and V/R. After adjustment for baseline clinical severity, A, T, and N remained independently prognostic; V/R did not. Clinical-state transitions were predominantly monotonic and demonstrated marked sojourn-time dependence. Biomarker-only anchored PPCA provides a biologically grounded representation of AD that validates against cognition, function, diagnosis, and clinical progression. A, T, and N may capture disease processes not fully reflected in cross-sectional clinical severity, whereas V/R appears more closely related to contemporaneous clinical status. Duration-dependent transitions support semi-Markov disease-progression and Shared Latent Disease Process models. External validation is warranted.