Machine Learning-Based Multiclass Classification of Cognitive Stages Using Plasma Biomarkers, Clinical Assessments,...

Background: Plasma biomarkers are promoted as scalable tools for the staging of Alzheimer's disease (AD), yet head-to-head comparisons against the clinical scales used to define diagnostic labels remain scarce. Reported gains from machine learning fusion of clinical and biomarker features may reflect label circularity rather than biological signals, and quantifying this circularity is a central aim of the present work. Methods: From the Alzheimer's Disease Neuroimaging Initiative (ADNI), we assembled 655 participants (CN = 296, MCI = 168, and AD = 191) with concurrent plasma biomarkers (pT217, Aβ42/40, NfL, and GFAP), clinical scales (MMSE, CDR-SB, and FAQ), APOE genotype, and demographics. Three pre-specified feature sets (clinical-only, biomarker plus demographic-genetic, and full fusion) were compared across four classifiers (Logistic Regression, SVM, Random Forest, and XGBoost) using repeated, nested cross-validation (5-fold × 3 outer, 5-fold inner) with balanced class weighting. Because the external Center for Neurodegeneration and Translational Neuroscience (CNTN) cohort (n=130) measures pT181 rather than pT217 and lacks Aβ42/40, external evaluation used a separate reduced feature panel (NfL, GFAP, APOE, age, sex, and education), not the proposed pT217-inclusive panel. Results: Clinical scales alone reached a three-class AUC-OVR of 0.9539±0.0041, and fusion reached 0.9559±0.0046, an indistinguishable gain. Because MMSE, CDR-SB, and FAQ partly determine ADNI diagnostic labels, both estimates are circularity-inflated upper bounds and do not reflect independent classification power. Independent of this circularity, the internal plasma plus demographic-genetic model still achieved AUC-OVR =0.7455±0.0150, with pT217 as the dominant contributor. Pairwise discrimination was excellent for CN vs. AD (1.0000) and MCI vs. AD (0.9739) but markedly weaker for CN vs. MCI (0.9302 for fused and 0.6972 for plasma only). The separate reduced-feature model, which contains neither pT217 nor Aβ42/40, transferred to CNTN with AUC-OVR =0.702 (95% CI 0.635-0.764). Conclusions: Apparent fusion gains in ADNI are largely a consequence of label circularity. After removing the circular clinical features, the internal pT217-inclusive plasma model supports three-class CN/MCI/AD screening at AUC ≈0.74 and a reduced panel without pT217 transfers to an independent cohort at AUC ≈0.70. These values provide a realistic performance estimate for blood-based AD staging under the current feature set, diagnostic label structure, and cohort design, and richer feature sets or pathology-anchored labels may shift this estimate. MCI detection remains the principal bottleneck.