Blood-based staging of Alzheimer's disease — and the label circularity that inflates it. Machine-learning multiclass classification of cognitive stages using plasma biomarkers, clinical assessments and genetic features: a repeated, nested cross-validation study in ADNI with external evaluation in CNTN.

Jiayuan Xu, Andrew J. Doig, Sofia Michopoulou, Petroula Proitsi, Fumie Costen.

Background: Plasma biomarkers are promoted as scalable tools for staging Alzheimer's disease (AD), but head-to-head comparisons against the clinical scales used to define diagnostic labels are scarce, and reported machine-learning "fusion" gains may reflect label circularity rather than biological signal.

Methods: From ADNI we assembled 655 participants (CN=296, MCI=168, AD=191) with concurrent plasma biomarkers (pT217, Aβ42/40, NfL, GFAP), clinical scales (MMSE, CDR-SB, FAQ), APOE genotype and demographics. Three pre-specified feature sets (clinical-only; biomarker + demographic–genetic; full fusion) were compared across four classifiers (Logistic Regression, SVM, Random Forest, XGBoost) using repeated, nested cross-validation (5-fold × 3 outer, 5-fold inner). External evaluation used the CNTN cohort (n=130) on a reduced shared panel (NfL, GFAP, APOE, age, sex, education).

Results: Clinical scales alone reached three-class AUC-OVR 0.954 and fusion 0.956 — statistically indistinguishable (Wilcoxon p=0.33). Because MMSE, CDR-SB and FAQ partly define the ADNI labels, both estimates are circularity-inflated. The de-confounded plasma + demographic–genetic model achieved AUC-OVR 0.746, with pT217 dominant (mean|SHAP| 0.106, 1.8× NfL; best subset pT217+NfL 0.753); SHAP attributed 78% of the fusion model's importance to the clinical scales. Pairwise discrimination: CN vs AD ≈1.00, MCI vs AD 0.97, CN vs MCI 0.93 (0.70 plasma-only). The reduced panel transferred to CNTN at AUC-OVR 0.702 (95% CI 0.635–0.764).

Conclusions: Apparent fusion gains are largely label circularity. Realistic blood-based CN/MCI/AD staging is ≈0.74 (internal) / 0.70 (external); pT217 carries the plasma signal; MCI detection is the principal bottleneck.

Keywords: Alzheimer's disease; mild cognitive impairment; plasma biomarkers; pT217; machine learning; nested cross-validation; multiclass classification; ADNI; CNTN; SHAP; label circularity.