Feature harmonization reveals that apparent cross-cohort differences in tau-positivity prediction reflect variable availability and cohort stage structure

Background: Cross-cohort machine-learning studies of tau positivity are increasingly used to assess generalizability of Alzheimer's disease biomarker models, but apparent differences between cohorts may reflect clinical severity structure and cohort stage composition rather than genuine biological divergence. Methods: We evaluated this problem across subject-level cohorts from the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Open Access Series of Imaging Studies (OASIS3), and the National Alzheimer's Coordinating Center (NACC) using harmonized tabular predictors and a feature-block ablation framework. For ADNI and OASIS3, we performed matched-feature analyses restricted to shared harmonized variables (age, APOE e4, education, sex, the CDR(R) Dementia Staging Instrument sum-of-boxes score (CDR-SB), and the Functional Activities Questionnaire (FAQ)). Models were trained with five-fold stratified cross-validation using balanced logistic regression as the primary classifier. Random forests and extra-trees classifiers were evaluated as robustness checks and are reported in supplementary analyses. Results: In the primary matched-feature comparison, reference areas under the receiver operating characteristic curve (AUCs) were 0.851 for ADNI tau90, 0.860 for ADNI tau180, and 0.933 for OASIS3 tau180. Severity-only models retained 86-99% of reference performance, whereas demographic/genetic-only models retained 85-87%; bootstrap confidence intervals showed that the reference-minus-demo contrast remained positive across all primary cohorts. Before feature matching, OASIS3 showed substantially weaker demographic-only prediction (AUC approx. 0.54), attributable to missing APOE and age variables. After harmonization with a matched feature set including age, the pre-harmonization demographic gap was no longer attributable to feature availability; the residual difference was most consistent with cohort disease-stage composition asymmetry, with CDR-SB = 0 prevalence of 84.5-88.7% in OASIS3 versus 49.3-51.7% in ADNI. Conclusions: The original large ADNI-OASIS3 performance gap was largely a feature-availability artifact; after strict harmonization including age-field repair, the absolute gap narrowed and the signed ADNI-minus-OASIS3 difference reversed, with the residual difference most consistent with cohort disease-stage composition asymmetry. Clinical severity remains an important predictor across all cohorts, but the unique contributions of severity and demographic/genetic information depend strongly on stage composition. Cross-cohort biomarker modeling should account for both feature-set completeness and disease-stage composition heterogeneity, not only headline performance metrics.