Cross-cohort generalization limitations in gut microbiome-based Alzheimer's disease classifiers: evidence from four supervised cohorts with batch-correction analysis

Gut microbiome dysbiosis has been proposed as an Alzheimer's disease (AD) biomarker, but the transportability of microbiome classifiers across independent cohorts remains uncertain. We reprocessed five public 16S rRNA cohorts (509 samples, China, Kazakhstan, South Korea) with a common DADA2/genus-level pipeline and evaluated logistic regression and LightGBM classifiers using nested within-cohort cross-validation, leave-one-cohort-out (LOCO) transfer, pairwise transfer, permutational multivariate analysis of variance (PERMANOVA), batch adjustment, and out-of-fold SHapley Additive exPlanations (SHAP) analyses. Within four labeled cohorts, using strict training-only genus selection as the primary classifier-performance pipeline (Section 2.5.1), within-cohort area under the receiver operating characteristic curve (AUC-ROC) ranged from 0.611 to 0.996. The near-perfect Li 2019 result was stable across 50 repeated nested-cross-validation (CV) partitions and exceeded a 1,000-permutation label null (p = 0.000999), though a technical metadata confound cannot be excluded (Section 4.3). In LOCO evaluation, AUC ranged from 0.506 to 0.812 (mean decreases of 0.170 for logistic regression and 0.133 for LightGBM); paired bootstrap intervals excluded zero for four of eight cohort-model differences. LOCO estimates agreed closely with a fixed common-universe reference retained for secondary analyses (max |Δ| = 0.066), though one pairwise-transfer cell changed materially (Kazakhstan→Ling 2020, LightGBM: 0.58 → 0.44), reported without smoothing. In the four-cohort marginal PERMANOVA model, cohort-associated compositional variation exceeded diagnosis-associated variation (R² = 0.172 vs. 0.014), decreasing to 0.059 vs. 0.023 when restricted to three more technically comparable Chinese paired-end MiSeq cohorts; dispersion differed significantly among cohorts. Label-blind transductive ComBat-seq and MMUPHin adjustment changed mean LOCO AUC by at most 0.007 and did not recover transferability. Cross-cohort overlap of top predictive taxa was low: logistic regression overlap remained above a prevalence-restricted null, while LightGBM did not. A descriptive coefficient-stability screen identified eight logistic-regression genera with opposite, stable fitted-coefficient directions across cohorts. These findings demonstrate substantial, heterogeneous cross-cohort generalization limitations among the evaluated public 16S datasets and show that, under the tested label-blind transductive design, the two batch-adjustment methods did not restore transferability. Prospective validation in larger, geographically and technically diverse cohorts is needed before microbiome-based AD classifiers can be considered clinically actionable.