Safeguards for Integrative Re-Analysis of Public Blood Transcriptomes: Undeclared Patient Overlap, Cell Composition, and Batch-Confounded Covariates in Longitudinal Stroke

Background: Pooling public blood transcriptome studies never designed for integration introduces failures that standard diagnostics miss. Objective: To define safeguards for confounding-aware integration and establish what an integrated stroke dataset can support. Methods: Eleven GEO series (1885 samples) were integrated under six safeguards: expression fingerprinting for undeclared patient overlap; staged ComBat with leave-one-study-out PCA; scale-correct deconvolution; composition-adjusted differential expression; an estimability audit with leakage-safe cross-validation; and a provenance audit against the source studies’ deposited tables. Results: Fingerprinting found 33 shared patients in two series. After ComBat, study and platform explained ~0% of PC1, and cell composition explained 61.4%. Linear rather than log-scale deconvolution input changed the neutrophil fraction (0.07 to 0.25) but did not affect the level of cross-method agreement (r = 0.81 vs. 0.78). Neutrophil adjustment removed 65–84% of acute differentially expressed genes, and co-expression modules re-correlated with composition even among composition-robust genes. Stroke subtype was fully confounded with study (classifier AUC 0.882 from single-class folds). The provenance audit reclassified our novel candidate finding in cell-sorted data as replicative and confirmed the depositors’ differential expression (monocytes: r = 0.92, neutrophils: 0.90). The integrated dataset supported a direction-resolved acute signature, three temporal response programs in a 513-gene core, and seven candidate stroke-versus-trauma discriminators with divergent kinetics. Conclusions: Passing integration diagnostics does not guarantee estimability. Estimability must be assessed before correction removes the evidence, and novelty must be checked against deposited data.