Diagnostic tools for detecting and quantifying non-biological signal in biomarker classifier evaluation. These are generic miRNA/omics QC checks that every biomarker study should pass before any AUC claim. They expose confounding from:
cohort identity (institution, batch, dataset),
demographic covariates (age, sex, BMI),
pre-analytical asymmetry (storage, collection site, tube type),
specimen duplication across accessions.
Details
The motivating problem is simple: on the four public 3D-Gene ovarian-cancer cohorts (GSE106817, GSE211692, GSE113486, GSE113740), a logistic regression using only dataset-identity dummy variables reaches AUC 0.72 against cancer/healthy labels without touching a single miRNA value. Any classifier trained on the same cohort structure therefore inherits this "bias floor" even if its apparent performance is higher.
The functions here make that floor visible before any biological claim is made, and extend to:
covariate-only AUCs (age, sex, etc.),
clustered bootstrap CIs that respect specimen duplication,
feature-level batch-signal ANOVA within healthy samples,
case/control pre-analytical asymmetry tables.
References
Leek JT, Scharpf RB, Bravo HC, et al. (2010). Tackling the widespread and critical impact of batch effects in high-throughput data. Nature Reviews Genetics, 11(10), 733-739.
Pritchard CC, Kroh E, Wood B, et al. (2012). Blood cell origin of circulating microRNAs: a cautionary note for cancer biomarker studies. Cancer Prevention Research, 5(3), 492-497.
Kirschner MB, Kao SC, Edelman JJ, et al. (2011). Haemolysis during sample preparation alters microRNA content of plasma. PLoS ONE, 6(9), e24145.