Hidden Dependencies in Endometriosis Microarray Classifiers: Sample Reuse, Outcome-Informed Batch Correction, and Diagnostic Reassessment

Distinct Gene Expression Omnibus (GEO) accessions do not necessarily represent independent transcriptomic samples. We investigated exact array reuse, access to test diagnoses during batch correction, and cohort and acquisition-date confounding in endometriosis classifiers. We compared decompressed CEL files using SHA-256, reconstructed a matrisome elastic-net workflow, changed test labels supplied to ComBat, and conducted 3,582 constrained training-label refits. We also evaluated the effect of removing matching training copies in a five-gene Huang-inspired random forest and performed a 324-probe nested internal benchmark within GSE51981. Of 37 test CEL files, 26 (70.3%) were identical to training arrays; all had been scanned in 2004-2005. Changing a test diagnosis supplied to ComBat moved the fixed classifier score toward the supplied label in all 37 interventions. In a deduplicated, batch-constrained comparison, median nominal balanced accuracy was 0.881 when ComBat received true test outcomes and 0.549 without them; the outcome-informed arm performed better in 1,193 of 1,194 matched refit comparisons. In the separate Huang-inspired reconstruction, removing each target's matching training copy lowered pooled AUC from 1.000 to 0.323 (conditional 95% interval 0.105-0.571). Exploratory nested AUC was 0.733 for endometriosis versus other gynecological conditions, and date-held-out AUC in 2009 was 0.698. For endometriosis versus healthy controls, a date-only out-of-fold model achieved AUC 0.935, compared with 0.893 for the molecular model. Together, these findings show how exact array reuse, outcome-informed preprocessing, and acquisition-date structure can compromise the interpretation of nominally independent validation. They describe reconstructions and exploratory internal analyses, not corrected performance estimates for the published classifiers.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-17
DOI
https://doi.org/10.64898/2026.11.15.26368137
Primary Topic
Gene expression and cancer classification
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Hidden Dependencies in Endometriosis Microarray Classifiers: Sample Reuse, Outcome-Informed Batch Correction, and Diagnostic Reassessment

Abedalrhman Alkhateeb, Elena Krikun
medRxiv
Gene expression and cancer classification
preprint

Hidden Dependencies in Endometriosis Microarray Classifiers: Sample Reuse, Outcome-Informed Batch Correction, and Diagnostic Reassessment

Abedalrhman Alkhateeb, Elena Krikun
preprint en

Abstract

Distinct Gene Expression Omnibus (GEO) accessions do not necessarily represent independent transcriptomic samples. We investigated exact array reuse, access to test diagnoses during batch correction, and cohort and acquisition-date confounding in endometriosis classifiers. We compared decompressed CEL files using SHA-256, reconstructed a matrisome elastic-net workflow, changed test labels supplied to ComBat, and conducted 3,582 constrained training-label refits. We also evaluated the effect of removing matching training copies in a five-gene Huang-inspired random forest and performed a 324-probe nested internal benchmark within GSE51981. Of 37 test CEL files, 26 (70.3%) were identical to training arrays; all had been scanned in 2004-2005. Changing a test diagnosis supplied to ComBat moved the fixed classifier score toward the supplied label in all 37 interventions. In a deduplicated, batch-constrained comparison, median nominal balanced accuracy was 0.881 when ComBat received true test outcomes and 0.549 without them; the outcome-informed arm performed better in 1,193 of 1,194 matched refit comparisons. In the separate Huang-inspired reconstruction, removing each target's matching training copy lowered pooled AUC from 1.000 to 0.323 (conditional 95% interval 0.105-0.571). Exploratory nested AUC was 0.733 for endometriosis versus other gynecological conditions, and date-held-out AUC in 2009 was 0.698. For endometriosis versus healthy controls, a date-only out-of-fold model achieved AUC 0.935, compared with 0.893 for the molecular model. Together, these findings show how exact array reuse, outcome-informed preprocessing, and acquisition-date structure can compromise the interpretation of nominally independent validation. They describe reconstructions and exploratory internal analyses, not corrected performance estimates for the published classifiers.

medRxiv
Lakehead University (CA)
Good health and well-being
Gene expression and cancer classification
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon