A transparent 13-gene transcriptomic candidate signature for cervical neoplastic spectrum across GEO cohorts

Transcriptomic classifiers can prioritize molecular signals associated with cervical neoplastic lesions, but high-dimensional feature selection is vulnerable to overfitting, information leakage, and cohort-selection bias. We performed a retrospective computational reanalysis of six Gene Expression Omnibus microarray cohorts retained from the original project, with cervical tissue or epithelial expression profiles, extractable binary labels, and harmonizable gene annotation. GSE63514, GSE67522, and GSE9750 formed the development cohort (n = 200; 127 case samples and 73 control samples), whereas GSE64217, GSE6791, and GSE7803 were reserved for external evaluation. A glmBoost-derived 13-gene panel was coupled to a gradient boosting machine (GBM). The supervised stages of internal evaluation used 10 repeats of stratified five-fold nested cross-validation; scaling, feature selection, and GBM tuning were repeated within every outer-training split. The inherited expression matrix had undergone cohort-wide ComBat adjustment before this reconstruction, so the analysis does not establish fully end-to-end foldwise preprocessing validation. The aggregated out-of-fold AUC for the reconstructed selection pipeline was 0.9161 (95% CI 0.8731–0.9591). External discrimination was high but cohort dependent; the nominal AUC of 1.0000 in the six-sample GSE64217 cohort was too imprecise to support a claim of perfect performance. Five genes were selected in all 50 outer models. In a post hoc fixed-panel sensitivity analysis, the 13-gene and five-gene panels had similar repeated-CV out-of-fold AUCs (0.924 and 0.931; paired DeLong P = 0.390). The signature is therefore a candidate computational model for a heterogeneous cervical neoplastic spectrum. It does not establish biological validation, comparative clinical utility, or a deployable single-patient assay.