An interpretable Hallmark pathway activity classifier for distinguishing CIN3/HSIL from invasive cervical squamous carcinoma: Development, external validation, and feature stability assessment.

Abstract

Public transcriptomic cohorts rarely include paired biopsy, conization, and final surgical pathology labels, making direct modeling of post-conization pathological upgrading infeasible in most public datasets. We therefore examined whether pathway-level transcriptomic activity can distinguish CIN3/HSIL from invasive cervical squamous carcinoma. GSE63514 was used for model development and GSE7803 as the primary external validation cohort. Expression matrices were mapped to gene symbols and summarized as MSigDB Hallmark pathway activity scores using ssGSEA. An elastic net logistic regression classifier was trained in GSE63514 and applied to GSE7803 without refitting or threshold re-optimization. Bootstrap resampling was used to assess feature-selection stability. The final model retained eight Hallmark pathways. In GSE63514, the classifier achieved an AUC of 0.890 (95% CI, 0.810-0.971). In GSE7803, the locked model achieved an AUC of 0.815 (95% CI, 0.665-0.966). Bootstrap analysis showed recurrent selection of the major contributing pathways, including estrogen response early, KRAS signaling DN, TGF-beta signaling, estrogen response late, and epithelial-mesenchymal transition. This study provides an interpretable pathway-level transcriptomic classifier that separates preinvasive high-grade cervical disease from invasive squamous carcinoma in public cohorts. The model should be interpreted as a molecular characterization framework rather than a clinically deployable diagnostic assay.