Unsupervised 20-feature immunohistochemical immune profiling identifies exploratory cervical cancer phenotypes with unfavorable survival signals

Abstract

Objective To determine whether unsupervised integration of routinely accessible immunohistochemical immune features can identify clinically relevant tumor immune microenvironment phenotypes in cervical cancer. Methods We performed a retrospective exploratory analysis of surgically treated cervical cancer cases with complete scoring for a 20-feature immunohistochemical immune panel. The features were derived from approximately 11 antibody-defined markers, including separate intratumoral (IT) and stromal (ST) scores where applicable. Hierarchical clustering was performed using standardized feature values, Euclidean distance, and Ward.D2 linkage. The three-cluster solution was the primary unsupervised analysis. The comparison between C2 and combined C1/C3 was an exploratory, outcome-informed contrast generated after inspection of the primary three-cluster survival curves. Cluster-defining features, clinicopathological associations, overall survival, cause-specific Cox models, and competing-risk endpoints were assessed. Results The source dataset included 226 patients; 177 had pT1b-region disease and 169 had complete data for the 20-feature panel. Three tumor immune microenvironment clusters were identified: C1 (n=50), C2 (n=78), and C3 (n=41). The most discriminatory features were CD8-ST, CD8-IT, CD3-ST, PD-1-ST, PD-1-IT, CD3-IT, CD4-ST, CD68-IT, CD20-ST, and CD57-ST. C2 showed relative depletion of several stromal and intratumoral lymphoid and checkpoint-associated features. In the exploratory C2 versus C1/C3 contrast, C2 was associated with inferior overall survival by Kaplan-Meier analysis (log-rank p=0.012) and univariable Cox regression (HR 2.14, 95% CI 1.16-3.92, p=0.014). After adjustment for age, nodal status, and histology, the association was attenuated and borderline (HR 1.84, 95% CI 0.99-3.42, p=0.055). Fine-Gray analysis showed a similar adverse direction for neoplasm-specific death (SHR 2.37, 95% CI 0.95-5.87, p=0.063). Conclusion Routine compartment-aware immunohistochemical immune profiling identified exploratory cervical cancer immune phenotypes, including an immune-depleted C2 cluster with unfavorable survival signals. These findings are hypothesis-generating and require independent validation before clinical implementation.