Hematological Versus Clinical Predictors of Tumor Grade in Endometrial Cancer: A Comparative Machine Learning Study with SHAP-Based Explainability

Background: Systemic inflammatory indices derived from routine complete blood counts have been proposed as inexpensive biomarkers in gynecological malignancies. Their ability to discriminate tumor grade in endometrial cancer, however, has rarely been evaluated against routinely available clinical variables under rigorous validation. Methods: We retrospectively analyzed 225 women with endometrial cancer (166 low-grade, G1–G2; 59 high-grade, G3). Nine inflammatory indices were computed. Four feature sets were compared using a prespecified elastic-net logistic regression under repeated nested cross-validation (5 outer × 5 inner folds, 20 repetitions) with 5000 stratified bootstrap confidence intervals: a collinearity-reduced inflammatory panel, the full inflammatory panel, preoperative clinical variables (age, CA-125 and preoperative albumin), and their combination. Selection among nine classifiers was retained as an exploratory analysis. Because 66% of high-grade events were non-endometrioid, every analysis was repeated in three cohorts: the full cohort, endometrioid tumors only, and the subgroup with assessed hormone receptor status. Histologic subtype was deliberately excluded because non-endometrioid carcinomas are high-grade by definition. Model behavior was interpreted with TreeSHAP and cross-checked against permutation importance. Results: Of nine inflammatory indices, only the eosinophil-to-lymphocyte ratio remained statistically associated with grade after false discovery rate correction (AUC 0.626, 95% CI 0.543–0.705; q = 0.036), with all others showing negligible effect sizes (|r| < 0.08); its discrimination was nonetheless modest and insufficient for individual-level prediction. The inflammatory panel achieved AUC 0.533 (95% CI 0.443–0.621), within the range attainable by chance in this design (permutation p = 0.254). Preoperative clinical variables reached AUC 0.687 (95% CI 0.590–0.772) in the full cohort but 0.390 (0.257–0.524) in endometrioid tumors alone and 0.627 (0.485–0.765) in the receptor-assessed subgroup. Adding inflammatory indices to clinical variables did not improve discrimination (ΔAUC = −0.025, 95% CI −0.062 to +0.012; one-sided upper bound +0.007 against a prespecified margin of +0.05). The results were unchanged under an alternative grade dichotomization and after excluding patients with inconsistent blood-count entries. Conclusions: Most inflammatory indices did not demonstrate clinically useful discrimination between G3 and G1–G2 disease in this cohort, and did not add incremental discriminative value beyond the prespecified margin over routinely available preoperative clinical variables. The apparent discrimination of the clinical variables in the full cohort did not survive restriction to endometrioid histology, indicating that it reflected histologic composition rather than grade. Decision curve analysis on recalibrated probabilities showed that no model exceeded the treat-all strategy at thresholds below the prevalence of high-grade disease, and that, above it, any advantage over treating no one was small in the full cohort and absent within endometrioid tumors.