A Coupled Bayesian Sequential Decision-Support Framework for Personalising Fertility- and Uterus-Sparing Management of Early-Stage Endometrial Cancer: Calibration, Discrimination, and the Structure of Longitudinal Evidence Accumulation A Theoretical Hypothesis and Modelling Perspective

Background. Hysterectomy remains the oncologically definitive treatment for endometrioid endometrial cancer, but it permanently ends reproductive capacity. Progestin-based fertility-sparing treatment (FST) is recommended by international guidelines for carefully selected patients and achieves a pooled complete-response (CR) rate near 94%, yet is applied with essentially one-size-fits-all decision criteria despite strong evidence that molecular subtype and biomarker trajectory modulate response. Objective. We propose — as a theoretical, hypothesis-generating contribution, not a validated clinical protocol — a Bayesian sequential decision-support framework that integrates baseline molecular classification (POLE-mutated, mismatch-repair-deficient [MMRd], no specific molecular profile [NSMP], p53-abnormal), immunohistochemical markers (progesterone receptor, Ki-67, PTEN), and serial serum HE4 into a posterior probability of complete response that updates at each follow-up visit. Methods. We formalise the framework mathematically and, unlike prior purely elicited constructions, generate the ground-truth CR label of each synthetic patient from the same latent logistic structure that the fitted model targets, so that calibration is a genuine test rather than a tautology. Model coefficients are estimated from a seeded virtual calibration cohort (n = 4000) by maximum-a-posteriori inference with a Gaussian prior and a Laplace posterior; the cohort's subtype prevalence and subtype-conditional CR targets are calibrated to published, cited, aggregate statistics. We report calibration (reliability, Brier score, calibration slope and intercept), discrimination (area under the ROC curve), decision-curve net benefit, a Saltelli-scheme Sobol global sensitivity analysis with bootstrap confidence intervals, and a controlled comparison of two competing longitudinal-evidence update rules. Results. On an independent hold-out cohort the estimated model is well calibrated (calibration slope 0.97, intercept 0.12, expected calibration error 0.009, Brier score 0.068) and discriminates moderately (AUC 0.855). Two non-obvious structural findings emerge. First, longitudinal marker trajectory adds no discrimination beyond the calibrated baseline predictor (AUC 0.848 at every follow-up visit), because subtype-conditional decay renders the trajectory largely redundant given baseline. Second, the choice of evidence-accumulation rule, not any marker weight, governs the decision rule's stability: a naive rule that measures each marker decline against a fixed baseline and inflates its weight over time manufactures a spurious upward drift and a 22 percentage-point collapse of specificity between month 6 and month 12, whereas a corrected rule that accumulates incremental between-visit evidence limits that collapse to 8 points. The decision threshold dominates global sensitivity (total-order Sobol index ST ≈ 0.86), not serum HE4. Conclusion. The framework is mathematically coherent, internally consistent, properly calibrated under simulation, and yields falsifiable, prospectively testable predictions; but it is a decision-support hypothesis requiring dedicated prospective validation before any clinical use. It does not itself constitute new clinical evidence and must never substitute for multidisciplinary oncological judgement or informed patient choice.