Skip to main content
SSO 2026 Annual Meeting Main banner

P19. BayeSarc: Bayesian Update of Sarculator for Primary Retroperitoneal Sarcoma

Thursday, March 5, 2026
11:14 AM - 11:21 AM MST
Location: 101A-C
Introduction:  The objective of this study was to develop a geotemporally sensitive update of Sarculator (BayeSarc) for application to patients with primary retroperitoneal sarcoma (RPS), using Bayesian Sequential Learning (BSL). Unlike Sarculator, which is built on a traditional two-phase development–validation framework, using one cohort for model training and others for external validation, BayeSarc employs sequential updating to integrate information from each new cohort. This allows the model to dynamically adjust hazard ratios (HRs) and capture temporal or cohort-related variations in prognosis, thereby optimizing the use of available data.

Methods: BayeSarc applies a BSL adaptation of the Cox proportional hazards model to predict overall survival (OS) and disease free survival (DFS) in patients with primary localized RPS treated with surgery. Starting from the original Sarculator cohort (Milan, 2010-2017), BayeSarc sequentially integrated seven additional cohorts of patients with primary RPS treated with surgery during the same study period (London, Paris-IGR, Toronto, Boston, Warsaw, Paris-Curie, Amsterdam). At each iteration, the posterior distribution of all parameters becomes the prior for the subsequent step. This approach efficiently leverages prior knowledge, continuously updates the model, and simultaneously provides external validation at each step. BayeSarc performance at 5 years post-resection was evaluated using Harrell’s C-index and Brier score (BS), and compared with Sarculator.


Results: Overall 1101 patients from eight centers were included. BayeSarc was based on the Sarculator variables: age, tumor size, FNCLCC grade, histology, multifocality, completeness of resection and center volume (high volume if ≥13 patients/year). Compared with Sarculator, BayeSarc achieved equal or superior discrimination (higher C-index) and calibration (lower BS), as shown in Table 1.

For OS, the C ranged from 0.610-0.796 (mean 0.706) for Sarculator vs 0.674-0.821 (mean 0.741) for BayeSarc; the 5-year BS ranged 0.155-0.238 (mean 0.191) vs 0.144-0.201 (mean 0.167). DFS performance metrics were similarly impacted (Table 1).



Moreover, BayeSarc substantially reduced uncertainty: credibility intervals were ~60% narrower for the C-index (OS/DFS) and ~26% (OS) and ~69% (DFS) narrower for the 5-year BS.



Conclusions: BayeSarc provides more accurate and precise predictions for both OS and DFS in patients with primary RPS compared with Sarculator. It reduces uncertainty, refines variable weighting, and enhances model generalizability across centers.

Learning Objectives:

  • Understand the concept and methodological principles of Bayesian Sequential Learning and its application to clinical prognostic models.
  • Recognize how BayeSarc improves upon traditional modeling approaches like Sarculator by integrating new data iteratively rather than relying on separate validation cohorts.
  • Appreciate the potential of Bayesian updating to enhance prognostic precision and adaptability in rare cancers such as retroperitoneal sarcoma.