Bayesian sequential learning for prognostication in extremity soft tissue sarcoma (BayeSarc): a retrospective, multicentre cohort study.

Abstract

Sarculator is a widely validated prognostic tool that estimates overall survival and crude cumulative incidence (CCI) of distant metastasis in patients with resected soft tissue sarcomas in the extremities. Sarculator relied on external cohorts only for performance testing and could not incorporate new information or adapt to temporal changes. We aimed to develop BayeSarc, a prognostic model based on Bayesian sequential learning (BSL), which enables continuous updating by incorporating new clinical cohorts and provides more accurate estimates. In this retrospective, multicentre cohort study, eligible patients were adults (aged ≥18 years) with primary, localised, surgically treated soft tissue sarcomas in the extremities (excluding desmoid tumours, undifferentiated small round cell sarcoma of soft tissue, alveolar or embryonal rhabdomyosarcoma, dermatofibrosarcoma protuberans, and well differentiated liposarcoma). Data were retrieved from institutional databases at each participating hospital. BayeSarc used the same clinicopathological variables as Sarculator (age, size, grade, and histology) and was developed with a historical cohort of consecutive patients treated surgically at the Istituto Nazionale dei Tumori (Milan, Italy) and sequentially updated with five independent cohorts from Canada, France, the UK, USA, and Italy. Bayesian Cox (overall survival) and Fine-Gray (CCI distant metastasis) models were reformulated within a BSL framework combining Bayesian updating with prior-information adaptive borrowing. The primary objective was to compare the discrimination and calibration of BayeSarc versus Sarculator for predicting overall survival and CCI-distant metastasis. We evaluated the performance of BayeSarc at each update using prequential estimates, reflecting model transport to a new cohort without local recalibration, and post-update estimates, reflecting performance after sequential updating. We included a total of 4916 patients (2204 [44·8%] female, 2694 [54·8%] male, and 18 [0·4%] with sex not recorded) drawn from six cohorts: Istituto Nazionale dei Tumori, Milan, Italy (Jan 1, 1994-Dec 31, 2013; median follow-up 86 months [IQR 81-90]); Mount Sinai Hospital, Toronto, Canada (Jan 1, 1994-Dec 31, 2013; 85 months [81-90]); Institut Gustave Roussy, Villejuif, France (Jan 1, 1996-May 15, 2012; 75 months [68-82]); Royal Marsden Hospital, London, UK (Jan 1, 2006-Dec 31, 2013; 54 months [48-59]); Brigham and Women's Hospital, Boston, USA (Jan 1, 2014-Dec 31, 2021; 72 months [66-84]); and Istituto Nazionale dei Tumori, Milan, Italy (Jan 1, 2014-Dec 31, 2021; 61 months [57-64]). 12 patients from the UK were missing follow-up data and were excluded from survival analyses. At the final step of the BSL update, BayeSarc achieved higher discrimination than Sarculator for both overall survival (prequential mean C index: 0·784 [95% credible interval 0·759-0·794]; after update: 0·801 [0·790-0·809]; Sarculator: 0·773) and distant metastasis (prequential mean C index: 0·723 [0·704-0·738]; after update: 0·738 [0·730-0·743]; Sarculator: 0·718). Calibration improved consistently across updates, and uncertainty around estimates and predictions decreased. BayeSarc is a continuously updatable, accurate, and precise prognostic tool for soft tissue sarcomas in the extremities. It reduces uncertainty, adapts to temporal changes, and refines variable weights. Its incorporation into the Sarculator app enables immediate clinical use, with potential to improve patient counselling, guide treatment decisions, and refine trial design. More broadly, the BSL framework provides an innovative and generalisable approach to prognostication in rare cancers, moving beyond the traditional two-step development-validation paradigm, enabling more efficient use of patient data. Associazione Italiana per la Ricerca sul Cancro, Cancer Research UK, Fundacion Científica, and Asociacion Espanola Contra el Cancer.