Beyond Spigelman: time-adaptive duodenal cancer risk in familial adenomatous polyposis – results from a retrospective two-centre study (DRACO)

Background Duodenal cancer is a leading cause of cancer-related mortality in familial adenomatous polyposis (FAP), yet the Spigelman staging system—developed with equal weighting of its components and without formal outcome validation—poorly discriminates short-term progression risk. Objective To derive and validate two reweighted risk models for predicting progression to advanced duodenal neoplasia in FAP and to assess their transportability across institutions. Design Duodenal Risk assessment for APC surveillance intervals Optimisation (DRACO) is a retrospective, Strengthening the Reporting of Observational Studies in Epidemiology (STROBE)-compliant and Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)-compliant cohort study of 914 upper endoscopies from 362 genetically confirmed FAP patients treated at two tertiary referral centres in Milan, Italy. Data were partitioned into a training set, an internal validation set drawn from the same institution and a non-overlapping testing set from a second institution. Two multivariable logistic regression models—Spigelman 2.0 (reweighted canonical variables) and Milan (Spigelman 2.0 plus age and desmoid tumour status)—were derived to predict advanced duodenal neoplasia (high-grade dysplasia or carcinoma) at 1, 3 and 5 years. Results Both models improved 1-year sensitivity over the original Spigelman score (29.7% vs 75.7% (Spigelman 2.0) and 81.1% (Milan), p<0.05), without a significant loss in specificity and improved 3-year and 5-year specificity (up to 96.9% and 95.0%, respectively, p<0.05) without loss of sensitivity. Calibration was robust in all cohorts (Brier scores <0.05). Performance was consistent between the training institution’s internal validation set and the second institution’s testing set, although CIs were wide for several risk strata, reflecting a limited number of outcome events. Conclusion The DRACO models provide a biologically grounded, time-adaptive framework for tiering FAP surveillance and outperform Spigelman in this two-centre retrospective cohort. Given the retrospective design, limited geographic validation and few events at longer horizons, prospective multi-country validation is needed before clinical adoption. Trial registration number NCT07199127.