Machine-Learning Model Identifies New Diagnostic Criteria for Beckwith-Wiedemann Spectrum

Abstract

ObjectiveBeckwith-Wiedemann spectrum (BWSp) is an overgrowth and cancer predisposition disorder caused by genetic and epigenetic alterations of chromosome 11p15. The 2018 international consensus produced a clinical scoring system to capture the phenotypic variability of BWSp and guide genetic testing and clinical management, including tumor screening, in patients without molecular confirmation. In this study, we evaluated BWSp predictors to identify the most informative features. MethodsSupervised machine learning analyzed 25 phenotypic features in 555 patients with BWSp and 150 controls. Logistic regression, combined with a purposeful stepwise selection algorithm, identified a subset of features that can accurately classify subjects. Model performance was evaluated in a testing set and validated externally. ResultsThe final model included six predictors: macroglossia, lateralized overgrowth, midface flattening, hepatomegaly, omphalocele, and developmental delay. Developmental delay was the only negative predictor; macroglossia (OR 46.10) and lateralized overgrowth (OR 27.87) were the strongest predictors. The proposed model and 2018 system did not differ in classification performance for testing (P = .39) or external (P = .15) sets. ConclusionA simplified diagnostic model, driven by macroglossia and lateralized overgrowth, differentiates between patients with BWSp and controls with performance comparable to the 2018 system. And may help physicians prioritize BWSp evaluation.