Few-shot learning for expert consensus severity stratification in newly diagnosed crohn’s disease

Early and accurate assessment of disease severity is important for standardizing baseline risk stratification and supporting treatment planning in newly diagnosed Crohn’s disease. This study used electronic medical record data, a few-shot learning algorithm, and SHAP-based interpretability analysis to develop a decision-support framework for standardized expert-consensus cross-sectional severity labels at initial diagnosis of Crohn’s disease. Electronic medical records of 106 patients with newly diagnosed CD were included at the Zhongshan Hospital of Xiamen University. Three few-shot learning models: metric learning, matching networks, and prototypical networks were constructed with expert-consensus reference labels. For standardized evaluation, each trained feature embedding backbone was followed by a logistic-regression classifier fitted on training-set embeddings and applied to the corresponding validation fold. Key features were identified via SHapley Additive exPlanations (SHAP) value and Pearson correlation analysis. Model performance (accuracy, macro-F1, and AUC) was evaluated across feature combinations and preliminarily assessed in a temporally independent single-center validation cohort ( n = 23). In nested cross-validation, prototypical networks showed the highest mean macro-F1 (0.817), followed by matching networks (0.812) and metric learning (0.796). The prototypical network was emphasized in downstream feature-ablation analysis. After removal of CTE-derived features, it retained acceptable AUCs for mild and moderate disease in the temporally independent single-center validation cohort, but the severe-class AUC decreased to 0.842, indicating the continued importance of CTE-derived structural information in severe CD. This prototypical network-based few-shot learning framework showed preliminary ability to reproduce expert-consensus cross-sectional severity stratification of Crohn’s disease at first diagnosis. Its performance estimates remain subject to incorporation bias because some of the same clinical information informed both the reference labels and model predictors, warranting further evaluation in larger independent cohorts.