Abstract Background Diagnostic delay for Crohn’s disease in China is significant (median 34 months). To improve early detection, this study aims to develop and deploy an online, interpretable machine learning model using prospective cohorts. The tool will stratify risk based on baseline clinical features to aid screening centers and clinics in identifying high-risk individuals for further evaluation. Methods This study was conducted in four main steps: (1) identification of 19 clinical characteristic variables through literature review and consultation from 20 IBD experts; (2) enrollment of study subjects according to predefined criteria, with collection of symptom recall data (CD patients recalling symptoms from 6 months prior to diagnosis, others reporting recent symptoms) and laboratory indicators; (3) development of a diagnostic model using nine machine learning algorithms based on the selected variables, along with SHAP-based interpretation; and (4) completion of model validation and development of a web-based application. Results This study identified 14 clinical features significantly more frequent in Crohn’s Disease (CD) patients than non-CD individuals. Nine machine learning models were evaluated for CD prediction. Comprehensive evaluation demonstrated the XGB model as superior. It achieved top performance on both training and test sets across key metrics including accuracy, sensitivity, Negative Predictive Value (NPV), Youden’s index, and Area Under the Curve (AUC). Calibration curves and Decision Curve Analysis (DCA) confirmed XGB’s better calibration and superior clinical utility across all probability thresholds compared to alternatives like ADB. Crucially, XGB significantly outperformed existing screening methods (fecal calprotectin, red flag index and their combination). External validation on an independent cohort reinforced XGB’s robustness, showing the highest accuracy (0.907), sensitivity (0.956), NPV (0.965), and Youden’s index (0.829). While the RFI model had higher specificity, XGB’s markedly superior sensitivity is vital for minimizing false negatives in screening. This balance confirms XGB as a reliable tool for CD risk stratification in new populations. To facilitate clinical use, the final XGB model has been implemented as a web-based calculator. Inputting values for the 14 features generates an early CD risk prediction, offering a practical tool to reduce diagnostic delays. Conclusion The XGB model is a superior, robust tool for early Crohn’s disease detection, demonstrating high performance and key predictive features like perianal lesions and weight loss. Its implementation as a web calculator facilitates clinical use to reduce diagnostic delays. Conflict of interest: Mr. Kong, Zihao: No conflict of interest zheng, chang: No conflict of interest zhang, xiaoqi: No conflict of interest Peng, Chunyan: No conflict of interest
Kong et al. (Thu,) studied this question.