P0515Development and validation of a multimodal prediction model for endoscopic activity in Crohn’s disease

Abstract Background Accurate assessment of disease activity is crucial in Crohn’s disease (CD). While endoscopy remains the reference standard, its frequent use carries discomfort and risks. Laboratory indicators demonstrate limited predictive capability, while validated imaging scores provide valuable information but are typically used independently. This study aims to systematically integrate these noninvasive tools to explore potential synergistic effects, thereby developing a predictive model and reliably identifying patients with active endoscopic disease. Methods This retrospective study (2020-2024) enrolled 106 CD patients from West China Hospital of Sichuan University. We collected demographic, clinical, and laboratory data within a 6-week window of endoscopy, defining endoscopic activity as Simple Endoscopic Score for Crohn’s disease (SES-CD) 2. We scored imaging using Magnetic Resonance Index of Activity (MaRIA) and International Bowel Ultrasound-Segmental Activity Score (IBUS-SAS). The cohort was randomly divided into an 80% training set for model building and a 20% validation set. Predictors were selected via bootstrap resampling (100 iterations), and a multivariate logistic regression model was developed using backward elimination. We assessed model discrimination with AUC and calibration with the Hosmer-Lemeshow test. Results Among the 106 CD patients, 68 (64.2%) had endoscopically active disease (SES-CD 2). Univariate analysis identified ANC, WBC, PLT, ALB, CRP, ESR, IBUS-SAS, and MaRIA as significant predictors, which were consolidated into a multivariate model via bootstrap selection. Multivariate analysis identified IBUS-SAS, ALB, CRP, MaRIA, ANC, and PLT as independent predictors of endoscopic activity. (Table 1). The final predictive model is: logit(P) = 22.0474 + 0.0562 × IBUS-SAS - 0.4424 × ALB + 0.6047 × CRP + 0.1825 × MaRIA + 0.19 × ANC - 0.0421 × PLT. The integrated model demonstrated exceptional discrimination in the training set (AUC=0.978, 95%CI:0.954-1.000) and good performance in the validation set (AUC=0.833, 95%CI:0.656-1.000) (Figure 1). Calibration was excellent (Hosmer-Lemeshow test p = 0.924). The optimal threshold determined based on Youden’s index was 0.9, yielding a sensitivity of 0.889 and an accuracy of 0.762 at this threshold. Conclusion In conclusion, the integration of IUS, MRI, and routine laboratory markers provides a robust multimodal tool for noninvasive prediction of endoscopic activity in CD. Our model demonstrated excellent performance in internal validation. We are currently collecting prospective data for independent external validation, and future multicenter studies will further explore its generalizability and clinical translation. Conflict of interest: Li, Zhaoyang: Zhao, Jieying: No conflict of interest Cai, Lei: No conflict of interest Luo, Mengqi: This author declare no conflicts of interest. Zhong, Rui: No conflict of interest Jv, Jingyi: No conflict of interest Xu, Hongzhen: No conflict of interest Chen, Xiaoting: No conflict of interest Axi, Libumu: No conflict of interest Zhang, Qiuyue: No conflict of interest Wang, Yufang: No conflict of interest