{
  "abstract": "Background Early identification of short-term, phenotype-­specific complications remains a key unmet need in Crohn’s ­disease (CD). Existing prediction tools often focus on single endpoints and provide limited interpretability, restricting ­clinical adoption. We developed and externally validated an explainable machine-learning framework to predict three one-year complication phenotypes after discharge: bowel resection (Task 1), anal complications (Task 2), and abdominal complications (Task 3) ( IDDF2026-ABS-0097 Figure 1. Study design concept. This study integrated LASSO selection and SMOTETomek resampling to identify the Random Forest model as the optimal classifier among seven candidates).Methods A two–center CD cohort from Zhongshan Hospital of Xiamen University and the First Affiliated Hospital of Jinan University was utilized for model development and validation. Data were randomly partitioned into training and test sets with 10–fold cross–validation. Feature selection involved the LASSO algorithm. Seven ML models were benchmarked with ROC, accuracy, recall, precision, specificity, and negative predictive value. Precision–Recall curves, calibration curves and decision curve analysis (DCA) were also analyzed for model utility. Model interpretability was assessed using SHAP, with SHAP-based nomograms, heatmaps and a web-based calculator for clinical application. We further examined how Montreal anatomical phenotypes and CDAI score relate to the risk of the three complications.Results One-year incidences were 18.4%, 38.4%, and 22.7% respectively. Random Forest (RF) achieved the best overall ­performance, with an AUROC of 0.88 (95% CI 0.85–0.91) for bowel resection and a favorable net benefit on DCA ( IDDF2026-ABS-0097 Figure 2. Model performance evaluation for tasks 1-3 in the test cohort mean of 10-fold cross-validation, IDDF2026-ABS-0097 Figure 3. Web-based calculators for individual risk prediction of tasks 1-3). SHAP identified task-specific dominant predictors: Bowel Resections (Task 1)—disease behavior, nutritional support therapy (NST), and disease duration; Anal Complications (Task 2)—prior anal complication history, extra-intestinal manifestations, and CDAI score; Abdominal Complications (Task 3)—prior abdominal complication history, NST, and Montreal L3 phenotype. Montreal phenotype analyses showed marked heterogeneity, with isolated small-bowel involvement exhibiting higher bowel resection and abdominal complication rates, whereas isolated colonic disease showed the highest perianal complication burden.Conclusions An interpretable RF–SHAP framework enables accurate, externally validated prediction of three distinct one-year post-discharge CD complication phenotypes from routinely available clinical data. In addition to strong discrimination and favourable decision-curve net benefit, the framework offers transparent, patient-level explanations with practical visual outputs, supporting risk stratification at the point of care. Integration with Montreal phenotype and CDAI score heterogeneity analyses underscores clinically meaningful subgroup variation, informing phenotype-tailored surveillance, timely therapeutic optimization and multidisciplinary decision-making in routine practice.Abstract IDDF2026-ABS-0097 Figure 1Abstract IDDF2026-ABS-0097 Figure 2Abstract IDDF2026-ABS-0097 Figure 3",
  "authors": [
    {
      "affiliations": [
        "Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian Province, China"
      ],
      "name": "Kun Xia"
    },
    {
      "affiliations": [
        "Gastroenterology Department, First Affiliated Hospital of Jinan University, Guangzhou, Guangdong Province, China"
      ],
      "name": "Ying Shi"
    },
    {
      "affiliations": [
        "Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian Province, China"
      ],
      "name": "Shuntian Cai"
    },
    {
      "affiliations": [
        "Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian Province, China"
      ],
      "name": "Jianlin Ren"
    },
    {
      "affiliations": [
        "Department of Pathology, Tangshan Gongren Hospital, Tangshan, Hebei Province, China"
      ],
      "name": "Qingan Xia"
    },
    {
      "affiliations": [
        "Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian Province, China"
      ],
      "name": "Hongzhi Xu"
    },
    {
      "affiliations": [
        "Department of Gastroenterology, The National Key Clinical Specialty, Zhongshan Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian Province, China"
      ],
      "name": "Yanyun Fan"
    }
  ],
  "title": "IDDF2026-ABS-0097 Interpretable machine learning for prediction of one-year complications in crohn’s disease: a multi-center, retrospective study",
  "uid": "3dbb3779-733a-5941-8cbd-94b342d0a33f"
}