{
  "abstract": "Background Accurate identification of mucosal healing, defined as Mayo Endoscopic Score (MES) 0 is the primary treatment target in UC and a key endpoint in clinical trials. Manual remission assessment carries ~30% inter-observer variability, and access to expert endoscopists is inequitable across healthcare settings. We developed and validated a binary AI classifier to detect endoscopic remission from colonoscopy images.Methods An EfficientNet-B3 model was trained on 9,590 images (LIMUC dataset, 564 patients) to classify images as remission (MES 0) versus active disease (MES 1-3). Performance was evaluated on a held-out test set (n=1,686; 925 remission, 761 active) using AUC, sensitivity, specificity, PPV, and NPV. Confidence distributions were analysed to characterise decision boundary behaviour.Results The classifier achieved AUC=0.947, sensitivity=92.4%, specificity=80.1%, PPV=79.3%, and NPV=92.7% (IDDF2026-ABS-0258 Figure 1. ROC curve for remission vs active disease classification, IDDF2026-ABS-0258 Figure 3. Clinical performance summary across all metrics). The model correctly identified 703/761 active disease cases and 741/925 remission cases (IDDF2026-ABS-0258 Figure 2. Confusion matrix for binary remission classification on held out test set). The high NPV confirms that AI-confirmed remission is highly reliable, making it particularly suited to rule out active disease in treat-to-target monitoring.Confidence distributions demonstrated clear bimodal separation between remission and active disease predictions, with uncertainty appropriately concentrated at the MES 0|1 boundary (IDDF2026-ABS-0258 Figure 4. Prediction confidence distribution for true remission and true active disease), consistent with known inter-observer difficulty at this clinically significant threshold. The precision-recall curve (AP=0.941) confirmed sustained high precision across the full sensitivity range, with the operating point achieving PPV=79.3% and NPV=92.7% (IDDF2026-ABS-0258 Figure 5. Precision recall curve for remission classification), a clinically favourable trade-off prioritising safe remission confirmation over false reassurance.Conclusions An EfficientNet-B3 AI remission classifier achieves AUC=0.947 with 92.4% sensitivity and NPV=92.7%, supporting its role as a reliable screening tool to confirm mucosal healing in UC. The strong NPV makes it particularly suited for treat-to-target monitoring, reducing unnecessary endoscopy burden while maintaining safety. Uncertainty is appropriately highest at the MES 0 -1 boundary, consistent with the biological difficulty of this threshold. Prospective validation in diverse clinical settings is warranted to support equitable deployment across under-resourced IBD centres.Abstract IDDF2026-ABS-0258 Figure 1Abstract IDDF2026-ABS-0258 Figure 2Abstract IDDF2026-ABS-0258 Figure 3Abstract IDDF2026-ABS-0258 Figure 4Abstract IDDF2026-ABS-0258 Figure 5",
  "authors": [
    {
      "affiliations": [
        "New York Medical College, United States"
      ],
      "name": "Jeril Lasington"
    },
    {
      "affiliations": [
        "Rutgers University, United States"
      ],
      "name": "Lawin Steve Mathew Lasington"
    },
    {
      "affiliations": [
        "Boston University, United States"
      ],
      "name": "Swamynathan Umamaheshwaran"
    }
  ],
  "title": "IDDF2026-ABS-0258 Can AI standardise mucosal healing assessment in ulcerative colitis?",
  "uid": "21d225f5-1f90-5351-9abc-11fa733b2ad1"
}