P283 Patient-reported symptoms over a period of 14 days reliably predict endoscopic and histological disease activity in ulcerative colitis (UC)

Abstract Background Disease activity assessment in UC is most accurately evaluated by endoscopy and biopsy, which often correlates poorly with current symptoms. We investigated whether a machine learning classification algorithm using patient-reported Simple Clinical Colitis Activity Index (SCCAI) over 14 days and signature features could reliably distinguish endoscopic and histopathologic activity from remission. Methods The TrueColours ulcerative colitis (UC) monitoring platform was used to collect symptoms (SCCAI, daily) and endoscopic/histopathological activity, as per the UCEIS and Nancy indices (twice in 6 months) in 233 patients. The longitudinal data may be seen as a trajectory. The signature is a collection of statistics which efficiently summarises the trajectory and serves as a non-parametric hierarchical method for longitudinal data representation. Signature features were used as input to the extreme gradient boosting classification algorithm to categorise subjects into remission or active disease groups. Remission: UCEIS ≤1 AND Nancy ≤1; Active disease: UCEIS ≥4 AND Nancy ≥2. The advanced signature-based approach was compared with a baseline model with manually constructed features (mean SCCAI score over 14 days preceding endoscopy). Cross-validation was used to report metrics, Table 1. Table 1: Disease activity classification results using 14 days of SCCAI Mean Signatures AUC 0.79 (0.72–0.86) 0.87 (0.82–0.93) Sensitivity 0.87 (0.80–0.94) 0.89 (0.83–0.95) Specificity 0.56 (0.46–0.68) 0.78 (0.68–0.87) Positive Predicted Value 0.69 (0.63–0.75) 0.82 (0.75–0.88) Negative Predicted Value 0.80 (0.71–0.89) 0.87 (0.80–0.94) Accuracy 0.72 (0.66–0.79) 0.84 (0.76–0.90) Results Disease activity classification with signature features significantly outperformed mean-SCCAI based approach (AUC 0.87 vs. AUC 0.79 correspondingly, p = 0.011, Figure 1). All other statistical metrics (including sensitivity (0.89), specificity (0.78), PPV (0.82) and NPV) demonstrated the same trend, favouring the signature method. The decision-tree-based ensemble gradient boosting classification algorithm (XGBoost) has the intrinsic ability to rank feature importance. The signature feature of the fourth item within the SCCAI (blood in stool) dominated, followed by cross-correlation between the first (daily bowel frequency) and fourth item. Figure 1: Area under the ROC curve of both approaches (mean SCCAI vs. signatures) Conclusion Gradient boosting classification algorithm with sig