Predictive models for endoscopic disease activity in patients with ulcerative colitis: Practical machine...

Area of Science:

• Gastroenterology
• Medical Informatics
• Computational Biology

Background:

• Endoscopic monitoring is crucial for ulcerative colitis (UC) management.
• Current non-invasive methods lack effectiveness in predicting endoscopic disease activity.
• Reducing endoscopic procedures and associated costs is a clinical need.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting endoscopic disease activity in UC patients.
• To establish a non-invasive tool for assessing UC severity.
• To potentially decrease the frequency of invasive endoscopic examinations.

Main Methods:

• Retrospective study of 420 UC patients (January 2016 - January 2021).
• Collected 39 clinical and laboratory variables, categorized by Mayo Endoscopic Score (MES) or Ulcerative Colitis Endoscopic Index of Severity (UCEIS).
• Applied logistic regression and four ML algorithms (Random Forests, XGBoost), with feature selection and SMOTE, evaluated using AUC, accuracy, sensitivity, precision, and F1 score. SHAP for interpretability.

Main Results:

• Random Forests (23 variables) achieved AUC 0.8192 for MES prediction.
• XGBoost (21 variables) achieved AUC 0.8006 for UCEIS prediction.
• Albumin, rectal bleeding, and CRP/ALB ratio were key predictors identified by SHAP analysis.

Conclusions:

• ML models offer a promising non-invasive method for predicting endoscopic disease activity in UC.
• Random Forests and XGBoost demonstrate suitability for data-driven endoscopic activity prediction in UC.
• This approach could aid in optimizing patient management and reducing healthcare burdens.