EP180 - Development of an AI Model for Predicting Urgent Colonoscopy Findings Using Multimodal Clinical Data

Purpose/Background: Prediction of patients likely to have urgent colonoscopy findings (e.g., advanced adenomas or colorectal cancers) is critical for optimizing resource allocation in colorectal cancer screening. This study aims to develop machine learning models using demographic, laboratory, and clinical features to predict urgent colonoscopy findings, supporting a more prioritized screening process.

Methods/Interventions: A retrospective cohort of asymptomatic patients who underwent screening colonoscopy was analyzed to develop predictive models through supervised learning for colorectal cancer risk stratification. Urgent findings were defined as colorectal cancers or high-risk polyps (high-grade dysplasia, size >1cm, villous or tubulovillous histology, and intramucosal adenocarcinomas). Patients' colonoscopy outcomes were classified as 'urgent' or 'non-urgent' through Large Language Model (LLM) analysis of histopathology reports, with LLM classifications validated against clinician-labeled samples. Clinical history features such as smoking status were similarly extracted via LLM analysis of free-text clinical notes. Two XGBoost classifiers were developed: one using demographic, laboratory, and clinical predictors, and another incorporating Faecal Occult Blood Test (FOBT) results as an additional feature. Feature selection was conducted using mutual information scores, and model performance was evaluated via stratified 5-fold cross-validation. Contribution of each feature to model prediction was assessed using Shapley Additive Explanations (SHAP) analysis.

Results/Outcomes: The final cohort comprised 550 patients, of whom 107 had urgent findings. LLM classification of urgency achieved 99% accuracy when validated against clinician labeling of 100 randomly sampled patients. The baseline model without FOBT achieved a mean AUC of 0.5781 (Figure 1). Incorporation of FOBT in the second model modestly improved model discrimination, with a mean AUC performance of 0.6029 (Figure 2). SHAP analysis (Figure 1-2) of the top 10 contributing features revealed key shared predictors such as age, serum triglyceride level, total cholesterol level, MCV, MCHC and HbA1c abnormalities across both models. The baseline model uniquely emphasized high-sensitivity C-reactive protein, BMI, AST, and Vitamin D abnormalities, reflecting stronger reliance on metabolic and inflammatory markers and liver function indicators. In contrast, the FOBT-enhanced model emphasized FOBT and repeated FOBT measures, suggesting additional value from occult blood screening results.

Conclusion/Discussion: Machine learning models developed using multimodal clinical data may complement traditional FOBT testing and offer reasonable predictive performance for identifying patients at risk of more advanced neoplasms, serving as the foundation for a clinician-facing tool aiding colonoscopy prioritisation.

Figure 1
Receiver Operating Characteristic (ROC) Curve of Baseline Model (without FOBT results) and corresponding SHAP Summary Plot
Figure 2
Receiver Operating Characteristic (ROC) Curve of Second Model (with FOBT results) and corresponding SHAP Summary Plot