PP059: Prediction Of Post-Colonoscopy Colorectal Cancer In Statin Users With Machine Learning Model

Abstract

<p>Background & Aims: While statins may have chemopreventive effect on CRC, certain individuals remain at risk. We aimed to develop a machine learning (ML) model to predict postcolonoscopy (PCCRC) risk in statin users, and to determine high-risk subgroups. <br></p><p>Methods This was a retrospective cohort study based on a territory-wide electronic healthcare database in Hong Kong. Individuals aged at least 40 years who had undergone colonoscopy between 2005 and 2013, and had used statins for at least 90 days within 5 years before index colonoscopy were recruited. Statin users were divided into training (n=20,358) and validation set (n=5,089) in a random selection with bootstrap method of training of ML model. Data was used to construct CatBoost model to predict outcome of PCCRC-3y (defined as CRC diagnosed between 6 months and 3 years after index colonoscopy). The variables input into the CatBoost model included age, sex, history of colonic polyps, biopsy/polypectomy at index colonoscopy, smoking- and alcohol-related diseases, 12 comorbidities (including cardiovascular diseases, dementia and parkinsonism), use of concurrent medications (NSAIDs/aspirin/COX-2 inhibitors and metformin), annual colonoscopy volume and polypectomy rate of the center. The performances of the CatBoost model was assessed by the area under receiver operating characteristic curve (AUC) analysis. <br></p><p>Results 25,447 subjects, 114 (0.45%) developed PCCRC-3y. The AUROC of the Catboost Model was 0.77 (95% CI: 0.82-0.86) with a sensitivity of 74% (56-92%), specificity of 79% (77-80%), positive predictive value of 1.7% (0.9-2.5%) and negative predictive value of 99% (99-99%). Feature extraction showed that age (16.2%) and biopsy/polypectomy at index colonoscopy (14.0%) were two most important factors influencing CatBoost model, while other factors contribute to less than 5% individually. <br></p><p>Conclusions The Catboost model has satisfactory performance in predicting PCCRC-3y risk among statin users. Older age and history of colonic polyps/polypectomy are the most important factors in determining PCCRC-3y risk.<br></p>