Development of a web-based tool using machine learning algorithms to improve adherence to diagnostic colonoscopy among middle-aged and elderly adults in colorectal cancer screening program.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
BACKGROUND: Colorectal cancer represents a major concern in China, coupled with a low rate of diagnostic colonoscopy, emphasising the necessity for improved adherence to diagnostic colonoscopy. OBJECTIVE: We aimed to predict colonoscopy behaviour in middle-aged and elderly adults using machine learning, explore determinants, and develop a web-based tool to improve adherence to diagnostic colonoscopy. METHODS: We used 49,418 consultation data with positive results of colorectal cancer preliminary screening (defining a positive result in risk assessment or FOBT screening as a preliminary screening positive) from a total of 202,091 consultations in a community-based colorectal cancer screening program between 2013 and 2021 from Baoshan District, Shanghai, China. We established logistic regression (LR), adaptive boosting classifier (Adaboost), eXtreme Gradient Boosting (XGBoost), random forest (RF), gradient boosting machine (GBM), and multilayer perceptron (MLP) to predict adherence to diagnostic colonoscopy. Our predictors included demographic variables (gender, age, marital status, education level, and occupation), personal health background (history of chronic diarrhoea, chronic constipation, mucus or blood in stool, history of chronic appendicitis or appendectomy, history of chronic cholecystitis or cholecystectomy, cancer history, history of colon polyps), life experiences (whether there have been significant traumatic events causing mental distress or anguish in the past ten years), first-degree relatives' history of colon cancer (father, mother, siblings, children), and personal lifestyle factors (smoking status). We utilise the SHapley Additive exPlanations (SHAP) analysis to elucidate the importance of variables. Our primary outcome was colonoscopy behaviour. Evaluation metrics, such as ROC curves and AUC values, were used to evaluate model performance. RESULTS: The diagnostic colonoscopy completion rate was 11.5 % (5664/49,418). Adaboost, XGBoost, RF, GBM, MLP and LR reported moderate predictive performance on the testing data (AUC > 60.0 %). Important predictors for adherence to diagnostic colonoscopy included FOBT screening, history of cancer, gender, age, history of colorectal polyps and occupation. To show the potential applications of our predictive model, we created ColoAdhere, a web tool tailored for institutional purposes that could predict adherence to diagnostic colonoscopy. CONCLUSIONS: A low diagnostic colonoscopy completion rate among individuals at high risk for colorectal cancer suggests targeted interventions are needed. Our study proposed a new and promising approach of using an AI behaviour prediction tool to increase the low completion rate of diagnostic colonoscopy among middle-aged and elderly adults in colorectal cancer screening programs.