Applying Machine Learning on Customized Invitation to Improve Regular Repeated Fecal Immunochemical Test and Colonoscopy Referral Rates Based on Hospital-Based Colorectal Cancer Screening

Background: The nationwide colorectal cancer (CRC) screening using biennial fecal immunochemical test (FIT) has been launched since 2004. Both proportion of late stage CRC and CRC mortality were significant declined. To enhance the screening coverage rate, the model was extended from community-based to hospital-based model and the amount of CRC screening participants was dramatically increasing in hospital. It is challenge to enhance regular repeated screening rate and the colonoscopy referral rate for positive cases to sustain this screening achievement.Aims: First, this project was aimed to apply machine learning for regular repeated participants and colonoscopy referral based on hospital-based to build the predictive model/classification or tendency of screening participation by different hospital levels. Second, based on the first finding on the pattern or tendency classification as stratification, we conduct the randomized controlled trial to evaluate the effectiveness of text message reminders and on invitation repeated screening and colonoscopy referral completion for FIT positive cases.Materials and methods: Based on the retrospective cohort study design, both eligible population for regular repeated screening between 2011-2018 and the colonoscopy referral lists for FIT positive subjects between 2010-2019 were employed build up and validate the predictive models for pattern classification and tendency clustering using logistic regression analysis and machine learning approaches, including Artificial Neural Network, gradient boosting machine, SVM (Support Vector Machine), random forest methods, etc. The empirical data are randomly divided into training, validation, and testing datasets for model’s build-up, tuning, validation, and testing processes. The 5-fold cross validation method is conducted to model validation for general statistical logistic regression. The indicators of area under ROC and APR (Area under the precision-recall curve) are conducted for models’ performance comparison. The randomized controlled trial was used to evaluate the effectiveness of the mobile text message reminders on repeated screening participants and colonoscopy referral completion.Expected results: Applying machine learning approach on big data combining screening database and hospital-based databases, the predictive models for subjects’ pattern classification and tendency clustering would be generated for application, especially for different areas and hospital-level. Using the pattern classification as stratification, the customized mobile text message on invitation for repeated screening and reminding for colonoscopy referral completion can be recommended based on scientific evidenced evaluation. These results and experience can be applied for other disease screening and other hospitals use.

Project ID:PC10907-0936
External Project ID:MOST109-2314-B182-038-MY3