(197) Predicting CRC Screening Uptake using Machine Learning and Ensemble Models

Background: Colorectal cancer (CRC) remains a major health challenge, with nearly 76% of CRC fatalities linked to delayed screening adherence. Advances in digital health interventions, tailored to diverse populations and powered by machine learning, are proving crucial in improving screening rates and overcoming literacy barriers.

Objectives: Our objective is to use supervised learning to identify key factors in digital health literacy impacting CRC screening rates, aiming to develop targeted interventions for early detection and improved screening strategies.

Methods: We merged the HINTS, a biennial U.S. national digital literacy survey datasets, by adjusting variables, treating missing values, and enhancing its descriptive power. We then applied supervised learning models to a balanced dataset, split into 80% training and 20% testing sets for optimal performance analysis. Our study optimized AdaBoost and RandomForest classifiers using GridSearchCV, focusing on hyperparameters to balance model complexity and generalization, as indicated by high F1 scores and accuracy in CRC detection.

Results: For those who have not been screened, 33 individuals (representing 8.068% of one group) and 2467 individuals (27.145% of another group) are indicated. Conversely, for individuals who have undergone screening, 327 (79.95% of the former group) and 6142 (67.58% of the latter group) have been screened. Feature importance and chi-square tests identified key predictors, enhancing model efficiency and interpretability. The results demonstrated the superior performance of ensemble models, particularly RandomForest (RF) and AdaBoost (AB), with RF achieving 98% accuracy, 99% precision, 96% recall, an F1 score of 0.98, and an AUC of 0.98. These models significantly outperformed others like Logistic Regression and Perceptron, underscoring the effectiveness of ensemble approaches in CRC screening.

Conclusions: This study demonstrates the effectiveness of machine learning, particularly ensemble models like Random Forest and AdaBoost, in predicting CRC screening uptake by identifying crucial digital health literacy factors. This approach not only surpasses traditional models in accuracy but also offers a pathway for tailored, impactful screening strategies, suggesting a significant potential for reducing CRC incidence through targeted, data-driven interventions.