Abstract 1057: Addressing challenges in integrating dietary factors into machine learning models for gastrointestinal cancer prediction in South Korean cohort

Abstract Introduction: Dietary factors are integral to cancer etiology but are underrepresented in predictive models for gastrointestinal (GI) cancer risk. Challenges such as limited availability and variability of dietary intake data hinder the development of effective machine learning (ML) models. This study investigates these challenges and develops practical solutions to enhance the integration of dietary data into GI cancer risk prediction using a South Korean cohort. Methods: We analyzed data from the Cancer Screenee Cohort, following 7, 862 adults (158 GI cancer cases and 7, 704 controls) over 16 years. Dietary data were collected using a validated food frequency questionnaire (FFQ), supplemented by clinical biomarkers and anthropometric measurements. ML models—logistic regression, random forests, gradient boosting machines, and support vector machines—were trained and evaluated using precision and F1 scores, focusing on predicting the minority class (GI cancer cases). A tailored data sampling and model training strategy was developed to address class imbalance and improve predictive performance. Results: Initial ML models faced performance challenges due to class imbalance and limited sample size, achieving zero precision for the case class and strong performance only for control class (control precision = 0.98). The proposed sampling and model training strategy enhanced minority-class predictions achieving a case class prediction precision of 0.78—a 78% improvement—while maintaining an overall F1 score of 0.75, reflecting a strong balance between overall model precision and recall. Conclusions: This study highlights the challenges of incorporating dietary factors into GI cancer prediction and underscores the importance of tailored approaches to address class imbalance in ML models. The proposed strategy substantially improved predictions for the minority class, offering practical insights for enhancing GI cancer screening and prevention through data-driven methods. Future work could explore incorporating additional dietary factors and testing the approach in larger populations to further validate its applicability. Citation Format: Daina Baublyte, Jeonghee Lee, Madhawa Gunathilake, Jeongseon Kim. Addressing challenges in integrating dietary factors into machine learning models for gastrointestinal cancer prediction in South Korean cohort [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1057.