AUTHOR=Yang Xiaoyu , Xu Jinjian , Ji Hong , Li Jun , Yang Bingqing , Wang Liye 

TITLE=Early prediction of colorectal adenoma risk: leveraging large-language model for clinical electronic medical record data

JOURNAL=Frontiers in Oncology

VOLUME=Volume 15 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2025.1508455

DOI=10.3389/fonc.2025.1508455

ISSN=2234-943X

ABSTRACT=ObjectiveTo develop a non-invasive, radiation-free model for early colorectal adenoma prediction using clinical electronic medical record (EMR) data, addressing limitations in current diagnostic approaches for large-scale screening.DesignRetrospective analysis utilized 92,681 cases with EMR, spanning from 2012 to 2022, as the training cohort. Testing was performed on an independent test cohort of 19,265 cases from 2023. Several classical machine learning algorithms were applied in combination with the BGE-M3 large-language model (LLM) for enhanced semantic feature extraction. Area under the receiver operating characteristic curve (AUC) is the major metric for evaluating model performance. The Shapley additive explanations (SHAP) method was employed to identify the most influential risk factors.ResultsXGBoost algorithm, integrated with BGE-M3, demonstrated superior performance (AUC = 0.9847) in the validation cohort. Notably, when applied to the independent test cohort, XGBoost maintained its strong predictive ability with an AUC of 0.9839 and an average advance prediction time of 6.88 hours, underscoring the effectiveness of the BGE-M3 model. The SHAP analysis further identified 16 high-impact risk factors, highlighting the interplay of genetic, lifestyle, and environmental influences on colorectal adenoma risk.ConclusionThis study developed a robust machine learning-based model for colorectal adenoma risk prediction, leveraging clinical EMR and LLM. The proposed model demonstrates high predictive accuracy and has the potential to enhance early detection, making it well-suited for large-scale screening programs. By facilitating early identification of individuals at risk, this approach may contribute to reducing the incidence and mortality associated with colorectal cancer.