Artificial intelligence-based models for colorectal cancer diagnosis using laboratory tests: an exploratory retrospective case-control study.

BACKGROUND: Colorectal cancer (CRC) is frequently diagnosed at advanced stages owing to the insidious nature of its early clinical manifestations and the lack of validated screening modalities. This diagnostic delay presents a substantial therapeutic challenge, underscoring the need to develop population-based screening strategies. Based on laboratory data, this study constructed a diagnostic model for CRC to explore a screening strategy for population-based tumor census.

METHODS: In this retrospective case-control study, we analyzed anonymized laboratory parameters and baseline demographic and clinical characteristics of patients with histologically confirmed CRC and age-matched healthy controls undergoing routine health checkups at our tertiary care center between January 2021 and December 2022. Non-parametric comparisons between groups were conducted using the Mann-Whitney U test. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic performance of candidate biomarkers. Following variable selection via least absolute shrinkage and selection operator (LASSO), the dataset was partitioned into training and internal validation sets via stratified randomization at a 7:3 ratio. Multiple machine learning algorithms were evaluated, and the optimal predictive model for this study was selected based on this comparative assessment. Subsequently, the influence of individual features on the model predictions was analyzed using SHapley Additive exPlanations (SHAP) values.

RESULTS: After rigorous screening, this study included 1,068 patients with histologically confirmed CRC and 1,068 age- and sex-matched healthy controls. Comparative analysis using the Mann-Whitney U test revealed statistically significant intergroup differences (P<0.05) in age and serum biomarkers. ROC curve analysis identified four predictors with superior diagnostic performances for CRC: total protein (TP), carcinoembryonic antigen (CEA), age, and apolipoprotein A1 (ApoA1). LASSO regression selected nine clinically relevant predictors, and various machine learning models were used to classify the data samples. Based on a comprehensive judgment, the Random Forest model [area under the curve (AUC) =0.946, Brier score =0.101] was the best model for the diagnosis of CRC in this study.

CONCLUSIONS: An effective risk stratification model for CRC screening was developed using routine clinical laboratory parameters and the readily available demographic variable of age. However, being derived from a single-center retrospective cohort, the model requires further validation in multi-center, prospective studies to confirm its generalizability and mitigate potential overfitting.