Development and validation of a machine learning model for colorectal cancer status classification using NHANES data:...

Colorectal cancer (CRC) is a leading cause of cancer-related morbidity and mortality worldwide. Tools based on routinely collected variables may help identify prevalent CRC status and support clinical evaluation. Traditional approaches often rely on limited predictors and may not capture the multidimensional nature of CRC. Data were obtained from the National Health and Nutrition Examination Survey 1999-2018. Among 53,881 participants, 420 reported physician-diagnosed CRC (MCQ220). A 1:10 stratified case-control sample was constructed (420 cases, 4200 controls) and randomly split into training (70%) and internal validation (30%) sets. Missing data were imputed separately in the training and validation sets using a random forest-based method. SMOTE was applied only to the training set. Logistic regression, random forest, support vector machine, k-nearest neighbors, and extreme gradient boosting (XGBoost) were compared. Performance was primarily assessed by discrimination in the held-out validation set. For the final XGBoost model, exploratory post hoc probability calibration analyses were performed on validation-set predictions using raw probabilities, prior prevalence correction, and Platt scaling. Model interpretability was examined using Shapley Additive Explanations (SHAP), and a web-based CRC status classifier was developed. XGBoost showed the best discrimination in the validation cohort, with an area under the receiver operating characteristic curve of 0.787 (95% confidence interval 0.749-0.825). At the Youden-index cutoff, sensitivity was 77.0%, specificity 67.6%, PPV 19.2%, and NPV 96.7%. In exploratory probability-based analyses, decision curve analysis using Platt-scaled probabilities showed greater net benefit than treat-all and treat-none strategies across low-to-moderate threshold probabilities. post hoc calibration analyses fitted and assessed on the validation-set predictions showed improved apparent agreement after Platt scaling, with a Brier score of 0.191 and a Hosmer-Lemeshow P value of 0.086. SHAP identified key predictors, including alcohol use, hypertension, age, triglycerides, absolute lymphocyte count, blood lead, serum cotinine, and neutrophil-to-lymphocyte ratio. An interpretable machine learning framework integrating multidomain predictors enabled effective CRC status classification in a large population-based cohort. Discrimination was strong, whereas probability-based outputs after post hoc calibration should be considered exploratory pending independent confirmation. The model may support clinical evaluation and triage for individuals requiring further assessment.