Machine learning-based preoperative classification of colorectal cancer stage using systemic inflammatory and nutritional biomarkers.

BACKGROUND

Accurate preoperative staging in colorectal cancer (CRC) is critical for determining the appropriate treatment strategy. This study evaluated whether early-stage (TNM I-II) and advanced-stage (TNM III-IV) CRC could be distinguished using machine learning (ML) models in conjunction with biomarkers reflecting systemic inflammatory and nutritional status.

METHODS

We retrospectively analyzed data of 204 patients with CRC. Preoperative variables included tumor size, tumor markers (carcinoembryonic antigen [CEA] and CA 19-9), and inflammation and nutrition-based biomarkers derived from routine laboratory tests. Five supervised machine learning models (random forest, gradient boosting, logistic regression, neural network, and naive Bayes) were developed and evaluated using stratified cross-validation. Model performance was evaluated using the area under the curve (AUC), accuracy, F1 score, and Matthews correlation coefficient (MCC). Model interpretability was examined using SHAP analysis.

RESULTS

We evaluated 204 patients (103 early-stage and 101 advanced-stage). Advanced-stage disease was significantly associated with higher inflammatory indices and lower nutritional biomarkers (all p < 0.01). Ensemble-based models showed the strongest performance, whereas the random forest algorithm achieved the highest discriminatory performance (area under the curve [AUC]: 0.91; accuracy: 0.85). SHAP analysis revealed HALP, PLR, and LMR as the most effective predictors and suggested that systemic inflammatory and nutritional status may contribute more to stage differentiation than tumor size.

CONCLUSION

Machine learning models incorporating routinely available inflammatory and nutritional biomarkers may help differentiate early-stage from advanced-stage colorectal cancer in the preoperative setting. However, larger multicenter studies with external validation are needed to confirm the generalizability of these findings.