Interpretable machine learning models based on body composition and inflammatory nutritional index (BCINI) to predict early postoperative recurrence of colorectal cancer: Multi-center study.

BACKGROUND AND OBJECTIVE

Colorectal cancer (CRC) ranks among the most prevalent cancers worldwide, with early postoperative recurrence remaining a major cause of mortality. Body composition and inflammatory-nutritional indices (BCINI) have demonstrated potential in reflecting patients' physiological states; however, their association with early recurrence (ER) after CRC resection remains unclear. This study aimed to establish and validate interpretable machine learning (ML) models based on BCINI to predict ER after CRC resection.

METHODS

Data from three hospitals were collected, including CT-based body composition metrics and blood test variables. After variable selection, six ML algorithms-XGBoost, Complement Naive Bayes (CNB), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), and Gaussian Naive Bayes (GNB)-were used to construct ER prediction models. Optimal model selection was based on receiver operating characteristic (ROC) curve analysis. The selected model was externally validated using independent datasets to assess generalizability, while its accuracy and clinical utility were evaluated via calibration curves and decision curve analysis. Additionally, SHapley Additive exPlanations were employed to visualize prediction processes for clinical interpretability.

RESULTS

The XGBoost algorithm outperformed other methods in model selection, demonstrating superior accuracy and stability with area under the ROC curve (AUC) values of 0.837 and 0.777 in internal training and validation sets, respectively. This model achieved the lowest Brier score of 0.131 on calibration curves, surpassing the five other ML algorithms. External validation further confirmed its generalizability, yielding AUC values of 0.783 and 0.773 in two independent datasets. Consistent predictive performance was observed across age subgroups (<60 years: AUC 0.762-0.834; ≥60 years: AUC 0.777-0.800) and tumor location subgroups (colon: AUC 0.785-0.845; rectum: AUC 0.751-0.799).

CONCLUSIONS

The interpretable ML model developed based on BCINI shows promise in predicting ER of CRC. This approach may provide valuable insights for clinical decision-making, enabling early detection and intervention to improve patient outcomes.