Construction of prognostic models for colorectal cancer based on the perioperative modified Gustave Roussy Immune Score and machine learning.

OBJECTIVE

To evaluate the prognostic value of preoperative and postoperative modified Gustave Roussy Immune Score (mGRIm) and dynamic mGRIm trajectories in patients with colorectal cancer (CRC), and to develop internally validated machine-learning models based on perioperative dynamic features.

METHODS

This retrospective cohort study included 457 patients with CRC who underwent primary tumor resection at Hubei Cancer Hospital between January 2016 and December 2019. Preoperative and postoperative mGRIm scores were calculated using receiver operating characteristic curve-derived cut-off values for albumin, lactate dehydrogenase, and neutrophil-to-lymphocyte ratio. Dynamic mGRIm trajectory was defined as the transition between preoperative and postoperative mGRIm risk states. Associations with overall survival (OS) and disease-free survival (DFS) were evaluated using Kaplan-Meier analysis and Cox proportional hazards models. Machine-learning model development incorporated Boruta-based feature selection, training set-restricted 5-fold cross-validation, and training set-restricted resampling strategies.

RESULTS

Dynamic mGRIm trajectory showed strong and consistent prognostic stratification. Patients with persistently high mGRIm status had significantly higher risks of death (HR = 3.47, 95% CI: 1.76 -6.83, p < 0.001) and DFS events (HR = 2.76, 95% CI: 1.65 -4.61, p < 0.001) than those with persistently low status. In the prespecified evaluation framework integrating multiple performance metrics, the neural network model achieved the highest composite score for OS prediction (AUC = 0.801), whereas the radial basis function support vector machine achieved the best performance for DFS prediction (AUC = 0.856) in the independent test set.

CONCLUSION

Dynamic mGRIm trajectory provides robust prognostic information for patients with CRC. Machine-learning models based on perioperative dynamic features demonstrate satisfactory internal predictive performance; however, external validation is required prior to clinical application.