Research on prognostic prediction of colorectal cancer based on multi-dimensional biomarker features and machine learning models.

Background: The traditional TNM staging system for colorectal cancer (CRC) prognosis has significant limitations, with prognostic differences exceeding 40% among patients with the same stage. This study aims to construct and validate a machine learning prognostic prediction model based on routinely available clinical, laboratory, inflammatory, nutritional, and treatment-related variables for hospitalized colorectal cancer patients. Given that the training cohort (MIMIC-IV) primarily comprises hospitalized and critically ill patients, the proposed model is intended for risk stratification in hospitalized CRC patients. Methods: A retrospective cohort study used the MIMIC-IV database (n = 2049) for training/internal validation and the Second Affiliated Hospital of Soochow University (n = 561) for external validation. The study integrated characteristics including clinical pathology, inflammatory response, immune function, nutritional status, tumor markers, and treatment information. Through correlation analysis, collinearity testing, and the Boruta algorithm, 16 core features were selected. Nine machine learning algorithms were employed to construct the model, with hyperparameters optimized via five-fold cross-validation and grid search. Model performance was evaluated using AUROC, AUPR, Brier score, survival consistency index, and decision curve analysis, with feature contribution explained by SHAP values. Results: The RSF model achieved the highest normalized comprehensive score among the nine algorithms evaluated. In the external validation set, the RSF model achieved an average AUROC of 0.818 (95% CI: 0.761-0.866), ranking first in this composite metric. The full RSF model showed statistically significantly higher AUROC than TNM-only baseline models (DeLong test P < 0.001), suggesting potential for incremental prognostic information beyond traditional staging, pending prospective validation. Feature importance analysis revealed that age-corrected Charlson Comorbidity Index (ACCI), distant metastasis stage (Mstage), and surgical treatment were core prognostic factors, exhibiting time-dependent dynamic changes. Inflammatory markers (CRP), nutritional markers (ALB, TC), and tumor markers (CEA) also demonstrated significant predictive value. Conclusions: The RSF model demonstrated favorable discrimination and calibration in the evaluated cohorts; clinical utility requires prospective evaluation in the available cohorts, showing incremental prognostic value beyond TNM staging alone. These findings suggest potential for assisting risk stratification in similar hospitalized populations, although further prospective and multicenter validation remains necessary before clinical implementation.