Artificial intelligence for the prediction of prognosis in colorectal cancer patients using routine blood indices

Overall survival (OS) of colorectal cancer (CRC) patients remains suboptimal, especially in advanced disease. This study aimed to construct and validate explainable machine learning (ML) models using routine blood indices for accurate CRC prognosis across multicenter cohorts. The training cohort included 850 CRC patients (demographic and routine blood data) from Union; validation cohorts were 403 patients (Hefei) and 217 (Shihezi). Seven time-to-event models and SHapley Additive exPlanation (SHAP) (for interpretation) were used. Among the evaluated models, the random survival forest (RSF) algorithm demonstrated superior predictive performance. RSF algorithm demonstrated high discriminatory performance in the Union test cohort with AUCs of 0.768, 0.775, and 0.731 for 1-year, 2-year and 3-year OS, which was sustained in external validation cohorts: Hefei (0.820, 0.805, 0.775) and Shihezi (0.651, 0.706, 0.747). SHAP analysis identified CEA, CA125, age, MPV, CA19-9, INR and monocyte that contributed to the accurate prediction of RSF model. This study provides an innovative strategy for the convenient and accurate prediction of survival outcome of CRC individuals based on routine blood laboratory indices. RSF model helps oncologists to early identify CRC patients with high risk of death and provides a basis for personalized treatment.