Prognostic stratification in metastatic early-onset colorectal cancer: development and independent external validation of a survival model

Metastatic early-onset colorectal cancer (mEO-CRC) is a clinically distinct subgroup with substantial survival heterogeneity, yet prognostic tools specifically validated for this population remain limited. We aimed to develop and independently externally validate a survival model for prognostic stratification in patients with mEO-CRC and to explore molecular features associated with survival heterogeneity. A total of 7061 patients aged 18–49 years with mEO-CRC from the SEER database (2010–2020) were randomly assigned to a training cohort ( n = 4942) and an internal validation cohort ( n = 2119). An independent cohort from Nanfang Hospital ( n = 182) was used for external validation. A Cox regression model incorporating 16 clinicopathological and treatment-related predictors was developed and presented as a nomogram. Model performance was assessed using the C-index, time-dependent AUC, quantitative calibration metrics, Brier score, and decision curve analysis (DCA). Missingness, LODDS categorization, proportional hazards assumptions, and multiple imputation by chained equations (MICE) sensitivity analyses were further evaluated. Exploratory bioinformatic analyses based on GEO and TCGA datasets were performed as hypothesis-generating analyses. Sixteen predictors were incorporated into the model. The C-index was 0.720 (95% CI 0.710–0.729) in the training cohort, 0.728 (95% CI 0.713–0.743) in the internal validation cohort, and 0.685 (95% CI 0.632–0.737) in the independent external validation cohort. The 1-, 3-, and 5-year AUCs were 0.775, 0.798, and 0.836 in the training cohort; 0.776, 0.808, and 0.845 in the internal validation cohort; and 0.713, 0.778, and 0.747 in the external validation cohort. Calibration and decision-curve analyses supported moderate prognostic utility, and multiple-imputation sensitivity analyses yielded broadly consistent discrimination. Exploratory molecular analyses identified 81 differentially expressed genes; PRL and ATP2A1 showed survival associations in TCGA/GEPIA analyses and were considered hypothesis-generating findings. We developed and independently externally validated a survival model for prognostic stratification in mEO-CRC. The model showed moderate external discrimination and may support individualized prognostic assessment, but its use should be considered in light of missing registry data, treatment-selection bias, proportional-hazards limitations, and the need for prospective multicenter validation. The exploratory molecular findings provide biological context for future investigation rather than direct mechanistic validation.