Development and internal validation of an immune-based prognostic modeling of early-onset colorectal cancer via machine learning

BACKGROUND Early-onset colorectal cancer (EOCRC) is an aggressive malignancy with rising incidence and poor prognosis in young adults. Circulating immune cells may hold prognostic value, yet their role in EOCRC outcomes remains unclear. AIM To develop machine learning-based prognostic models using peripheral immune markers in a retrospective cohort of EOCRC patients. METHODS A cohort of 123 EOCRC patients undergoing radical resection, from January 2017 to December 2020 was included. Data were extracted from medical records with a follow-up till July 2025. Blood samples were processed for flow cytometry to assess immune markers. RESULTS Univariable screening identified disease stage and CD16+CD56+ natural killer (NK) cell percentage as top predictors. A parsimonious Cox model integrating stage and high NK cells outperformed random survival forests (concordance index 0.693 vs 0.256). High-risk patients (stage III/IV, high NK cells) had inferior 5-year progression-free survival (61.2%; 95% confidence interval: 49.0-76.5) vs low-risk (86.4%; 95% confidence interval: 78.9-94.6; log-rank P = 0.001). Time-dependent areas under the curve ranged from 0.671 to 0.693, with robust calibration. CONCLUSION This two-factor model offers moderate accuracy for personalized EOCRC risk stratification, highlighting systemic NK cell dysfunction as a potential immunotherapy target. External validation is warranted.