Comparison of three malnutrition assessment criteria for predicting postoperative complications and overall survival in colorectal cancer

Background Malnutrition is highly prevalent in patients with colorectal cancer and is associated with adverse postoperative outcomes and shortened survival. However, the comparative predictive performance of commonly used nutritional assessment tools—PG-SGA, ESPEN, and GLIM criteria—remains unclear. Methods A total of 1,258 patients who underwent radical resection for colorectal cancer were retrospectively enrolled. Malnutrition was diagnosed simultaneously using the PG-SGA, ESPEN, and GLIM criteria. Agreement among the three tools was evaluated using Cohen’s kappa coefficient. Multivariable logistic regression, an exploratory random forest analysis, and Cox proportional hazards models were applied. For the logistic regression models, ROC curve analysis and pairwise AUC comparisons using DeLong’s test were performed; for the Cox models, bootstrap-based C-index comparisons were conducted. Propensity score matching (PSM) was performed for the GLIM criterion as a post hoc sensitivity analysis to explore the association between GLIM-defined malnutrition and clinical outcomes. Results The prevalence of malnutrition defined by PG-SGA, ESPEN, and GLIM was 24.7%, 15.7%, and 29.9%, respectively, with pairwise agreements ranging from moderate to substantial. All three criteria were independent risk factors for postoperative complications, with GLIM having the highest odds ratio (OR = 1.869, p < 0.001). AUC comparison further showed that the GLIM-incorporated model achieved the numerically highest AUC (0.781), with a significantly higher AUC than the ESPEN-based model ( p = 0.023), whereas no significant difference was observed between GLIM and PG-SGA ( p = 0.172). For OS, PG-SGA (HR = 1.364, p = 0.014) and GLIM (HR = 1.415, p = 0.004) were independent prognostic factors, whereas ESPEN was not ( p = 0.129). Bootstrap C-index comparisons showed that the GLIM-based model provided modestly better prognostic discrimination than both the PG-SGA- and ESPEN-based models (both p < 0.05), although the absolute C-index improvement was small (approximately 0.015). After PSM balancing baseline characteristics, GLIM-defined malnutrition remained significantly associated with prolonged postoperative hospital stay, higher complication rates, and reduced 5-year overall survival. Conclusion The three malnutrition diagnostic criteria are not interchangeable. For postoperative complications, GLIM demonstrated better discriminative ability than ESPEN but was comparable to PG-SGA. For long-term OS, GLIM provided modestly better prognostic discrimination than both PG-SGA and ESPEN, although the absolute improvement was small. These findings support the potential clinical utility of GLIM for perioperative nutritional risk assessment; however, the clinical relevance of these relatively small performance differences warrants further validation.