Yong-ling Yang, Yi-Sheng Huang, Yu-long Bai, Xiao-xiang Huang, Zhao-Yin Fu, Zhi-wei Huang
Background Sepsis is a life-threatening condition with high mortality, necessitating early risk stratification. This study aimed to develop and validate a predictive nomogram for 28-day mortality in sepsis patients incorporating machine learning-selected biomarkers. Methods A total of 1,350 sepsis patients were retrospectively enrolled and divided into training ( n = 944) and testing ( n = 406) sets. LASSO and Random Forest (RF) algorithms were applied to screen key biomarkers associated with mortality. A logistic regression model was constructed using the selected features, and a nomogram was developed by integrating these biomarkers with APACHE II, SOFA score, and shock status. Model performance was evaluated by AUC, calibration, and decision curve analysis (DCA). External validation was performed in an independent cohort of 120 patients. Results Six biomarkers consistently selected by both LASSO and RF were: procalcitonin (PCT), prognostic nutritional index (PNI), red blood cell count (RBC), platelet count (PLT), alanine aminotransferase (ALT), and indirect bilirubin (IBIL). Non-survivors exhibited significantly higher levels of PCT, ALT, and IBIL, and lower levels of RBC, PLT, and PNI compared to survivors (all P < 0.05). The logistic regression model demonstrated strong discrimination [AUC: 0.841 (95% CI: 0.814–0.868) in training set; 0.808 (95% CI: 0.769–0.847) in testing set]. The nomogram showed good calibration and favorable net clinical benefit across a wide range of threshold probabilities. In the external validation cohort, the model maintained excellent predictive performance with an AUC of 0.921 (95% CI: 0.876–0.966). Conclusion We developed and validated a clinically useful nomogram incorporating nutrition-related biomarkers, particularly PNI, for predicting 28-day mortality in sepsis patients. The model demonstrates robust performance and highlights the importance of nutritional status in sepsis outcomes.