Related Experiment Videos
Serum Inflammatory-Nutritional Markers Refine Machine-Learning Prognostic Stratification for Resectable
Ao Sun1, Changlei Li2, Junzhe Zhang1
1Department of Hepatobiliary and Pancreatic Surgery, The Affiliated Hospital of Qingdao University, Qingdao, People's Republic of China.
Background:
Outcomes after curative resection for hepatocellular carcinoma (HCC) remain heterogeneous, and conventional staging systems have limited resolution for individual postoperative risk prediction. Serum inflammatory-nutritional markers are inexpensive and routinely available, but their value as integrated inputs for machine-learning survival modeling after HCC resection remains insufficiently defined.
Methods:
This retrospective cohort study included 880 patients with pathologically confirmed HCC who underwent curative resection at the Affiliated Hospital of Qingdao University between 2013 and 2024. Overall survival (OS) was the primary endpoint, and recurrence-free survival (RFS) was secondary. Twelve preoperative serum inflammatory-nutritional indexes were assessed after adjustment for key clinical covariates. Retained markers were combined into a serum inflammatory-nutritional risk score (SINRS). LASSO regression and Boruta algorithm were then used to identify clinicopathological predictors. The SINRS and selected variables were entered into 10 machine learning algorithms and their 101 combinations. Model performance was evaluated using C-index, time-dependent AUC, calibration, decision curve analysis, Kaplan-Meier curve and SHapley Additive exPlanations (SHAP).
Results:
Neutrophil percentage-to-albumin ratio (NPAR), prognostic nutritional index (PNI), albumin-to-alkaline phosphatase ratio (AAPR), fibrinogen-to-albumin ratio (FAR) and hemoglobin-albumin-lymphocyte-platelet (HALP) score, were retained for SINRS construction. Nine clinicopathological predictors were selected: age, alpha-fetoprotein, alkaline phosphatase, eosinophil count, lymphocyte count, fibrinogen, tumor size, microvascular invasion and satellite nodules. Among all candidate strategies, the random survival forest (RSF) model showed the best predictive performance. For OS prediction, 2-, 3- and 5-year AUCs were 0.814, 0.792 and 0.816 in the training cohort and 0.796, 0.787 and 0.790 in the validation cohort. SHAP analysis ranked microvascular invasion, SINRS and tumor size among the features with the highest model-derived importance for OS prediction.
Conclusion:
A serum inflammatory-nutritional marker-informed RSF model refined OS stratification after curative HCC resection in this internally validated cohort, which may provide additional information for postoperative risk stratification but requires external validation before clinical implementation.