Preoperative gut microbiota combined with machine learning for predicting post-hepatectomy liver failure in HBV-related hepatocellular carcinoma: a pilot study.

OBJECTIVE: This study investigated the association between preoperative gut microbiota (GM) profiles and post-hepatectomy liver failure (PHLF) in patients with newly diagnosed HBV-related hepatocellular carcinoma (HBV-HCC), aiming to explore non-invasive, modifiable GM biomarkers for perioperative risk stratification. METHODS: A total of 233 fecal samples underwent 16S rRNA sequencing and were stratified into three distinct sets: a training set (n = 179), an internal validation set (n = 32), and an independent external validation set (n = 22). Preoperative GM features were compared between patients with and without PHLF to identify diagnostic biomarkers and construct machine learning-driven risk models (XGBoost, Random Forest [RF], and Gradient Boosting Machine [GBM]). RESULTS: Significant β-diversity differences (Weighted UniFrac) were observed between the groups, identifying 11 differential genera; PHLF patients exhibited a characteristic enrichment of Faecalibacterium, and Blautia, alongside a profound depletion of Bacteroides. Functional inferences revealed distinct metabolic remodeling in the PHLF group, characterized by the upregulation of amino acid biosynthesis (specifically lysine and arginine) and nucleoside catabolism, coupled with a prominent downregulation of the hepatic-interactive urea cycle, central carbon metabolism (glycolysis and tricarboxylic acid [TCA] cycle), and short-chain fatty acid [SCFA] (butyrate) metabolism. In the independent external validation set, the XGBoost, RF, and GBM models achieved areas under the receiver operating characteristic curve (AUCs) of 78.12%, 71.88%, and 63.54%, respectively. Importantly, the models showed potent exclusionary performance, with specificities of 81.25% for XGBoost and RF versus 72.22% for GBM, while all three maintained a stable negative predictive value (NPV) of 81.25% and exceptionally stable negative F1-scores (F1- negative).Decision curve analysis (DCA) confirmed notable clinical net benefit of all models. CONCLUSION: We are among the first to characterize distinct GM signatures in PHLF patients with cross-geographic reproducibility. GM demonstrates strong potential as a non-invasive tool for preoperative risk stratification targeting the primary endpoint of PHLF. Characterized by high specificity and potent exclusionary capacity, this signature establishes a clinically robust framework for accurate low-risk patient identification and triage, thereby optimizing perioperative management.