Preoperative gut microbiota combined with machine learning for predicting post-hepatectomy liver failure in HBV-related hepatocellular carcinoma: a pilot study.
Peng Y, Xu J, et al. • Frontiers in cellular and infection microbiology • 2026
Preoperative gut microbiota profiles combined with machine learning models demonstrated cross-geographic reproducibility in identifying distinct microbial signatures associated with post-hepatectomy liver failure in HBV-related hepatocellular carcinoma, achieving AUCs up to 78.12% with high specificity and negative predictive values of 81.25% for low-risk patient identification.
Key Findings
Results
Significant beta-diversity differences were observed between PHLF and non-PHLF patients based on Weighted UniFrac distances.
Beta-diversity analysis using Weighted UniFrac metric revealed distinct GM community composition between PHLF and non-PHLF groups.
A total of 233 fecal samples underwent 16S rRNA sequencing and were stratified into a training set (n=179), internal validation set (n=32), and independent external validation set (n=22).
The cross-geographic reproducibility of these differences was demonstrated across the independent external validation set.
This finding supports that gut microbial community structure, not just individual taxa, differs systematically between risk groups.
Results
Eleven differential genera were identified between PHLF and non-PHLF patients, with PHLF patients showing enrichment of Faecalibacterium and Blautia alongside depletion of Bacteroides.
A total of 11 genera showed differential abundance between PHLF and non-PHLF groups.
PHLF patients exhibited 'a characteristic enrichment of Faecalibacterium, and Blautia, alongside a profound depletion of Bacteroides.'
These taxa were identified from 16S rRNA sequencing of preoperative fecal samples.
These genera served as candidate biomarkers for constructing machine learning-driven risk models.
Results
Functional inference revealed distinct metabolic remodeling in the PHLF group, including upregulation of amino acid biosynthesis and downregulation of key metabolic pathways.
The PHLF group showed upregulation of amino acid biosynthesis, specifically lysine and arginine, and nucleoside catabolism.
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' was observed in PHLF patients.
These functional inferences were derived from the gut microbiota compositional data using predictive functional profiling.
The downregulation of butyrate metabolism is notable given SCFAs' known role in gut barrier integrity and hepatic health.
Results
The XGBoost model achieved the highest AUC of 78.12% in the independent external validation set among the three machine learning models tested.
In the independent external validation set (n=22), AUCs were: XGBoost 78.12%, Random Forest 71.88%, and Gradient Boosting Machine 63.54%.
All three models were trained on preoperative gut microbiota features from the training set (n=179) and internally validated (n=32) before external validation.
XGBoost and RF achieved specificities of 81.25% versus 72.22% for GBM.
All three models maintained a stable negative predictive value (NPV) of 81.25% and 'exceptionally stable negative F1-scores (F1- negative)'.
Results
Decision curve analysis confirmed notable clinical net benefit for all three machine learning models across a range of threshold probabilities.
Decision curve analysis (DCA) was applied to evaluate the clinical utility of XGBoost, Random Forest, and Gradient Boosting Machine models.
DCA 'confirmed notable clinical net benefit of all models,' supporting their clinical applicability beyond statistical performance metrics.
The models demonstrated 'potent exclusionary performance,' suggesting particular utility for ruling out PHLF risk in low-risk patients.
This finding supports the use of these models for perioperative risk stratification and patient triage.
Discussion
The study is described as among the first to characterize distinct gut microbiota signatures in PHLF patients with cross-geographic reproducibility.
The authors state they 'are among the first to characterize distinct GM signatures in PHLF patients with cross-geographic reproducibility.'
Cross-geographic validation was achieved through the use of an independent external validation set (n=22) separate from the training (n=179) and internal validation (n=32) sets.
The study population consisted of patients with newly diagnosed HBV-related hepatocellular carcinoma (HBV-HCC).
This is described as a pilot study, acknowledging the preliminary nature of the findings.
Results
The gut microbiota-based models demonstrated strong specificity and negative predictive value, making them particularly suited for identifying low-risk patients.
Specificities were 81.25% for both XGBoost and RF, and 72.22% for GBM in the external validation set.
All three models achieved a negative predictive value (NPV) of 81.25%.
The authors characterize this profile as 'a clinically robust framework for accurate low-risk patient identification and triage.'
High NPV and specificity suggest the models are better suited for ruling out PHLF than for ruling it in.
What This Means
This research suggests that the bacterial community living in the gut before surgery can provide meaningful information about which liver cancer patients are at risk of developing liver failure after having part of their liver removed (a procedure called hepatectomy). The researchers collected stool samples from 233 patients with hepatitis B virus-related liver cancer before their surgery and analyzed the types and amounts of bacteria present. They found that patients who later developed post-surgery liver failure had distinctly different gut bacterial profiles — particularly higher levels of Faecalibacterium and Blautia bacteria and lower levels of Bacteroides — compared to patients who recovered without this complication. Additionally, the metabolic functions encoded by these bacteria differed, with liver failure patients showing disrupted pathways related to energy metabolism and reduced production of beneficial short-chain fatty acids like butyrate.
Using these gut bacteria profiles as inputs, the researchers trained three different machine learning (AI) models to predict which patients would develop liver failure after surgery. When tested on an independent group of patients from a different geographic location, the best-performing model (XGBoost) correctly identified non-PHLF cases with a specificity of 81.25% and a negative predictive value of 81.25%, meaning the model was particularly good at correctly ruling out patients who were not at high risk. The AUC (a measure of overall model accuracy) reached 78.12% for the XGBoost model in this external validation.
This research suggests that analyzing gut bacteria through a simple stool sample before liver surgery could offer a non-invasive way to help clinicians better identify patients who are unlikely to develop post-surgical liver failure, potentially improving how surgical candidates are selected and managed. Because the findings held up in patients from a different geographic region, the gut microbiota signatures appear to be broadly relevant rather than population-specific. The authors note this is a pilot study, and larger studies will be needed to confirm these findings before clinical implementation.
Peng Y, Xu J, Lu H, Bi Y, Qi L, Chen Y, et al.. (2026). Preoperative gut microbiota combined with machine learning for predicting post-hepatectomy liver failure in HBV-related hepatocellular carcinoma: a pilot study.. Frontiers in cellular and infection microbiology. https://doi.org/10.3389/fcimb.2026.1877450