Development and validation of a machine learning-derived dual-indicator tool for identification of hepatitis B virus-related pre-acute-on-chronic liver failure.

Background & aims: Patients who do not meet the criteria for Acute-on-Chronic Liver Failure (ACLF) at admission still face a high risk of disease progression and mortality. This study aimed to develop and validate a machine learning-derived clinical tool for the early identification of HBV-pre-ACLF. Methods: We analyzed 1,682 patients experiencing acute deterioration of HBV-related chronic liver disease but without ACLF. By random forest and SHAP analysis, we identified key predictors of ACLF onset within a 7-day window. We then validated the diagnostic thresholds in a test cohort of 260 patients. Results: Total bilirubin (TBIL) and INR were identified as the most critical predictors of ACLF progression. We established optimal diagnostic thresholds at TBIL ≥131.5 μmol/L and INR ≥1.35. Two sequential criteria were developed: pre-ACLF-O (meeting either threshold) for highly sensitive screening, and pre-ACLF-A (meeting both thresholds) for risk confirmation. In the derivation cohort, pre-ACLF-O achieved 98.3% sensitivity, while pre-ACLF-A predicted a 33.09% positive predictive value. These results were successfully validated in the test cohort, where pre-ACLF-O maintained 100% sensitivity and pre-ACLF-A demonstrated a 37.04% positive predictive value. Conclusions: This dual-indicator tool easily and effectively identifies HBV-pre-ACLF. Using these criteria sequentially provides a practical strategy for early risk stratification, allowing clinicians to initiate timely interventions for high-risk patients.