Novel HEP-FIB score combined with machine learning models for non-invasive prediction of significant hepatic fibrosis in chronic hepatitis B patients: a single-center, internally validated study

Chronic hepatitis B (CHB) remains a major global health burden. Assessment of hepatic fibrosis is important for risk stratification and guiding treatment in CHB. Liver biopsy, the reference standard for fibrosis staging, is invasive, and fraught with potential complications, creating the need for alternative non-invasive tools. Existing non-invasive fibrosis scores demonstrate variable diagnostic accuracy, and the role of lipids in fibrosis prediction remains insufficiently explored. We developed and internally validated HEP-FIB, a parsimonious score based on routinely available laboratory parameters, and evaluated machine learning (ML) models for non-invasive prediction of significant hepatic fibrosis in CHB patients. A retrospective cohort of 382 CHB patients was analyzed. Least Absolute Shrinkage and Selection Operator (LASSO) regression followed by multivariable logistic regression was used to identify independent predictors of fibrosis. A novel HEP-FIB score was developed based on platelet count, liver function parameters, and lipid biomarkers. Diagnostic performance was evaluated using receiver operating characteristic (ROC) curve analysis, bootstrap internal validation and calibration analysis. The performance of HEP-FIB was compared with established non-invasive indices, including APRI, FIB-4, and GPR. In addition, ten ML algorithms were constructed and evaluated for predictive performance. ML model performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC), average precision (AP), accuracy, sensitivity, specificity, positive predictive value (PPV, precision), negative predictive value (NPV), F1 score, Matthews Correlation Coefficient (MCC), balanced accuracy, false positive rate (FPR), false negative rate (FNR), and Brier score to allow a comprehensive evaluation of discrimination, classification performance, calibration, and overall predictive ability. Decision curve analysis (DCA) was performed separately for both HEP-FIB score and ML models to assess clinical utility. The HEP-FIB score demonstrated excellent diagnostic performance, achieving an AUC of 0.892 (95% CI: 0.859–0.925).At the optimal cutoff value (≥ 4), the score showed a sensitivity of 78%, specificity of 89%, and overall accuracy of 88%. HEP-FIB outperformed APRI (AUC 0.724), FIB-4 (AUC 0.654), and GPR (AUC 0.773) (all p < 0.001 by DeLong test). Bootstrap validation showed stable performance with an optimism-corrected AUC of 0.881. Calibration analysis showed good agreement between predicted and observed fibrosis probabilities, while DCA demonstrated superior net clinical benefit over default strategies. Among ML approaches, XGBoost achieved the best predictive performance, with an AUC of 0.914 (95% CI: 0.857–0.971), accuracy of 0.88, sensitivity of 0.72, specificity of 0.93, balanced accuracy of 0.83, and MCC of 0.74 in the validation dataset. The model demonstrated favorable calibration, evidenced by the low Brier score (0.095), together with reduced misclassification errors as reflected by low FPR (0.062) and FNR (0.275). DCA demonstrated a greater net clinical benefit for the ML models compared with both the test-all and test-none strategies. The ML models showed a marginal performance advantage over HEP-FIB score. HEP-FIB demonstrated strong diagnostic performance for identifying significant hepatic fibrosis in CHB and outperformed conventional non-invasive fibrosis scores in this retrospective, single-center cohort after internal bootstrap validation. Ensemble ML models achieved marginally superior predictive accuracy over HEP-FIB score. However, external validation in independent and diverse CHB populations is required before the score can be recommended for routine clinical use.