Extreme gradient boosting-based explainable machine learning model for predicting significant fibrosis in autoimmune hepatitis

BACKGROUND: Accurate assessment of liver fibrosis is crucial for patients with autoimmune hepatitis (AIH). AIM: We developed and validated a non-invasive explainable machine learning (ML) model for the prediction of liver fibrosis in patients with AIH. DESIGN: A retrospective multicenter study of patients with AIH with liver biopsy was conducted. METHODS: Patients were randomly divided into a training set and a test set. Nine ML models were built, including logistic regression, k-nearest neighbors, Support vector machine, random forest, extreme gradient boosting (XGBoost), gradient boosting, Adaboost, decision tree, and Gaussian naive bayes. The best model was compared with aminotransferase to platelet ratio index (APRI) and fibrosis index based on four factors (FIB-4) on the test set by area under receiver operating characteristic curves (AUC). Shapley additive explanation (SHAP) analysis and local interpretable model-agnostic explanations (LIME) were used for model explanation. RESULTS: A total of 261 patients with AIH with a median age of 54.0 (interquartile range: 47.0-62.0) years and 82.8% of female sex were included. Among nine ML models, the XGBoost model exhibited superior predictive performance. The model achieved an AUC of 0.791 (95% confidence interval [CI]: 0.668-0.890) in the test set which was higher than APRI (AUC: 0.557, 95% CI: 0.380-0.732, P < 0.001) and FIB-4 (AUC: 0.625, 95% CI: 0.452-0.789, P < 0.001). SHAP and LIME analysis revealed that platelet was the most important predictive variable of significant liver fibrosis. CONCLUSIONS: The non-invasive interpretable XGBoost model surpasses APRI and FIB-4 for predicting significant liver fibrosis, contributing to better management of patients with AIH.