Machine Learning and SHAP Value Interpretation for Predicting Hepatic Steatosis Using Vibration-Controlled Transient...

Hepatic steatosis (HS) is a common liver condition characterized by excessive fat accumulation in hepatocytes. In the present study, we developed a machine learning (ML)-based predictive model for HS defined by vibration-controlled transient elastography (VCTE). Key predictive factors were further interpreted using SHapley Additive exPlanations (SHAP).

This study used data from two NHANES cycles, 2017-2020 and 2021-2023. The Boruta algorithm and LASSO regression were applied for feature selection. Six ML models, including logistic regression (LR), random forest (RF), multilayer perceptron (MLP), support vector machine (SVM), K-nearest neighbor (KNN), and extreme gradient boosting (XGBoost), were used to build prediction models. SHAP was used to interpret model predictions and identify the most influential predictors of HS. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

A total of 12,177 participants were included in the current study. Boruta and LASSO feature selection identified 14 significant predictors, comprising age, BMI, gender, glycohemoglobin, HDL cholesterol, hemoglobin, hypertension, lymphocyte, manganese, monocyte, race, segmented neutrophils, selenium, and total cholesterol. XGBoost achieved the highest AUC (0.832), followed by LR (0.826), RF (0.825), SVM (0.818), MLP (0.792), and KNN (0.737). SHAP analysis of the XGBoost model highlighted BMI, glycohemoglobin, age, and HDL cholesterol as key contributors to HS prediction.

The XGBoost model showed the highest discriminative ability for predicting HS, with BMI, glycohemoglobin, age, and HDL cholesterol identified as the most important predictors. This interpretable model may support early identification and risk stratification of HS.