Feature Importance Analysis of Clinical Variables in Liver Disease Prediction Using Machine Learning Algorithm

Liver disease is a major global health concern, responsible for over two million deaths annually. Detecting disease before symptoms become apparent is usually hindered by non-specific early symptoms and limited diagnostic infrastructure, especially in low-resource settings. This study proposes a machine learning-based approach to predict liver disease and identify the most influential clinical variables. Seven supervised classification algorithms were used—Random Forest, Logistic Regression, XGBoost, Gradient Boosting, Support Vector Classifier, LightGBM, and CatBoost. To ensure model reliability and avoid overfitting, training was conducted utilizing stratified 5-fold cross validation, with performance evaluated through recall, precision, accuracy, and F1-score. CatBoost outperformed the other models, achieving 92.18% accuracy, 92.17% precision, 92.18% recall, and a 92.17% F1-score. SHAP (SHapley Additive exPlanations) analysis was applied to the CatBoost model to interpret feature contributions, revealing that Alcohol Consumption and Liver Function Tests had the highest average SHAP values (1.609 and 1.559, respectively), making them the most influential predictors. SHAP beeswarm plot visualization further showed that higher values for most features increased the predicted risk of liver disease, while higher physical activity was associated with a lower risk. These findings demonstrate that machine learning models have strong potential in supporting medical decision-making for liver disease diagnosis and that SHAP improves model interpretability, enabling more trustworthy clinical insights.