Explainable Ensemble Machine Learning for Liver Cirrhosis Mortality Prediction Using Clinical Data

Abstract

Liver cirrhosis is a progressive and life-threatening chronic liver condition associated with high morbidity and mortality worldwide. Early and accurate risk assessment using non-invasive clinical and laboratory parameters can assist healthcare professionals in timely intervention and personalized patient management. Existing machine learning-based approaches have demonstrated the potential for liver disease classification; however, limitations remain in terms of prediction reliability, clinical relevance, and model interpretability. The existing study developed an ensemble machine learning framework for liver disease stage prediction using clinical features and reported an accuracy of 83% using a hyperparameter-tuned Logistic Regression model. However, multi-class histological stage prediction remains challenging due to overlapping clinical characteristics among different disease stages. In this work, a clinically meaningful binary mortality prediction framework is developed using the Mayo Clinic Primary Biliary Cirrhosis (PBC) dataset to classify patient outcomes as Death versus Alive/Censored. A CatBoost-based ensemble learning model is proposed and evaluated using stratified five-fold cross-validation and an independent test set. The proposed model achieved a cross-validation accuracy of 79.90% and obtained an improved test accuracy of 86.90%, with precision, recall, and F1-score values of 0.87, 0.87, and 0.87, respectively. Furthermore, SHAP (SHapley Additive exPlanations) is integrated to enhance model transparency by identifying influential clinical factors and generating patient-specific explanations for individual predictions. The proposed explainable ensemble framework provides a reliable and interpretable approach for mortality risk assessment in cirrhosis patients, supporting data-driven clinical decision-making while maintaining reproducibility and transparency.