Federated learning with explainable AI for liver disease prediction: A privacy-preserving approach
Deepak Kumar, Chaman Verma, Zoltán Illés · Intelligence-Based Medicine · 2025
Liver disease represents a major global health concern, demanding early and accurate detection to improve patient outcomes. Traditional machine learning (ML) models for liver disease prediction often require centralized data collection, raising privacy concerns and lacking sufficient interpretability for clinical adoption. The objective of this study is to develop a framework that preserves data privacy while providing transparent and reliable diagnostic insights. We present a Federated Learning with Explainable AI (FL-XAI) framework, integrating an ensemble of calibrated ML models—Random Forest (RF), Gradient Boosting (GBC), AdaBoost, Logistic Regression (LR), and Decision Tree (DT) — trained across five decentralized client nodes on stratified partitions of a real-world liver disease dataset. Model interpretability is enhanced via Shapley Additive Explanations (SHAP), and probability calibration is performed using isotonic regression to improve confidence reliability. The FL-XAI framework achieves 99% classification accuracy, 98% F1-score, a Brier Score of 0.01, and Expected Calibration Error (ECE) of 0.59, demonstrating strong predictive performance and reliable probability estimates. SHAP analysis identifies Direct Bilirubin, SGOT, and Alkaline Phosphatase as key predictive features, aiding clinical trust. Compared to centralized models, our approach matches or exceeds performance metrics while preserving privacy. This FL-XAI system offers a scalable, privacy-preserving and interpretable solution for liver disease prediction. Its calibrated and explainable predictions address key clinical adoption challenges, making it well-suited for deployment in regulated healthcare environments. • FL-XAI enables secure, decentralized ML for liver disease with 99% accuracy. • Model shows 98% F1-score, 0.01 Brier Score; isotonic regression improves calibration. • Direct Bilirubin, SGOT, and Alkaline Phosphatase are key predictors with SHAP explainability.