An explainable ensemble learning approach for liver disease risk prediction with interpretable clinical summaries
Open Access
- 10 November 2026
- journal article
- research article
- Published by Frontiers Media SA in Frontiers in Artificial Intelligence
- Vol. 9, 1988857
Abstract
Background: Liver disease causes more than 2 million annual deaths, often asymptomatic until the early stages. Routine liver function tests (LFTs) are typically interpreted univariately. We developed an explainable machine learning (ML) model for liver disease risk prediction using LFTs. Methods: Models (logistic regression, random forest, gradient boosting, soft-voting ensemble) were trained on the Kaggle Liver Disease Patient Dataset (LPD, n = 28,948 after removing 1,743 overlaps) and externally validated on the UCI Indian Liver Patient Dataset (ILPD, n = 583). Performance was assessed via AUC and DeLong’s test; SHAP provided interpretability and patient-level summaries. Results: The gradient boosting model achieved the best discrimination, with an AUC of 0.96 in the internal holdout set and 0.967 (95% CI 0.953–0.980) in the ILPD validation set, significantly outperforming random forest and the soft-voting ensemble (both AUC 0.939) and logistic regression (AUC 0.759, all p < 0.001). At the conventional 0.50 threshold, the model identified 410/416 liver disease cases (sensitivity 98.6%, accuracy 88.5%). At an optimised threshold of 0.55, accuracy improved to 91.6%, with sensitivity 97.4%, specificity 77.2%, precision 91.4%, and F1-score 0.943; average precision was 0.99 on ILPD. Five-tier risk stratification showed a stepwise increase in observed prevalence from 0% (Very Low) to 98.7% (Very High); the two lowest tiers contained only 1/85 cases, while the two highest tiers captured 97.4% of all cases. SHAP analysis identified direct bilirubin (mean |SHAP| 0.427), alanine aminotransferase (0.367), and alkaline phosphatase (0.355) as the strongest predictors; age and sex contributed minimally. Automated SHAP-based summaries highlighted key contributing biomarkers per patient and flagged discordant findings for review. Conclusion: An explainable gradient boosting model using routine LFTs delivers high-accuracy liver disease prediction, with transparent risk stratification and SHAP-based clinical summaries to aid early detection and triage.Keywords
This publication has 36 references indexed in Scilit:
- Chronic liver disease classification using deep learning with SHAP-optimized hybrid featuresiScience, 2025
- Feature Importance Analysis of Clinical Variables in Liver Disease Prediction Using Machine Learning AlgorithmProcedia Computer Science, 2025
- A focal loss and sequential analytics approach for liver disease classification and detectionHealthcare Analytics, 2025
- Machine Learning and Explainable AI for Liver Disease Prediction: An Integrated Interpretability FrameworkBiomedical Materials & Devices, 2025
- Leakage and the reproducibility crisis in machine-learning-based sciencePatterns, 2023
- Global burden of liver disease: 2023 updateJournal of Hepatology, 2023
- Prediction of chronic liver disease patients using integrated projection based statistical feature extraction with machine learning algorithmsInformatics in Medicine Unlocked, 2022
- Explaining individual predictions when features are dependent: More accurate approximations to Shapley valuesArtificial Intelligence, 2021
- A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction modelsJournal of Clinical Epidemiology, 2019
- A new framework to enhance the interpretation of external validation studies of clinical prediction modelsJournal of Clinical Epidemiology, 2015