Comparative evaluation of standard machine learning models for liver fibrosis detection

Background: Liver fibrosis progressing to cirrhosis is common in chronic liver disease, often leading to severe complications. Early detection is crucial, but current serological markers are inadequate. This prompted the analysis of explainable machine learning models to improve fibrosis detection.

Materials and Methods: We analyzed 655 patients who underwent liver biopsies, with clinical and laboratory parameters extracted retrospectively. Machine learning models, including tree-based models as well as classical and deep learning methods, were used for binary and three-stage liver staging. Independent validation was conducted on 302 patients from an independent hospital. Models were trained, hyperparameter-tuned, and tested on the collected data, demonstrating robust performance in fibrosis classification.

Results: The accuracy of machine learning models for predicting moderate liver fibrosis, severe fibrosis, and cirrhosis using blood markers was robust, with accuracies reaching up to 88.46%, 92.31%, and 82.69% respectively. The tree-based models, LightGBM, XGBoost, and Random Forest, performed best across various classification tasks with an accuracy range of 82.69% to 92.31% for binary classification, and an accuracy of 76.95% for three-stage classification, significantly outperforming FIB-4. SHAP analysis of the best ensemble models identified platelets, MCV, and INR, as the most influential biomarkers, with models using only these parameters achieving comparable performance to those using the full set of biomarkers for the cirrhosis classification (ACC max. 86.52%).

Conclusion: Machine learning models can significantly improve the prediction of liver stages compared to serum-based tests alone. Platelets, MCV, and INR are considerably more important than previously thought.

Publication History

Article published online:
20 January 2025