A comparative analysis of machine learning classifiers for chronic liver disease prediction in rats.

Abstract

OBJECTIVE: The present study investigates machine learning (ML) models to predict the stage of chronic liver disease (CLD) in experimental rats using non-invasive biomarker levels. It aims to establish a reliable ML-based framework for staging CLDs in experimental rats, enabling real-time disease monitoring and reducing reliance on histopathological assessments. METHODS: Data was compiled from previously published preclinical research on liver fibrosis, cirrhosis, and hepatocellular carcinoma (HCC). Experimental studies used thioacetamide (TAA) as the inducing agent and employed rats in the dataset. The dataset was pre-processed, and features such as TAA dose, dosing frequency, aspartate aminotransferase, and alanine aminotransferase levels were chosen for training the ML models. The six ML classifiers: K-nearest neighbors, decision tree, random forest, support vector machine, XGBoost, and CatBoost were trained and tested using stratified k-fold cross-validation. RESULTS: CatBoost showed the best prediction performance with 99.3% accuracy and an area under the receiver operating characteristic curve of 0.999. The model was able to distinguish between fibrosis, cirrhosis, and HCC with high sensitivity and specificity. CONCLUSION: Use of ML classifiers for disease prediction offers a promising alternative to traditional animal experimentation. Integrating ML into preclinical disease investigations can improve experimental repeatability, reduce variability, and optimize study designs.