Evaluation of Machine Learning Algorithms for Early Prediction of Liver Disease

Abstract

Liver disease contributes to major morbidities and mortalities globally. Around 2 million people die annually worldwide from liver diseases. Diagnosing liver disease is usually made by measuring levels of biomarkers that could be enzymes linked to liver function tests and/or proteins. However, not often the levels of these biomarkers change in cases of liver diseases. The majority of diagnosis based on biomarkers occur when the liver is partially or fully damaged. On the other hand, early diagnosis of the disease can prevent further complication, decrease the burden on healthcare systems and save lives. Subsequently, this research evaluated machine learning algorithms for early prediction of liver disease. These algorithms included logistic regression, decision tree, random forest, adaptive boosting, extreme gradient boost, support vector machine and Naïve Bayes and were applied to a dataset of patients with and without liver diseases after data preprocessing. Metrics for evaluation included accuracy, precision, recall, AUC-ROC and F1-score. Based on these metrics and after hyperparameter tuning, support vector machine, random forest, adaptive gradient boost and extreme gradient boost showed to be the best performing models. However, the performance of these models did not exclude overfitting that could be related to the low sample size of the dataset. Future work involves applying these algorithms to a larger sample size of patients and more features.