Assisting the Non-invasive Diagnosis of Liver Fibrosis Stages using Machine Learning Methods

Area of Science:

• Hepatology
• Medical Informatics
• Machine Learning in Medicine

Background:

• Liver fibrosis, a key indicator of chronic liver disease, is often caused by hepatitis C Virus (HCV).
• The increasing prevalence of HCV infection globally highlights the need for cost-effective and accurate fibrosis assessment methods.
• Current fibrosis assessment methods are expensive and have limitations.

Purpose of the Study:

• To identify key features for staging liver fibrosis non-invasively.
• To generate clinical decision rules for physicians using these features.
• To compare the performance of machine learning classifiers for fibrosis staging.

Main Methods:

• Utilized machine learning algorithms including Multi-layered Perceptron (MLP), Random Forest, and Logistic Regression.
• Employed a Decision Tree approach to generate a reduced set of diagnostic rules.
• Evaluated classifier performance on full and reduced feature sets for predicting fibrosis stages.

Main Results:

• The Decision Tree model generated 28 rules with a prediction accuracy of 97.45% for liver fibrosis staging.
• This contrasts with previous methods generating over 98,000 rules with 99.97% accuracy.
• Multi-layered Perceptron (MLP) demonstrated the highest accuracy among the evaluated machine learning classifiers.

Conclusions:

• A concise set of rules derived from machine learning can accurately stage liver fibrosis non-invasively.
• This approach offers a promising alternative to invasive diagnostic procedures.
• MLP shows superior performance for liver fibrosis staging using the identified features.