Machine Learning-Based Prediction of Decompensation in Hepatitis B Virus-Related Cirrhosis

TL;DR

This study uses machine learning to predict decompensation in hepatitis B-related cirrhosis patients, showing promising accuracy with electronic health records.

Contribution

The novel use of self-developed machine learning models to predict decompensation in HBV-related cirrhosis patients using electronic health records.

Findings

SVM and RF models achieved AUROCs of 0.85-0.93 and accuracy scores of 0.77-0.88 for predicting decompensation complications.

SVM and LR models performed best in predicting ascites among entecavir users with AUROCs of 0.93 and 0.92.

Antiviral treatment details and clinical data are key predictors for decompensation.

Abstract

Background/Objectives: Fatality of cirrhotic patients greatly increases when they progress to the decompensated state. Only a few studies to date have applied machine learning (ML) methods to predict decompensation in cirrhosis patients. In the present study, we attempted to apply self-developed ML models for validating their capability of predicting different complications in hepatitis B virus (HBV)-related cirrhosis patients. Methods: Data were extracted from electronic health records of 50,047 patients who were tested and diagnosed with HBV in a tertiary hospital. Four different algorithms (Support Vector Machine (SVM), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF)) were utilized, and a total of 32 ML models were trained and tested to predict variceal bleeding, ascites, jaundice, and multiple complications (≥2 complications) in HBV-related cirrhosis patients. The use of two antiviral drugs were considered: entecavir (ETV) and lamivudine (LAM). Performance of the models was assessed using area under receiver operating characteristic curve (AUROC) and accuracy score. Results: SVM and RF classifications produced the best overall predictions for decompensation in HBV-related cirrhosis patients, with AUROCs ranging from 0.85 to 0.93 and accuracy scores between 0.77 to 0.88 for ascites, jaundice, and multiple complications. The SVM and LR algorithms generated the best performance in differentiating ascites among ETV users, with AUROC of 0.93 and 0.92 and accuracy of 0.88 and 0.86, respectively. Antiviral treatment (type, length of use, adherence), and other routinely collected clinical information may serve as informative markers in differentiating decompensated cirrhosis. Conclusions: ML-based prediction of decompensation using electronic health records may assist clinicians in decision making. Findings of this study also underline the impact of antiviral therapy as a key predictor for decompensation.

Figures