Validation of a machine learning approach using FIB-4 and APRI scores assessed by the metavir scoring system: A cohort study

Abstract

Background and study aim: The study aim was to improve and validate the accuracy of the fibrosis-4 (FIB-4) and aspartate aminotransferase-to-platelet ratio index (APRI) scores for use in a potential machine-learning (ML) method that accurately predicts the extent of liver fibrosis.

Patients and methods: This retrospective multicenter study included 69,106 patients with chronic hepatitis C planned for antiviral therapy from January 2010-December 2014 with liver biopsy results. FIB-4 and APRI scores were calculated and their performance for predicting significant liver fibrosis (F3-F4) assessed against the Metavir scoring system. ML was used for feature selection and reduction to identify the most relevant attributes (CfsSubseteval/best first) for prediction.

Results: In this study, 57,492 (83.2%) patients were F0-F2, and 11,615 (16.8%) patients were F3-F4. The revalidation of FIB-4 and APRI showed lower accuracy and higher disagreement with the biopsy results, with AUCs of 0.68 and 0.58, respectively. FIB-4 diagnosed fewer (14%) F3-F4 patients, and the high specificity and negative predictive values of FIB-4 and APRI reflected the low prevalence of F3-F4 in the study population. Out of 15 attributes, age (>35 years), AFP (>6.5 ng/ml), and platelet count (<150,000/mm³) were the most relevant risk attributes, and patients with one or more of these risk factors were likely to be F3-F4, with a classification accuracy of ≤ 92% and receiver operating characteristics area of 0.74.

Conclusion: FIB-4 and APRI scores were not very accurate and missed diagnosing most of the F3-F4 patients. ML implementation improved medical decisions and minimized the required clinical data to three risk factors.

Keywords: APRI score; Attribute reduction; FIB4 score; Hepatitis C virus; Liver fibrosis; Machine learning; Metavir.