Assisting the diagnosis of cirrhosis in chronic hepatitis c patients based on machine learning algorithms: a novel non-invasive approach

Abstract

Aim: This study aimed to determine the important features and cut-off values after demonstrating the detectability of cirrhosis using routine laboratory test results of chronic hepatitis C (CHC) patients in machine learning (ML) algorithms. Methods: This retrospective multicenter (37 referral centers) study included the data obtained from the Hepatitis C Turkey registry of 1164 patients with biopsy-proven CHC. Three different ML algorithms were used to classify the presence/absence of cirrhosis with the determined features. Results: The highest performance in the prediction of cirrhosis (Accuracy = 0.89, AUC = 0.87) was obtained from the Random Forest (RF) method. The five most important features that contributed to the classification were platelet, αlpha-feto protein (AFP), age, gamma-glutamyl transferase (GGT), and prothrombin time (PT). The cut-off values of these features were obtained as platelet < 182.000/mm3, AFP > 5.49 ng/mL, age > 52 years, GGT > 39.9 U/L, and PT > 12.35 s. Using cut-off values, the risk coefficients were AOR = 4.82 for platelet, AOR = 3.49 for AFP, AOR = 4.32 for age, AOR = 3.04 for GGT, and AOR = 2.20 for PT. Conclusion: These findings indicated that the RF-based ML algorithm could classify cirrhosis with high accuracy. Thus, crucial features and cut-off values for physicians in the detection of cirrhosis were determined. In addition, although AFP is not included in non-invasive indexes, it had a remarkable contribution in predicting cirrhosis. Trial Registration: Clinicaltrials.gov identifier: NCT03145844.