Value of XGBoost machine learning model for diagnosis of hepatitis B cirrhosis

Abstract

BACKGROUND The progression of chronic hepatitis B into cirrhosis is slow and easily ignored, and the construction of a noninvasive diagnostic model for cirrhosis based on routine clinical indicators has become a hot research topic. However, there is still a lack of machine learning models regarding the early diagnosis of cirrhosis. AIM To investigate the performance of the extreme gradient boosting (XGBoost) machine model in the diagnosis of hepatitis B cirrhosis. METHODS A retrospective analysis was performed on 1087 patients with chronic hepatitis B virus infection (CHBV) diagnosed for the first time at the Department of Infection, The First/Second Affiliated Hospital of Anhui Medical University from 2010 to 2018. The patients were divided into training and validation sets in a 3:1 ratio according to the randomization principle. Clinical data of all study participants were collected and prediction models were constructed using XGBoost machine learning model. Meanwhile, the aspartate aminotransferase/platelet ratio index (APRI) and fibrosis-4 index (FIB-4) scores were calculated and compared with the XGBoost machine learning model. Area under the curve (AUC) was used to assess the model discrimination, and calibration curve (CA) and decision curve analysis (DCA) were used to assess the model calibration and benefit. RESULTS A total of 1087 CHBV patients were included, including 817 in the training set and 270 in the validation set. There was no statistical difference between the training and validation sets for all predictor variables (P > 0.05). Cirrhosis occurred in 103 patients in the training set, and APRI and FIB-4 scores were significantly higher in cirrhotic patients than in non-cirrhotic patients (P < 0.05). The relative importance of platelets was the highest among all predictors. The AUCs of the model in the training and validation sets were 0.95 and 0.86 (P < 0.05), respectively, and the Kappa values were 0.78 and 0.74, which suggested that the model was reproducible. CA curve analysis indicated that the model predicted a high degree of agreement with the true situation fit. DCA of the training and validation sets implied that the developed model could result in a high degree of benefit for patients. XGBoost machine learning model was significantly more efficient for the diagnosis of cirrhosis than APRI and FIB-4 scores. CONCLUSION The XGBoost machine learning model constructed in this study based on common clinical information of CHBV patients has an excellent performance for the diagnosis of cirrhosis and deserves further clinical promotion.