Machine learning model for predicting hepatitis C seroconversion in methadone maintenance patients in China.

Introduction: Hepatitis C virus (HCV) infection is a substantial public health concern, particularly among individuals with opioid addiction. The methadone maintenance treatment (MMT) programmes serve as a harm reduction strategy to mitigate HIV disease spread, yet the risk of HCV infection remains high within these settings. Accurate risk prediction for HCV seroconversion is therefore crucial for improving patient outcomes. Methods: We collected data from 1547 individuals with opioid use disorder who entered the MMT programme from May 2005 to October 2023 in Shanghai, China, and 283 individuals from July 2006 to October 2023 in Mianyang, China, whose HCV infection status was monitored. Shanghai data were divided into training and validation sets in a 7:3 ratio, with 70% of the Shanghai samples used for model training and the remaining 30% reserved for internal validation. Additionally, the Mianyang dataset was employed as an independent external validation cohort to assess the model's generalisability. Four machine learning models were developed. We then validated the predictive performance of the model using C-index, receiver-operating characteristic curves, calibration plots and decision curve analysis. Results: 13 predictive factors, including sex, age, ethnicity, education, occupation status, marriage status, living status, financial resource, drug use method in the past half year, injected drug last month, condom use, average methadone dosage and positive rate of drug urine tests, were all incorporated into the predictive model. The eXtreme Gradient Boosting (XGBoost) model exhibited superior performance in both discrimination and calibration compared with the other three models. Specifically, it achieved C-indices of 0.793 (95% CI: 0.771 to 0.813) in the training cohort, 0.744 (0.709 to 0.779) in the internal validation cohort and 0.756 (0.712 to 0.799) in the external validation cohort for predicting HCV seroconversion. A publicly accessible web tool was generated for the model. Conclusions: The developed XGBoost model has the potential to accurately predict individuals on MMT programmes at high risk of HCV seroconversion.