A machine learning-based predictive model for 48-week hepatitis B surface antigen seroclearance in chronic hepatitis B patients treated with pegylated interferon α-2b: prediction at week 24.

Background: Chronic hepatitis B (CHB) is an infectious disease mainly affecting the liver, caused by the hepatitis B virus (HBV). In the treatment of CHB, pegylated interferon α-2b (PEG-IFNα-2b) is one of the important therapeutic options. However, there are significant individual differences in patients' responses to this treatment and only a few patients can achieve hepatitis B surface antigen (HBsAg) seroclearance. Therefore, an effective method to identify patients with a high likelihood of favorable response at an early stage is urgently needed. Methods: In this study, we analyzed data from CHB patients who received antiviral treatment with PEG-IFNα-2b and completed 48 weeks of follow-up in the "OASIS" Project. Patients were divided into the seroclearance group and the non-seroclearance group based on whether HBsAg seroclearance was achieved at week 48.Five distinct machine learning feature selection algorithms were used to identify the optimal predictive variables for HBsAg seroclearance. These key variables were then incorporated into 12 machine learning algorithms to build predictive models for HBsAg seroclearance. The best-performing model was selected, and its performance was evaluated. Results: A total of 680 subjects were included in this study, comprising 165 in the 48-week seroclearance group and 515 in the 48-week non-seroclearance group. Through five different machine learning feature selection algorithms, 11 variables were identified and used to construct 12 distinct machine learning models. Comparative analysis of these models, based on the Area Under the Receiver Operating Characteristic Curve (AUC) and Decision Curve Analysis (DCA) results from the training set, indicated that the Random Forest model was the optimal model for predicting HBsAg seroclearance. Conclusion: The Random Forest model effectively predicted the 48-week HBsAg seroclearance rate using indicators measured at 24 weeks of PEG-IFNα-2b therapy. This model can provide a reliable reference for optimizing clinical treatment strategies.