Machine learning for predicting treatment response in hepatitis C: integrating viral genomics and clinical data

Despite the high effectiveness of direct-acting antivirals (DAAs), a small but clinically important subset of patients with hepatitis C virus (HCV) infection experience treatment failure, relapse, suboptimal treatment duration, or incomplete post-treatment recovery. Machine learning may improve risk stratification by integrating clinical, virologic, genomic, and treatment-related data. This systematic review was conducted according to PRISMA guidelines and registered on the Open Science Framework. PubMed, Scopus, and Web of Science were searched from inception to July 2026. Eligible studies developed or evaluated artificial intelligence or machine learning models predicting SVR, treatment failure, relapse, treatment duration, or closely related therapeutic outcomes in HCV antiviral therapy. Risk of bias was assessed using a PROBAST-informed framework adapted for prediction-model studies. Eight study records met the inclusion criteria, representing seven largely distinct populations. Prediction targets included DAA failure/SVR, short-duration treatment response, prolonged treatment duration, interferon/ribavirin response, and post-SVR liver-stiffness regression. Large registry-based models showed modest-to-good discrimination, with validation C-statistics or AUROCs ranging from approximately 0.64 to 0.83. Viral whole-genome models reported high internal accuracy in selected cohorts but were limited by small sample size and lack of external validation. Recurrent predictors included HCV RNA, genotype, DAA class, treatment duration, adherence or discontinuation, cirrhosis, hepatocellular carcinoma, liver function markers, platelet count, FIB-4, and viral sequence-derived variants. Machine learning models show promise for HCV antiviral-response prediction and treatment optimization, but current evidence shows that these remain at the development stage. Robust external validation, calibration, clinical utility assessment, and multimodal integration are required before routine implementation.