Proteomics-based machine learning model for predicting secondary infection in HBV-related liver failure

Patients with Hepatitis B Virus-related liver failure are highly vulnerable to secondary infections (SI), yet early predictive tools remain limited. In this work, we aim to develop and validate a plasma proteomics-based model for early SI risk assessment. In a prospective multicenter study, 114 patients are enrolled in the discovery cohort, 60 each in two validation cohorts. Untargeted proteomics is used to identify SI-related proteins, followed by Minimum Redundancy Maximum Relevance based feature selection and logistic regression modeling. Targeted proteomics and ELISA are applied for external validation. Inflammatory and coagulation pathway dysregulation is strongly associated with SI. A final model including Lysozyme (LYZ), Calmodulin 1 (CALM1), Serpin Family D Member 1 (SERPIND1), Dermatopontin (DPT), total bilirubin, and AST show excellent discrimination (area under the receiver operating characteristic curve (AUROC) 0.980 in discovery; 0.873 in validation), outperforming C-reactive protein (CRP), white blood cell (WBC), and Neutrophil percentage (NE%). It also predicts 28-day mortality better than Chronic Liver Failure-Consortium Acute-on-Chronic Liver Failure score (CLIF-C ACLF) and Model for End-Stage Liver Disease (MELD). ELISA measurements in validation cohort 2 yield consistent trends, and an ELISA-based model achieve an AUROC of 0.883. This proteomics-derived model reliably identifies patients at high SI risk and supports early clinical intervention.