Leveraging serum CHI3L1 and machine learning for non-invasive liver fibrosis diagnostics

This study was aimed to develop non-invasive liver fibrosis diagnostic model using chitinase 3-like protein 1 (CHI3L1) based on machine learning (ML).1,011 follow-up patients (556 training/455 validation) were enrolled, staging fibrosis via transient elastography (F01, F2, F23, F34, and F4). Demographic, clinical, and laboratory parameters were collected. Fibrosis-related indicators were selected by differential analysis and correlation analysis. ML models were developed, optimized, and validated. Of 23 initial laboratory indicators, 15 indicators were selected to developed a 5-stage liver fibrosis diagnostic model (XG-Boost, FIB5-15, area under the curve [AUC] = 0.93, sensitivity = 0.70, specificity = 0.93). This was refined to a 7-indicator model (FIB5-7, area under the ROC curve [AUC] = 0.98). By merging F2/F23 and F34/F4, a 3-category optimized model (FIB3-7, AUC = 0.94) and a 2-category model (FIB2-7, AUC = 0.95) were created to identify reversible fibrosis. The final FIB2-7 model significantly outperformed aspartate aminotransferase‑to‑platelet ratio index (APRI) and FIB-4 (p < 0.001). The CHI3L1-related liver fibrosis models are applicable across multiple clinical scenarios.