Explainable machine learning incorporating CHI3L1 enhances liver fibrosis staging in chronic hepatitis B: The CHILI multicenter study.

BACKGROUND

Identifying significant liver fibrosis (≥F2) is critical for initiating antiviral therapy in chronic hepatitis B (CHB). Chitinase-3-like protein 1 (CHI3L1) is a promising fibrosis biomarker, but its staging value remains unclear. We evaluated CHI3L1 (chemiluminescent immunoassay, CLIA) alone and with machine learning (ML) for differentiating mild (F0-F1) from significant (F2-F4) fibrosis.

METHODS

We included 347 participants (303 primary, 44 external validation), most CHB patients with biopsy-confirmed METAVIR staging. Serum markers were measured, and 5 ML models were developed.

RESULTS

CHI3L1 levels increased progressively with fibrosis severity and discriminated significant fibrosis (≥F2), with AUC 0.84 (95% CI 0.772-0.908) in the overall CHB cohort and 0.839 (95% CI 0.755-0.924) in CHB patients with normal ALT, outperforming conventional biomarkers. Incorporating CHI3L1 into ML further improved diagnostic performance; the Decision Tree model achieved optimal accuracy AUCs of 0.923 [95% CI 0.891-0.976] and 0.895 [95% CI 0.831-0.960], respectively. SHapley Additive exPlanations (SHAP) analysis identified CHI3L1 as the most influential feature (62.2%) in this model. The model performed consistently in the external validation cohort.

CONCLUSIONS

CHI3L1 is a strong noninvasive biomarker for identifying clinically significant fibrosis in CHB. Integration with ML enhances fibrosis staging accuracy, supporting timely antiviral therapy.