A minimalist and robust diagnostic model for neonatal biliary atresia: Harnessing MMP-7 and machine learning in a...

Early diagnosis is the most critical determinant of native liver survival in infants with biliary atresia (BA). Accurate differentiation of BA from other neonatal cholestatic disorders remains particularly challenging within the first 28 days of life, a period during which clinical and biochemical features are often nonspecific. This study aimed to develop and validate a minimalist, highly interpretable diagnostic model that integrates serum matrix metalloproteinase-7 (MMP-7) with routinely available clinical parameters to facilitate early BA diagnosis in neonates.

This retrospective diagnostic study enrolled neonates ≤28 days of age admitted for conjugated hyperbilirubinemia between 2021 and 2025. The final cohort consisted of 50 infants (29 BA, 21 non-BA) with complete data for four pre-specified variables: stool color, gamma-glutamyl transferase (GGT), direct bilirubin (DBIL), and serum MMP-7. Complete-case analysis was used (no imputation). Two interpretable models-a logistic regression and a random forest classifier-were developed. To ensure robust performance estimation, we performed repeated stratified 5-fold cross-validation (50 repeats; 250 iterations). A single 70/30 train-test split was retained solely for illustrative purposes. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, and decision curve analysis (DCA).

Serum MMP-7 levels were significantly higher in BA infants compared to controls (66.6 ± 25.4 vs. 19.2 ± 16.7 ng/mL, P < 0.001) and remained the strongest independent predictor in multivariable analysis (adjusted odds ratio 1.13 per ng/mL increase, P = 0.002). In repeated cross-validation, the logistic regression model achieved a mean AUC of 0.978 ± 0.047, sensitivity 0.955 ± 0.096, specificity 0.922 ± 0.134, and F1-score 0.949 ± 0.072, consistently outperforming random forest (mean AUC 0.951 ± 0.087). The single hold-out test set produced perfect discrimination for logistic regression (AUC = 1.00, 100% sensitivity and specificity), illustrating its potential but representing an optimistic upper bound. MMP-7 was the dominant contributor in both models (logistic regression standardized coefficient = 1.819; random forest Gini importance = 0.515). DCA confirmed positive net benefit across clinically relevant thresholds. To enable bedside use, we provide the explicit logistic regression equation and two types of nomograms.

In this single-center cohort of neonates aged ≤28 days, a minimalist logistic regression model integrating serum MMP-7 with three routine parameters demonstrated excellent and stable discriminative performance (mean AUC 0.978) and is readily deployable as a formula or nomogram. The observed performance estimates, while robust in internal cross-validation, are likely optimistic and require external validation in large, prospective multicenter cohorts before widespread clinical implementation. If confirmed, this model could advance BA diagnosis into the first month of life and improve native liver survival.