Developing an explainable machine-learning model for short-term identification of neonatal cholestasis in preterm infants: a retrospective matched case-control study

Neonatal cholestasis (NC) is an important hepatobiliary complication in preterm infants, but its short-term identification before the first clinical diagnosis remains challenging. We aimed to develop an explainable machine-learning model for short-term identification of NC using routinely collected clinical information available before the first documented diagnosis. This single-center retrospective matched case-control study included 280 preterm infants with NC and 280 controls matched by sex, gestational age, and year of hospitalization. Matched pairs were allocated to derivation (196 pairs) and held-out test (84 pairs) datasets. Boruta feature selection and hyperparameter optimization were restricted to the derivation dataset. Eight algorithms were evaluated using grouped five-fold cross-validation. Discrimination, calibration, classification performance, paired bootstrap model comparisons, sensitivity analyses, prevalence-dependent predictive values, and SHapley Additive exPlanations (SHAP) were assessed. Boruta confirmed 13 predictors spanning enteral and parenteral nutrition, inflammatory markers, hepatobiliary measurements, and hematological status. CatBoost had the highest derivation-set cross-validated area under the receiver operating characteristic curve (ROC-AUC) (0.803) and was selected before test evaluation. In the held-out test dataset, its ROC-AUC was 0.867 (95% CI, 0.815–0.919), area under the area under the area under the precision–recall curve (PR-AUC) was 0.875, and Brier score was 0.150. At a threshold of 0.50, sensitivity was 0.762 and specificity was 0.798. Results were broadly stable in sensitivity analyses. At the source-cohort NC prevalence of 7.55%, the expected positive and negative predictive values were 23.5% and 97.6%, respectively. Parenteral lipid exposure had the largest SHAP contribution. The internally validated model showed useful discrimination for short-term identification of NC in hospitalized preterm infants. It should be regarded as an exploratory clinical-warning model, not a diagnostic or causal tool. Recalibration and prospective external validation in unmatched multicenter cohorts are required before clinical implementation.