Development of a clinical prediction model for parenteral nutrition-associated cholestasis in preterm infants: a retrospective cohort study

Abstract

Abstract Background Parenteral nutrition-associated cholestasis (PNAC) remains a severe and potentially life-threatening complication in preterm infants requiring extended parenteral nutrition. This study aimed to identify independent risk factors for PNAC and develop a robust predictive model to guide clinical decision-making. Methods This retrospective study reviewed medical records of preterm infants admitted to a tertiary NICU between July 2020 and June 2025. Inclusion criteria comprised a gestational age (GA) < 37 weeks, birth weight (BW) < 2000 g, and a requirement for parenteral nutrition (PN) for ≥ 14 consecutive days. Infants diagnosed with necrotizing enterocolitis (NEC) or those who underwent major gastrointestinal surgery were excluded. Out of 389 screened neonates, 329 met the criteria and were included in the final analysis. Demographic, clinical, and comprehensive nutritional parameters were analyzed. Following collinearity diagnostics to ensure model stability, multivariable logistic regression was employed to identify independent risk factors. Model performance was evaluated using receiver operating characteristic (ROC) curves, bootstrap internal validation, subgroup analysis, and decision curve analysis (DCA). Results Among 329 included preterm infants, 41 (12.5%) developed PNAC. After adjusting for confounders, multivariable logistic regression identified lower BW, treated patent ductus arteriosus (PDA), maximum amino acid (maxAA) dosage, and Day 14 enteral feeding volume (Day 14 EF) < 25 mL/kg/day as independent predictors of PNAC. The predictive model demonstrated excellent discriminative performance with an area under the curve (AUC) of 0.859 (95% CI: 0.806–0.912, p < 0.001). At the optimal logit cutoff of -2.720, sensitivity was 85.4% and specificity was 77.1%. Individual optimal thresholds were ≤ 1375 g for BW (sensitivity: 69.4%, specificity: 82.9%) and ≥ 4.01 g/kg/day for maxAA (sensitivity: 48.8%, specificity: 83.0%). DCA demonstrated that the model consistently maintained a higher net benefit than “treat-all” and “treat-none” strategies within a threshold probability range of 0% to 45%. Conclusion Lower BW, treated PDA, higher maxAA dosage (4.01 g/kg/day), and delayed enteral feeding (Day 14 EF < 25 mL/kg/day) are independent risk factors for PNAC. The validated predictive model offers high diagnostic accuracy and clinical utility in guiding targeted prevention and management strategies.