Abstract
BACKGROUND: Pediatric acute liver failure (PALF) can require emergent liver transplantation (LT, >25%) or lead to death (~15%). Existing models cannot predict clinical trajectory or survival with native liver (SNL). We aimed to create a predictive model for PALF clinical outcomes based on admission variables. METHODS: A retrospective, single-center PALF cohort (April 2003 to January 2022) was identified using International Classification of Disease codes, selected using National Institutes of Health PALF Study Group (PALFSG) criteria, and grouped by clinical outcome (SNL, LT, or death). Significant admission variables were advanced for feature selection using least absolute shrinkage and selection operator regression with bootstrapping (5000×). A predictive model of SNL versus LT or death was created using logistic regression and validated using PALFSG data. RESULTS: Our single-center cohort included 147 patients (58% SNL, 32% LT, 10% expired), while the PALFSG validation cohort included 492 patients (50% SNL, 35% LT, 15% expired). Admission variables associated with SNL included albumin (odds ratio [OR], 16; P < 0.01), ammonia (OR, 2.37; P < 0.01), and total bilirubin (OR, 2.25; P < 0.001). A model using these variables predicted SNL versus LT or death with high accuracy (accuracy [0.75 training, 0.70 validation], area under the curve [0.83 training, 0.78 validation]). A scaled score (CHLA-acute liver failure score) was created that predicted SNL versus LT or death with greater accuracy (C statistic 0.83) than Pediatric End-Stage Liver Disease (C statistic 0.76) and admission liver injury unit (C statistic 0.76) scores. CONCLUSIONS: The CHLA-acute liver failure score predicts SNL versus LT or mortality in PALF using admission laboratories with high accuracy. This novel, externally validated model offers an objective guide for urgent referral to a pediatric LT center.