Predicting neonatal mortality among very preterm infants: a comparison of three versions of the CRIB score

Objective: To validate Clinical Risk Index for Babies (CRIB) and CRIB II mortality prediction scores in a UK population of infants born at <= 32 weeks' gestation, and investigate CRIB II calculated without admission temperature. Methods: Infants born at 22-32 weeks' gestation to mothers resident in a UK region in 2005-2006 admitted for neonatal care were identified. Predictive probabilities for mortality were calculated using CRIB, CRIB II and CRIB II without admission temperature (CRIB II(-T)) using published algorithms and after recalibration. Predictive performance was investigated overall and for groups defined by gestation and admission temperature and summarised by area under receiver-operating curve, Cox's regression, Brier scores and Spiegelhalter's z-scores. Results: 3268 infants were included: 317 (9.7%) died before discharge. Using published algorithms each score showed excellent discrimination (area under the curve = 0.92). The total number of deaths was predicted well for CRIB (324.4) but for both versions of CRIB II the number of deaths was underpredicted (255.2 and 216.6). All scores performed poorly for subgroups. After recalibration CRIB II displayed excellent predictive characteristics overall (Spiegelhalter's z-score p = 0.52) and in the gestation groups (p = 0.44 and 0.57) but not for the temperature groups (p = 0.026 and 0.97). CRIB II(-T) displayed excellent predictive characteristics for all groups: overall p = 0.53; gestation groups p = 0.64 and 0.42; temperature groups p = 0.42 and 0.66. Conclusions: The published algorithm for CRIB 11 was poorly calibrated but simple linear recalibration provided good results. The CRIB II score without admission temperature showed good predictive characteristics once recalibrated and this version of the score should be used when benchmarking mortality in neonatal intensive care units.