Birthweight-adjusted mortality rates for assessing the effectiveness of neonatal intensive care

Mortality rates for very-low-birthweight infants vary significantly among different neonatal intensive care units (NICUs). Computational models and computer simulation are used to predict the performance of an algorithm for identifying individual NICUs within a network that have greater than 110% of the expected birthweight-adjusted mortality risk. The algorithm maintains high sensitivity and specificity with as few as three moderately heterogeneous risk categories when applied to large health care networks; the model parameters were based on preliminary data from a real NICU network. The performance of the algorithm depends on the number of admissions at the individual NICU. A NICU with a center-specific risk 130% of the network average would be correctly identified as an outlier 50% of the time if it had 35 admissions, 59% of the time if it had 70 admissions, and 77% of the time if it had 280 admissions. A NICU with average risk would be incorrectly identified as an outlier 16%, 12%, or 2% of the time if it had 35, 70, or 280 admissions, respectively. Severity-of-illness casemix adjustment did not improve these results. It is concluded that the sensitivity and specificity of the algorithm in determining which facilities have higher-than-expected mortality will be less in typical NICU networks than in large health care networks that treat adult patients. It is unlikely that severity-of-illness adjustments will overcome the problem of the small numbers of admissions at individual NICUs.