The regression and CART models identified factors associated with death and/or BPD in premature infants with respiratory failure. The magnitude of improvement in PaO2
in response to iNO was not found to be associated with death and/or physiologic BPD, indicating that the initial response to iNO in premature infants with severe respiratory failure may not be a good gauge of whether iNO should be continued. The main trial1
did not demonstrate any benefit of iNO on BPD/death in these infants despite initial improvement in oxygenation, and our study extends these observations by showing that even if there were a greater initial response to iNO, there was no reduction in death/BPD. In addition, we confirmed the previous finding that higher birth weight was a predictor in the iNO but not control infants. In general, both the regression and CART models identified smaller infants with a greater severity of illness as more likely to either die or develop BPD. While the findings that smaller and sicker infants have a worse outcome is not surprising, the CART analysis ranks the importance of these predictors and develops a prognostic algorithm using the predictors, which is a novel result of clinical relevance. The variables identified as important in these models and the optimal cut-points for these variables may prove useful in risk-stratifying premature infants with respiratory failure for future clinical trials. The models may also help in assessment of prognosis, in discussions with parents, and in generating hypotheses that can be tested in clinical trials.
The strengths of this study include the relatively large sample, recruitment from multiple tertiary care NICUs, and prospective data collection by trained observers. The outcomes of death or physiologic BPD at 36 weeks’ post-menstrual age are also relatively well defined. In addition, rather than empirical or expert-opinion derived decision trees, we used variables associated statistically with the outcomes of interest and cut-points for continuous variables which optimized discrimination between those with and without these outcomes. For example, cut-points were identified at 1072 g for birth weight, 90% or more for FiO2, and 2 or more doses of surfactant. The cut-off of 1072g is closer to 1100g rather than 1000g, suggesting that infants who are between 1000g and 1100 g may not be ELBW by definition but may be at similar risk to ELBW infants. These cut-points may prove useful in stratifying infants for future clinical trials, instead of using empirical criteria such as <1000 g or <1250 g birth weight. A limitation is that despite the relatively large sample size for this population, it was not feasible to use a split-half cross-validation approach in which we could develop the model in half of the data set and test it in the other half. Therefore, these models need to be validated using other data sets. Also, it is difficult to distinguish between illness severity and aggressive therapy, as indicators of illness severity (e.g. OI, FiO2) primarily reflect the aggressiveness of therapy rather than the magnitude of underlying lung disease.
Other investigators have developed prediction models for BPD, which have also identified similar variables (lower birth weight or gestational age and increased severity of respiratory illness) as risk factors for BPD.7,8,9
Recently completed trials of iNO in premature infants2-4
indicate that iNO may benefit some premature infants, generally in less ill populations. However, infants evaluated in the current study were smaller and sicker with a very high oxygenation index, and a large proportion either died or developed BPD. Therefore, our models may be more suitable for use in premature infants with severe respiratory failure and may be less suitable for less sick infants. These models are suitable for risk stratification or assessment of prognosis but should not be a basis for decisions regarding withdrawal of support, unless validated on individual center data.
In any prognostic system, certain variables (e.g. pH, PaCO2) are relatively objective while others dependent on clinical examination or clinical judgment (e.g. use of surfactant or high-frequency ventilation) may be subjective, leading to significant inter-center and inter-rater variation. Although the center variable was significantly associated with death and/or BPD by preliminary regression analysis, the center variable was not included for the final analysis, as this limits generalizability of the prediction models, and the interpretation is limited as the clinical practices that lead to center variation are unknown.
The predictor variables that were selected by both regression models and the CART models were very similar, although the exact order in which they were identified was different. Regression analysis and classification tree models in head-to-head comparisons have a comparable performance.10,11,12
Logistic regression is a standard statistical technique in medical literature and is useful in determining the magnitude of the association between each risk factor and the outcome.12
However, estimation of the likelihood of poor outcome for individuals using the logistic regression equation is not practical in the clinical setting. The CART model may be simpler to use for clinicians and ancillary staff, as use of the model involves following a decision tree which provides a qualitative answer (likely or not likely to develop death or BPD).
An in-depth look at the variables selected by CART and by regression analysis provides new data for evaluation in future clinical studies. Birth weight was the variable that was most associated with outcome. This is consistent with literature indicating that birth weight and gestational age perform better for predicting the combined outcome of BPD/death as compared to indices of respiratory failure.13
Male gender was identified as a risk factor for death/BPD, confirming the increased susceptibility of premature male neonates to poor outcomes is well known.14
A higher OI and mean FiO2
are consistent with an increased severity of respiratory failure. A higher peak OI has been associated with death from respiratory failure in preterm infants,13
and in older pediatric patients.16
An earlier age at randomization was also associated with worse outcome. It is likely that extremely premature infants with an increased severity of respiratory distress syndrome would have been randomized earlier, and hence, an earlier age of randomization may be consistent with an increased severity of respiratory failure, as are more surfactant doses. It is interesting to note that conventional ventilation (as compared to high frequency ventilation) was associated with a higher risk of death and/or BPD by regression analysis but not by CART analysis. The association of conventional ventilation with death and/or BPD was also identified in the post hoc
analyses of the main trial.1
This is consistent with meta-analyses demonstrating that although there is not a convincing benefit to elective high frequency ventilation,17,18
there may possibly be some benefit with high frequency ventilation used for “rescue” of infants with hypoxemic respiratory failure. 19,20
In summary, stepwise regression and CART models identified variables and the optimal cut-points of the variables that are associated with death and/or physiologic BPD in premature infants with severe respiratory failure. These models may prove useful in assessment of prognosis as well as in stratification of infants for future clinical trials. We also observed that the magnitude of initial response to iNO as used in the trial did not correlate with outcome. Future studies are required to identify short-term indicators of response to iNO, if any, that may be associated with long-term benefit and can be used to adjust iNO dosing and duration.