Our large, multicenter study identified the change in the relative contributions of a variety of risk factors to the prediction of BPD with advancing postnatal age. This knowledge is important because the benefits of postnatal treatment strategies might depend on baseline BPD risk. For example, the benefits of postnatal steroids for the prevention of BPD seem to outweigh the risks only in populations of premature infants at high risk of BPD. A meta-regression analysis of randomized, controlled trials of BPD demonstrated that, as the BPD risk increased above 65% in the control groups, postnatal steroid use was associated with a decrease in the risk of mortality or cerebral palsy; at lower baseline risk, treatment was associated with increased risk of mortality or cerebral palsy (35
). In infants with more severe lung disease, corticosteroids seemed to reduce risk of adverse neurodevelopmental outcome (15
). In addition, steroids seem to be of greatest benefit if used after the first postnatal week (36
). Our models provide a precise and objective estimate of BPD risk at various postnatal ages, which is essential information for conducting trials to determine if preventive therapies, such as postnatal corticosteroids or inhaled nitric oxide, are beneficial in high-risk populations. These models might serve as a foundation for researchers to assess long-term respiratory risk including the development of asthma, long-term pulmonary function test status, and potential development of emphysema in adult life.
Previously identified risk factors for BPD include birth weight, gestational age, male sex, oxygen therapy at 24 hours, mechanical ventilation at 48 hours, and duration of assisted ventilation (16
). These factors, measured in greater detail and at different time points, were included in our model. We added the contribution of these factors relative to each other and their relative importance at different postnatal ages. For example, although oxygen exposure is a risk factor, its independent contribution to prediction of BPD risk is small relative to gestational age on Postnatal Day 1 and to respiratory support thereafter. Risk factors that were found to be significant in previous studies but that we did not include in our final models include PDA, NEC, sepsis, and postnatal corticosteroids (16
). None of these factors improved prediction of the risk of BPD after adjustment for the six critical risk factors.
Previous BPD prediction scoring systems have been described, but none has been widely adopted. Some have satisfactory sensitivity and specificity but use oxygen therapy at 28 postnatal days as a definition of BPD, which is now outdated (5
). Some include radiographs as part of the scoring system, which introduces subjectivity and reduces generalizability (8
). Other problems limiting the use of these models are the inclusion of ventilated infants only, the lack of categorization of BPD by severity, the exclusion of infants who die, and underuse of antenatal corticosteroids and surfactant therapy (8
). Most importantly, none examined models by postnatal day through the first 28 postnatal days. In addition, we internally and externally validated our models; our models were able successfully to classify infants in the internal validation sample into the correct level of BPD or death in more than 8 out of 10 cases. The models performed similarly in subjects in the SUPPORT trial who were born between 2004 and 2009. The C statistic is lower for both the subset of subjects in the reduced Benchmarking group and SUPPORT group because the excluded subjects had higher and lower gestational ages, and outcome is more accurately predicted at the extreme values (i.e., subjects with higher gestational age have better outcomes, with more accuracy; subjects with lower gestational age have worse outcomes, again with more accuracy).
We developed an on-line application of the prediction models. Caution should be exercised when using the on-line estimator to identify BPD risk in populations in which the demography or care practices differ markedly from our cohort. Our models and the on-line application were based on data from Level III neonatal intensive care units at large, mostly urban academic medical centers. Because our BPD definitions depended in part on various respiratory therapies, center-specific use of these therapies might influence the prediction models. For example, we speculate that high-frequency ventilation was most likely used in our study centers as a “rescue” therapy for infants with profound respiratory failure or pulmonary interstitial emphysema (40
), and therefore BPD risks were higher among infants exposed to high-frequency ventilation. In centers where high-frequency ventilation is used as a primary therapy, BPD risk might be overestimated using our model (41
In conclusion, we found that gestational age conveyed the most predictive information for BPD risk on Postnatal Days 1 and 3, and respiratory support on Days 7, 14, 21, and 28. The BPD estimator will provide prognostic information for families and clinicians, and will assist in identifying the optimal candidates for enrollment in clinical trials to prevent BPD.