Developing a Machine Learning-based Prediction Tool to Extend the Usability of the NICHD BPD Outcome Estimator to the Asian Population - A Pilot Study

Background: Bronchopulmonary dysplasia (BPD) is a complication of extreme prematurity and is associated with poor respiratory and neurodevelopmental outcomes. Despite significant improvement in the survival of extremely preterm infants, the BPD rate has not seen an improvement. The NICHD BPD Outcome Estimator is a tool widely used to guide corticosteroid use, which has been shown to reduce BPD severity but may also worsen neurodevelopmental outcomes. A limitation of the NICHD BPD Outcome Estimator is that it does not apply to infants of Asian descent. We sought for a physiologically plausible approach to extend the usability of the Estimator to Asian extremely premature infants. We hypothesize that susceptibility to respiratory prognoses in extremely premature infants of various racial/ethnic groups are interrelated. In this study, we aimed to develop an outcome prediction tool for extremely preterm infants of Asian descent based on prediction results for the other three major racial/ethnic groups, namely white, black, and Hispanic, of the same demographic and respiratory parameters using the NICHD BPD Outcome Estimator.

Methods: We conducted a retrospective study with prospective data collection following institutional review board approval. Patients born at ≤ 30 weeks gestational age (GA) with a birth weight (BW) of <1,250 grams and a race/ethnicity designation as Asian in the admission facesheet were included. Demographic and respiratory data including GA, BW, sex, ventilator type, and FiO2 on day of life 1, 3, 7, 14, 21, and 28 were entered into the Estimator. For race/ethnicity, white, black, and Hispanic were used separately to calculate risk probability for each severity category (death, severe, moderate, mild, no BPD). The severity risk-stratified probability results for each race/ethnicity as features and respiratory support needs (refer to Table 2 for definition) at 36, 37, and 40 weeks corrected GA (CGA) as the outcome were used in random-forest modeling. Ten-fold cross-validation repeated 10 times was performed for bias-variance balance check. Contingency tables and receivers’ operating characteristic curve area under the curve (ROC-AUC) were used for model performance assessment.

Results: We included 168 respiratory sets of data points from 29 Asian infants in this study, of which 114 sets of data points were used for training, and the remainder for testing. Table 1 shows patient characteristics. Table 2 summarizes model performance metrics. Notably, the final model showed ROC-AUC of 0.876, 0.828, and 0.733 for 36, 37, and 40 weeks CGA, respectively, in the testing data set. A web app for model demonstration is accessible at https://neostat.shinyapps.io/BPD_Asian/.

Conclusion: This pilot study successfully developed a random forest model with good performance to extend the usability of the NICH BPD Outcome Estimator to the Asian population. The study also provided insight into the interrelationship of respiratory outcomes among extremely preterm infants of various racial/ethnic groups.