Machine learning for prediction of bronchopulmonary dysplasia-free survival among very preterm infants

Bronchopulmonary dysplasia (BPD) was first described by Northway et al. in 1967 as a new lung disease of preterm infants following respiratory distress syndrome [1]. Over the past 50 years since its first characterization, medical technology and clinical management have evolved. Accordingly, the pathology of BPD evolved from primarily necrotic bronchiolitis and fibrotic changes in the lung tissues to alveolar simplification in the post-surfactant era [1, 2]. BPD is associated with long-term cardiopulmonary complications as well as neurodevelopmental disadvantages, including cerebral palsy, vision and hearing deficits, mental and psychomotor impairments [3‐9].

Due to how its defined clinically, the diagnosis of BPD is rather subjective. Most contemporary practices follow the 2001 NICHD Workshop definition to diagnose BPD at 36 weeks postmenstrual age (PMA) [10]. Severity stratification was based on the duration of supplemental oxygen and respiratory support use. A revised operational definition was proposed after the 2016 Workshop [11]. In 2019, a systematic approach correlating 18 operational definitions of BPD with toddler-age respiratory and neurodevelopmental outcomes suggested that the best way to diagnose and grade BPD was by respiratory support mode at 36 weeks PMA. Notably, the extent of supplemental oxygen use was not needed [12].

To our knowledge, the concept of a developmental origin of BPD was first explicitly mentioned by Thébaud et al. recently in their extensive review article on BPD [13]. Indeed, gestational age (GA) is the best predictor of BPD [14]. Additionally, genetic factors such as race/ethnicity and sex also play a role [15‐18]. Moreover, modifiable perinatal factors such as maternal smoking, chorioamnionitis, placental insufficiency, as well as volutrauma, barotrauma, oxygen toxicity, and inflammation from prolonged mechanical ventilation use have been associated with BPD [19‐23]. The modifiable factors may reflect on the trajectory of respiratory support in the early postnatal period. As a result, distinct patterns of lung disease based on supplemental oxygen use have been characterized as early as the first two weeks of life [24, 25]. These early predictors led to the development of the BPD outcome estimator using 4 demographic and 2 respiratory factors which allowed risk stratification at six different days of life in the first 4 weeks of life [26, 27]. After external validation has been performed, the predictive power was found to be lacking, with the area under the receiver’s operating characteristic curves ranging only between 0.73–0.76 [26, 28, 29]. We and others also did not find it particularly useful clinically [30, 31]. Center effect likely plays an impactful role, and a cutoff probability level needs to be individualized for each unit [29, 30, 32].

In recent years, machine learning algorithms have become more accessible to clinical researchers. Machine learning algorithms are able to detect patterns in data that are invisible to the human eye, and it can be used as a more appropriate data analysis strategy for multifactorial pathologies such as BPD. In this project, we sought to investigate whether perinatal features and the trajectory of early-life respiratory support can be used to predict BPD-free survival using a machine learning algorithm.

Out of a total of 1,191 infants who met the gestational age criteria, 128 died before 28 days of life and were excluded. Among the remaining 1,063 infants, outcome data were available for 935 infants, complete data for perinatal features were available in 847 infants, and respiratory data were available in 689 infants. Perinatal and respiratory data from these infants were used in the study (Fig. 1). Data from randomly selected 552 (80%) infants were used for model training, and data from the remaining 137 (20%) infants were used for validation (Fig. 1). Demographic information and the number of infants with each respiratory support mode at various DOL are summarized in Tables 1 and 2, respectively.

The performances of the random forest and the ensemble models were listed in Table 3. Model 1 (perinatal features only) had a ROC AUC of 0.861 with the training dataset and 0.786 with the testing dataset. Using the probability cutoff threshold calculated for binary outcome prediction based on the training dataset (55%), we found positive and negative predictive values to be 0.773 and 0.802, respectively, upon model validation with the testing dataset. The ROC AUC increased with more respiratory data available for training, from 0.724 with only 1 day of data to 0.900 with all 14 days of data. The ensemble model combining perinatal features and all 14 days of respiratory data (Model 7) provided the best prediction with ROC AUC of 0.921 in the training dataset, and 0.899 in the testing dataset. Using a cutoff threshold of 48.8% for binary outcome prediction, the overall predicting accuracy was 85.1% and 81.0% in the training and the testing dataset, respectively. In the testing dataset, the positive predictive value was 0.855 and the negative predictive value was 0.773.

Using permutation to assess the relative importance of each feature, we found that gestational age was the most influential among all perinatal features, followed by birth weight z-score, male sex, maternal smoking, and race/ethnicity in the order of importance (Fig. 2A). For respiratory feature, CPAP/HFNC use was the most predictive of the outcome (Fig. 2B-D).

It is well established that lower gestational age at birth and longer intubation time are associated with higher risk of BPD. To confirm these facts in our model, Model 7 (perinatal features and all 14 days of respiratory data) was used to predict the probability of BPD-free survival (Scenario 1). As shown in Fig. 3, lower birth gestation and longer intubation time were individually associated with a reduced probability of BPD-free survival, providing assurance for the validity of the model.

The model continued to be utilized for further clinical simulations. In scenario 2, we found that extubating to NIMV did not lead to a better respiratory prognosis, noting that the probability of BPD-free survival remained less than 20% even after only one day of intubation following birth, and that no statistically significant difference was shown between extubation and no extubation. On the other hand, if an infant were able to maintain adequate respiration on CPAP after extubation, the probability of BPD-free survival was significantly higher, although the probability decreased with more time spent intubated (Fig. 4).

In Scenario 3, we observed that the longer an infant remained extubated before reintubation, the higher the probability of BPD-free survival (Fig. 5). After staying extubated for 9 complete days, there was no statistically significant difference in the probabilities of BPD-free survival between the reintubated infant and the non-reintubated (control) infant.

The literature suggests a developmental origin of BPD [13, 17, 18, 24, 26, 39‐41]. In this study, we expanded on this concept to develop a machine learning BPD-free survival prediction model. Five non-modifiable perinatal factors combined with respiratory support mode data from the first 14 days of life were used in this model. The five perinatal factors selected for inclusion are based on the elegant work of Morrow et al. where biological, social, and antenatal exposure determinants were assessed for their association with BPD. While sex and race/ethnicity are strictly non-modifiable, the impact of maternal smoking, fetal weight gain, and birth gestation are factors that may be modifiable from the obstetric standpoint. Together, these factors indicate that low lung tissue volume at birth plays a determinant role in the development of BPD. We did not include maternal conditions (chorioamnionitis, preeclampsia, prolonged preterm rupture of membrane, etc.) or maternal medications (antenatal steroid, magnesium sulfate, antibiotics, etc.) due to inconsistent reliability of the information. Moreover, responses to these exposures were expected to be reflected on the trajectory of respiratory support mode use, including the reversibility of their influences.

For respiratory data, we chose to only include respiratory support mode data instead of other respiratory parameters such as the fraction of inspired oxygen, oxygen saturation, mean airway pressure, blood gas results, etc., because the respiratory support mode is the most reliable variable guided by physician orders and is less likely to be influenced by clinical subjectivity. Additionally, documentation of respiratory mode change is less likely to be associated with errors. Moreover, respiratory parameters with normal ranges or goal ranges are not useful by themselves because there would not be a clear distinction between the two outcomes groups to guide prediction. On the other hand, the decision to choose a respiratory support mode over another is heavily influenced by local respiratory protocols. Consequently, a model developed with local data may not be as easily generalizable. It is important to note that the study was not designed to assess the superiority of one specific respiratory mode over another within the same group.

The most well-developed and widely used prediction tool for BPD is the NICHD Neonatal Bronchopulmonary Dysplasia Outcome Estimator [27]. The model was developed with data from 2,415 infants that underwent internal and external validation with additional 1,214 and 722 infants, respectively, during development [26]. The model was further validated by two additional study cohorts (PreVILIG and PREMILOC), both showing fair discrimination [28, 29]. We and others showed that the Estimator performed well in predicting death or severe BPD [30, 31]. Unfortunately, the Estimator does have limitations as it does not have an option for infants of Asian descent nor an option for NIMV. One model was developed for each DOL which does not take into consideration time as a continuous factor. In their assessment, demographic factors play a more crucial role during the early days of life, which they justified based on a sequential addition of variables and the related ROC AUC values. Notably, this method has trouble assessing the contribution of respiratory trajectories to BPD risks across days of life because different models were built for each DOL.

In our design, we separated perinatal factors and respiratory data into two models during initial training. We combined the two models into a final ensemble model to ensure both perinatal and respiratory features were taken into consideration with equal weight. By taking this approach, the final model does not add more weight on the respiratory data for prediction. From the clinical standpoint, this approach not only provided us with an opportunity to assess the developmental origin of bronchopulmonary dysplasia, but will also allow us to assess BPD risks at four time points: at birth, at the end of the first 24 h, as well as at 7 and 14 complete days of life. The most appropriate model for risk stratification can be picked for use in future studies comparing efficacy of interventional therapies to prevent BPD development depending on the timing of the intervention. Although there is not an easily accessible and suitable algorithm in classic machine learning which considers repeated measurement in a way that is reminiscent of mixed-modeling, we felt that using a decision tree-based algorithm on a dataset with each DOL as one feature allows the model to assess the interrelationship of respiratory support modes across all days of life for each infant. This serves as an alternative approach to assessing changes in respiratory support mode over time.

The lack of improvement in BPD-free survival in infants extubated to a NIMV mode was unexpected (Scenario 2). Notably, more than 80% of the infants receiving NIMV received unsynchronized NIPPV instead of non-invasive NAVA. A multi-center prospective trial comparing NIPPV and CPAP use in infants born < 1,000 g (about 2.2 lbs), and < 30 weeks’ gestation showed no difference in the risks of BPD-free survival [42]. A subgroup analysis showed no difference in BPD risks among infants with prior intubation. The use of NIPPV did not appear to provide benefit for preventing subsequent reintubation which occurred in about 60% of the infants in both groups. The severity of illness from morbidities or comorbidities of prematurity did not differ between the two groups. These findings were different from other trials which assessed the early use of NIPPV compared to CPAP in infants born at a later gestational age, which showed reduced intubation needs and a lower risk of BPD [43‐45]. A Cochrane review also found no difference in BPD risks between NIPPV and CPAP use [46]. In the input data for model training, we observed a higher percentage of infants without BPD who received CPAP/HFNC rather than NIMV (the majority received NIPPV) or an invasive ventilatory mode. Clinically, sicker infants were more likely to require longer NIMV support due to unreliable respiratory drives in order to avoid intubation/reintubation. Our model reflected that, which was different from the above-mentioned randomized controlled trial. Interestingly, a recent meta-analysis on mixed treatment comparisons found that surfactant administration followed by CPAP use provided superior respiratory outcomes to NIPPV use [47]. In clinical practice, it would be reasonable to trial CPAP with early use of pharmacological intervention to prevent apnea of prematurity on infants who are able to sustain a reliable respiratory drive, and reserve NIPPV to those infants who require artificial breath.

Prolonged intubation may be associated with barotrauma and volutrauma, while invasive ventilation may be necessary in poor pulmonary compliance or intolerance of non-invasive support. In the literature, extubation failure in preterm infants was inconsistently defined as reintubation between 48 h and 7 days [48‐51]. Reintubation may be associated with respiratory setback, leaving us with the question of whether the interval between first extubation and reintubation conveys any significance in influencing long-term respiratory outcomes. Our prediction model provided an objective way of assessing what constitutes as the best definition for “successful extubation” (Scenario 3). We explored this by comparing the probability of BPD-free survival to a control simulated infant who did not require reintubation after initial extubation. Based on our simulation using BPD-free survival probability as outcome correlation, the duration required to maintain extubated is at least 48 h longer than the definitions used in most studies.

One major drawback of our model is that the source of the data was from infants cared for by the same group of neonatologists, so the study was considered as a single-center study, making external application of the model difficult. Nonetheless, studies have already suggested a strong influence of center effect on BPD rate and predictability [29, 32]. Plus, the goal of a prediction model is different from that of a statistical model. A prediction tool may be considered useful if it provides a reasonable balance between bias and variance, and is generalizable to a target population (e.g., preterm infants receiving care in one NICU or by one group of neonatal providers). Additional limitation of our prediction model was binary prediction of BPD-free survival, rather than prediction of BPD grade or severity. Also, the definition of BPD remained operational and subjective, affecting the accuracy of outcome labeling.

In conclusion, we developed a prediction tool for BPD-free survival based on the developmental origin of BPD. The predictability of BPD using early-life predictors provides an opportunity for personalized early intervention targeting only the at-risk subpopulation to minimize potential harm to the low-risk group. Moreover, the ability to simulate respiratory support use provides clinicians an objective way of assessment for an optimal respiratory support mode to endow actual benefit to the infant. A web app created to demonstrate the use of the prediction tool developed in this study is accessible at https://​neostat.​shinyapps.​io/​Resp_​Mode_​ML/​.