452 - Predicting Daily Bronchopulmonary Dysplasia Risk with Artificial Intelligence and Deep Learning

Poster Number: 452A
Publication Number: 452A.431

Andrew Beam, Harvard School of Public Health - Boston, MA, Waltham, MA, United States; Kristyn Beam, Beth Israel Deaconess Medical Center, Waltham, MA, United States; Anil Palepu, Harvard Medical School, Cambridge, MA, United States; Richard B. Parad, Brigham and Women's Hospital/Harvard Medical School, Boston, MA, United States; Karna Murthy, Northwestern University, Feinberg School of Medicine, Chicago, IL, United States; Matthew Laughon, University of North Carolina at Chapel Hill School of Medicine, Chapel Hill, NC, United States; Kaashif A. Ahmad, MEDNAX, Houston, TX, United States; Reese H. Clark, MEDNAX, Myrtle Beach, SC, United States; Veeral Tolia, Texas A&M Health Science Center College of Medicine, Dallas, TX, United States

Background: Bronchopulmonary dysplasia remains the most common morbidity of premature infants.

Objective: Develop and validate an AI system based on deep learning to predict BPD that provides daily forecasts of BPD risk using routinely collected data.

Design/Methods: All data was from the Pediatrix Clinical Data Warehouse, which included 63,533 preterm infants born at < = 30 weeks gestation from 311 NICUs in the Mednax network from 2009 to 2018 (Table 1). We split the data into 3 random partitions: a training set of 41,237 from 248 NICUs for model development, an internal validation set of 10,301 from the same 248 NICUs in the training set, and an external validation set of 11,986 infants born at 63 NICUs not seen by the model during training. The model used 20 static variables collected at the time of birth (e.g. gestational age (GA), birthweight, maternal age) and 842 variables representing the infant’s longitudinal medical history (e.g. daily weights, laboratory values, procedures, medications). Variables for each infant were converted to vectors using a deep learning “transformer” approach and the model was trained to predict daily risk of BPD grade at 36 weeks GA or mortality by 28 days of life.

Results: Demographics of the cohort are reported in Table 1. On the internal validation dataset, the model had an AUC as measured on different days of life of 0.883 (Day 0), 0.910 (Day 3), 0.917 (Day 7), 0.929 (Day 14), 0.936 (Day 21), and 0.940 (Day 28). On the external validation dataset, the model had an AUC on different days of life of 0.890 (Day 0), 0.914 (Day 3), 0.919 (Day 7), 0.931 (Day 14), 0.938 (Day 21), and 0.94 (Day 28) (Figure 1). We compared the deep learning approach to two baseline models which used the static variables only: a logistic model achieved an AUC of 0.823 & 0.822 on the internal and external sets, and a random forest model achieved an AUC of 0.807 & 0.805 on the internal and external sets respectively. The model was also able to provide longitudinal risk trajectories for individual infants (Figure 2).Conclusion(s): In this work, we present the first use of AI based on deep learning to improve daily risk prediction of BPD or mortality using a large, multi-center dataset. Our model takes advantage of longitudinal medical data in addition to traditional risk factors to improve predictive accuracy and provide longitudinal risk trajectories. We plan to utilize this model to better understand the risk and trajectory of BPD natural history and help clinicians make better informed clinical decisions.
Table 1Demographic summary of infants in the dataset used to train and validate the model.
Figure 1Model AUC on the external validation data (n=11,986 from 63 NICUs) for predicting different subtypes of BPD and mortality on each day of life. The AUC for the overall model is shown in the pink/purple line.