445 - Development of a Peripheral Blood Transcriptomic Signature for Bronchopulmonary Dysplasia

Poster Number: 445
Publication Number: 445.431

Maria M. Tovar-Hernandez, The University of Texas Health Science Center at San Antonio Joe R. and Teresa Lozano Long School of Medicine, San Antonio, TX, United States; Przemko Kwinta, Jagiellonian University, Kraków, Malopolskie, Poland; Axel Moreira, Baylor College of Medicine, Houston, TX, United States; Zoya Cheema, The University of Texas Health Science Center at San Antonio Joe R. and Teresa Lozano Long School of Medicine, San Antonio, TX, United States; Alvaro Moreira, University of Texas Health Science Center San Antonio, San Antonio, TX, United States

Background: Bronchopulmonary dysplasia (BPD) is a common sequela among very preterm neonates and is associated with lifelong morbidity. Identifying preterm newborns at high risk for developing BPD may impact medical decision making for clinicians and offer adequate patient selection for future clinical trials.

Objective: Develop an early transcriptional prediction model for BPD.

Design/Methods: Secondary analysis of whole blood microarray data obtained in the first week of life in 97 very low birth weight neonates. We randomly split the data into a 70:30 for model building and testing, respectively. Four machine learning models were used to reduce the dimension of transcripts. Predicting BPD (per the NICHD definition) was assessed for each model via C statistics. Performance of models was compared to two clinically important BPD risk factors: gestational age and birthweight.

Results: Overall, the median birthweight was 1000 grams (IQR 760,1245) with a median gestational age of 28.0 weeks (IQR 26.0, 30.0). Sixty-three percent of newborns were diagnosed with BPD at 28 days of life. From 33,352 transcripts, 5,065 were differentially expressed (FDR value < 0.01) between neonates with/without BPD. Each machine learning model reduced the number of discriminatory genes to < 5 and had a C statistic that ranged between 81.7% (95% CI, 65.9-97.5) to 89.4% (95% CI, 77.1-100). Comparatively the C statistic for gestational age and birthweight were 80.6% (95% CI, 64.4-96.7) and 87.2% (95% CI 73.5-100), respectively. A hybrid model including birthweight and two genes (SLC28A3 and PELO-ITGA1) yielded a C statistic of 93.9% (95% CI, 85.4-100). The top two gene ontology processes were transcription regulation and SRP-dependent co-translational protein targeting to membrane, while the top BioCarta pathway pointed towards T cell receptor signaling. Conclusion(s): Blood transcriptomics may play a promising role in the early prediction of BPD, thus enabling timely interventions.
Machine Learning Models for Predicting BPD in Test Cohort: Receiver-Operator Characteristic (ROC) Curves.The graph shows that MARS model has the lowest AUC=81.7% (95% Cl, 65.9-97.5), followed by SVM with AUC=84.4% (95% Cl, 69.4-99.5), Xgb Tree with AUC=86.7% (95% Cl, 73.5-99.9), and the Nnet has the highest AUC=89.4% (95% Cl, 77.1-100). Area Under the Curve (AUC), Multivariate Adaptative Regression Splines (MARS), Support Vector Machine (SVM), eXtreme, Gradient Boosting Tree (Xgb Tree), Neural Networks (Nnet).