Time-Series Machine Learning for Prediction of Bronchopulmonary Dysplasia

Area of Science:

• Neonatal Medicine
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Bronchopulmonary Dysplasia (BPD) is a significant complication in extremely low gestational age newborns (ELGANs).
• Accurate early prediction of BPD is crucial for timely intervention and improved outcomes.
• Existing prediction models may not fully capture the dynamic nature of infant physiology.

Purpose of the Study:

• To develop and evaluate a time-series machine learning (ML) model for predicting BPD in ELGANs.
• To compare the predictive performance of the ML model against established online calculators.
• To identify key clinical factors influencing BPD development and severity.

Main Methods:

• Utilized a cohort of 92 ELGANs (gestational age 25.6 ± 1.4 weeks).
• Extracted time-series data from electronic medical records, including static and dynamic attributes.
• Developed and compared dynamic Long Short-Term Memory (LSTM) models with static regression models and the Neonatal Research Network (NRN) BPD calculator.

Main Results:

• Dynamic LSTM models outperformed static models and the NRN calculator across multiple time points (PND 1-28).
• Peak Area Under the Curve (AUC) reached 0.8400 with LSTM models on Postnatal Day 28.
• Principal Component Analysis identified respiratory support, ventilator settings, oxygen, medications, and growth as key predictive factors.

Conclusions:

• LSTM-based ML time-series analysis offers a substantial improvement over static methods for BPD prediction in ELGANs.
• This approach shows promise for enhancing real-time BPD trajectory mapping in clinical settings.
• Integrating advanced ML into neonatal care can lead to more precise and personalized patient management.