Abstract
We developed a new two-stage machine learning (ML) model reflecting crucial bronchopulmonary dysplasia (BPD) indicators, specifically respiratory support duration (RSd). We identified significant clinical variables for the early prediction of BPD and its severity with high predictive accuracy. Our model, TS-PMbpd, serves as an effective adjunctive predictive tool for practical application in Neonatal Intensive Care Units (NICUs).
Introduction
Bronchopulmonary dysplasia (BPD) remains a significant challenge in neonatal care, primarily affecting extremely premature infants and leading to long-term respiratory and neurodevelopmental complications. The current clinical standard for diagnosing BPD severity—the NICHD 2001 criteria—requires waiting until an infant reaches 36 weeks of postmenstrual age. This delay often misses the critical window for early preventive interventions. While previous ML models have attempted early prediction, they frequently lack the ability to account for disease severity or the infant's respiratory history in a clinically intuitive manner.
Methodology: The Two-Stage Approach
To address these limitations, this study proposes a novel two-stage Multilayer Perceptron (MLP) model named TS-PMbpd. The architecture is designed to mimic clinical reasoning by incorporating intermediate milestones of lung health.
- Stage 1 (RSd Prediction): The model first predicts the duration of invasive respiratory support (RSd) using 45 prenatal and early postnatal variables collected within the first week of life.
- Stage 2 (BPD Prediction): This stage utilizes the initial 45 variables combined with the predicted RSd from Stage 1 to forecast both the occurrence of BPD and its specific severity level (None, Mild, Moderate, or Severe).
By "pre-calculating" the respiratory support duration, the model gains a more sophisticated understanding of the patient's trajectory. To ensure interpretability, the researchers employed SHAP (Shapley Additive Explanations) to identify and rank the clinical variables driving these predictions.
Data and Cohort Selection
The study leveraged data from the Korean Neonatal Network (KNN), a comprehensive nationwide registry covering approximately 75%–80% of very low birth weight (VLBW) infants in South Korea. The initial cohort consisted of 16,384 infants born between 2013 and 2020. After excluding cases with major congenital anomalies or missing data, a final analysis set of 11,177 infants was established.
Results and Clinical Insights
The TS-PMbpd model demonstrated superior performance compared to traditional algorithms such as SVM, Random Forest, and XGBoost. Key results include:
- BPD Occurrence: Achieved an AUROC of 0.897 with 81.9% accuracy.
- Severity Prediction: Maintained a respectable AUROC of 0.786 for multi-class classification across four severity levels.
- Predictive Variables: SHAP analysis identified Gestational Age (GA) and Birth Weight (BW) as the strongest predictors. Notably, early postnatal factors like hypotension and pulmonary hypertension within the first week were highlighted as critical warning signs for severe BPD.
Conclusion and Future Directions
The TS-PMbpd model offers a significant advancement in early BPD prediction by successfully integrating respiratory history into a multi-class classification framework. While the study is limited by its reliance on registry data (lacking real-time waveforms or biomarkers) and its focus on a South Korean population, it provides a robust foundation for future clinical tools. Next steps include validating the model on international cohorts and developing a simplified digital interface for bedside use in the NICU.
Summary updated