Background Accurate prediction of bronchopulmonary dysplasia (BPD) development would allow targeted early treatment. This study aims to develop a machine learning (ML) model incorporating respiratory and oxygenation timeseries to predict BPD development within 1 week after birth. Methods Data was collected retrospectively from a neonatal intensive care (2009–2015). Readily available clinical data and respiratory and oxygenation timeseries (mode of respiratory support, FiO2, SpO2) were gathered. Descriptive features were extracted from respiratory and oxygenation timeseries. Respiratory and oxygenation timeseries were compressed and trained on a long short-term memory neural network, combined with a neural network on clinical data. For comparison, logistic regression (LR), support vector machine and XGBoost models were trained on clinical and descriptive data. Results A total of 513 patients were included, of whom 102 (19.8%) developed BPD at 36 weeks postmenstrual age. Models based on clinical data and advanced timeseries features performed best (AUC 0.83, 95% CI 0.81–0.84). Performance was significantly better than the best performing models based on clinical data only (LR, AUC 0.80, 95% CI 0.79–0.82, p = 0.005) and clinical data supplemented with descriptive respiratory and oxygenation timeseries features (LR, AUC 0.81, 95% CI 0.79–0.82, p = 0.01). Conclusions Machine learning approaches that process respiratory and oxygenation timeseries can improve BPD prediction in preterm infants. Impact <UnorderedList Mark="Bullet"> <ItemContent> Existing BPD prediction models mainly use clinical data and do not fully utilize the predictive value of respiratory and oxygenation timeseries. </ItemContent> <ItemContent> This study shows that machine learning models using advanced features from respiratory and oxygenation timeseries significantly improve early prediction of BPD in preterm infants compared to models using only clinical or descriptive data. </ItemContent> <ItemContent> Incorporating prediction models which use advanced analysis of respiratory and oxygenation timeseries may enable earlier identification of infants at risk for BPD, supporting targeted and timely interventions in neonatal care </ItemContent> </UnorderedList>

Prediction of bronchopulmonary dysplasia seven days after birth using respiratory and oxygenation timeseries with machine learning

期刊论文
发表日期：

2026年7月17日

查看全文

Frank C. Bennis,
Wes Onland,
Joris P. van der Vorst,
Mark Hoogendoorn,
Gerard J. Hutten,
Anton H. van Kaam,
Jaap Oosterlaan,
Marsh Königs

摘要

Background

Accurate prediction of bronchopulmonary dysplasia (BPD) development would allow targeted early treatment. This study aims to develop a machine learning (ML) model incorporating respiratory and oxygenation timeseries to predict BPD development within 1 week after birth.

Methods

Data was collected retrospectively from a neonatal intensive care (2009–2015). Readily available clinical data and respiratory and oxygenation timeseries (mode of respiratory support, FiO₂, SpO₂) were gathered. Descriptive features were extracted from respiratory and oxygenation timeseries. Respiratory and oxygenation timeseries were compressed and trained on a long short-term memory neural network, combined with a neural network on clinical data. For comparison, logistic regression (LR), support vector machine and XGBoost models were trained on clinical and descriptive data.

Results

A total of 513 patients were included, of whom 102 (19.8%) developed BPD at 36 weeks postmenstrual age. Models based on clinical data and advanced timeseries features performed best (AUC 0.83, 95% CI 0.81–0.84). Performance was significantly better than the best performing models based on clinical data only (LR, AUC 0.80, 95% CI 0.79–0.82, p = 0.005) and clinical data supplemented with descriptive respiratory and oxygenation timeseries features (LR, AUC 0.81, 95% CI 0.79–0.82, p = 0.01).

Conclusions

Machine learning approaches that process respiratory and oxygenation timeseries can improve BPD prediction in preterm infants.

Impact

Existing BPD prediction models mainly use clinical data and do not fully utilize the predictive value of respiratory and oxygenation timeseries.

This study shows that machine learning models using advanced features from respiratory and oxygenation timeseries significantly improve early prediction of BPD in preterm infants compared to models using only clinical or descriptive data.

Incorporating prediction models which use advanced analysis of respiratory and oxygenation timeseries may enable earlier identification of infants at risk for BPD, supporting targeted and timely interventions in neonatal care