Analyzing respiratory patterns through machine learning models improved early identification of bronchopulmonary dysplasia (BPD) in preterm infants.
RT’s Three Key Takeaways:
- Enhanced Prediction Accuracy: Machine learning models that combine clinical data with advanced respiratory time series can predict bronchopulmonary dysplasia (BPD) more accurately than models using clinical data alone.
- Value of Continuous Data: The study indicates that analyzing how respiratory support and oxygenation change over time provides critical information that is often lost in basic measurement summaries.
- Earlier Clinical Intervention: Implementing these predictive tools could help healthcare providers identify vulnerable preterm infants within seven days of birth, allowing for more timely and targeted care.
Researchers have developed machine learning models that improve the prediction of bronchopulmonary dysplasia (BPD) within seven days of birth by analyzing respiratory and oxygenation patterns, according to a study published in Pediatric Research.
Accurately identifying preterm infants likely to develop BPD could help neonatal teams direct timely interventions toward those at greatest risk. Existing prediction models primarily rely on clinical characteristics and may not capture the full predictive value of continuously recorded data, researchers said in the study.
The retrospective study included 513 preterm infants treated in a neonatal intensive care unit (NICU) between 2009 and 2015. Of those infants, 102, or 19.8%, developed BPD at 36 weeks postmenstrual age.
The machine learning models combined routine clinical information with time series data collected during the first week after birth. This data included the mode of respiratory support, fraction of inspired oxygen, and peripheral oxygen saturation. Investigators evaluated both descriptive summaries of these measurements and advanced representations designed to capture changes and patterns over time.
The strongest prediction was achieved by models combining clinical data with advanced respiratory and oxygenation time series features. These models produced an area under the receiver operating characteristic curve (AUC) of 0.83. This performance was significantly better than the leading model based only on clinical information, which achieved an AUC of 0.80.
Adding simple descriptive respiratory and oxygenation features to clinical data produced a smaller improvement, resulting in an AUC of 0.81. This approach was also significantly outperformed by the model using advanced time series analysis.
The findings suggest that how respiratory support and oxygenation change over time contains clinically relevant information that may be lost when measurements are reduced to basic summaries, according to the researchers.
By processing these patterns, machine learning may strengthen early BPD risk assessment beyond models using clinical variables alone. Incorporating this type of analysis into neonatal prediction tools could support earlier identification of vulnerable preterm infants and more timely, targeted care in healthcare settings.
Further evaluation will be needed before these models can be incorporated into routine clinical practice.
Reference
Bennis FC et al. Prediction of bronchopulmonary dysplasia seven days after birth using respiratory and oxygenation time series with machine learning. Pediatr Res. 2026. doi:10.1038/s41390-026-05301-z. https://www.nature.com/articles/s41390-026-05301-z
This article was originally published by AMJ and was made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
