Related Experiment Video
Updated: Jul 25, 2025
07:57
Modeling Neonatal Intraventricular Hemorrhage Through Intraventricular Injection of Hemoglobin
Published on: August 25, 2022
3.0K
Machine Learning Detects Intraventricular Haemorrhage in Extremely Preterm Infants.
Minoo Ashoori1,2, John M O'Toole1,3, Ken D O'Halloran1,2
1INFANT Research Centre, University College Cork, T12 AK54 Cork, Ireland.
Children (Basel, Switzerland)
|June 28, 2023
Summary
Machine learning analysis of prolonged relative desaturations in regional cerebral oxygen saturation (rcSO2) effectively detected brain injury in extremely preterm infants. This automated approach shows promise for predicting intraventricular hemorrhage (IVH).
Area of Science:
- Neonatal neurology
- Biomedical engineering
- Machine learning applications in healthcare
Background:
- Extremely preterm infants (<28 weeks' gestational age) are at high risk for brain injury, specifically intraventricular hemorrhage (IVH).
- Current monitoring methods for detecting brain injury in neonates have limitations.
- Regional cerebral oxygen saturation (rcSO2) and peripheral oxygen saturation (SpO2) are physiological parameters that can be monitored non-invasively.
Purpose of the Study:
- To evaluate the utility of machine learning algorithms applied to rcSO2 and SpO2 signals for detecting brain injury in extremely preterm infants.
- To compare the performance of a data-driven approach using prolonged relative desaturations (PRDs) against traditional threshold-based methods.
Main Methods:
- Analysis of a subset of 46 infants (<28 weeks' gestational age) from the Management of Hypotension in Preterm infants (HIP) trial.
- Continuous rcSO2 monitoring for the first 72 hours and cranial ultrasounds within the first week.
- Extraction of quantitative features from rcSO2 and SpO2 signals, focusing on data-driven prolonged relative desaturations (PRDs) and their predictive value for IVH (grade II-IV) using a machine learning model with leave-one-out cross-validation.
Main Results:
- The machine learning model using PRDs from rcSO2 achieved an area under the receiver operating characteristic curve (AUC) of 0.846, significantly outperforming a threshold-based rcSO2 approach (AUC 0.593).
- Neither clinical models nor SpO2-based models showed a significant association with brain injury.
- A significant association was found between data-driven PRDs in rcSO2 and the presence of brain injury.
Conclusions:
- Automated analysis of PRDs in cerebral NIRS signals demonstrates potential for improved prediction of IVH in extremely preterm infants compared to threshold-based methods.
- The findings suggest that machine learning applied to rcSO2 signals, specifically focusing on PRDs, can aid in early detection of brain injury.
- Further research is needed to refine the definition of PRDs and understand the underlying physiological mechanisms.
Keywords:
extreme gradient boosting (XGBoost)near-infrared spectroscopy (NIRS)peripheral oxygen saturation (SpO2)prolonged relative desaturation (PRD)regional cerebral oxygen saturation (rcSO2)
