Early Prediction of Mechanical Ventilation Needs in Very Preterm Neonates Using Machine Learning.

BACKGROUND

The use of invasive mechanical ventilation (IMV) increases morbidity in patients but may be required for survival on a subset of patients. Delayed, as well as unnecessary, initiation of IMV increases morbidity and mortality among preterm neonates.

OBJECTIVE

To assess the use of machine learning in predicting need for IMV among patients on lower levels of respiratory support (CPAP failure) using the earliest clinical data captured in the electronic health records across multiple health systems in the United States. Accurate predictions may enable earlier proactive intervention to minimize CPAP failure and improve clinical outcomes.

METHODS

This study was conducted using the Oracle EHR Real-World Data (OERWD) database including preterm neonatal intensive care unit (NICU) admissions between 2012 and 2022. Demographics and the first set of vital signs and laboratory values were retrieved and used to train an extreme gradient boosting (XGBoost) machine learning model.

RESULTS

Twenty thousand three hundred and sixty-three neonates from 27 NICUs qualified for the study with CPAP failure rate of 69.0%. Fraction of inspired oxygen (FiO) was the strongest predictor of CPAP failure followed by systolic blood pressure, partial pressure of oxygen (PaO), birthweight, diastolic blood pressure, gestational age, and oxygen saturation. The resulting model attained an area under the receiver operator characteristic curve of 0.90 (95% CI: 0.89, 0.91) and an F score of 0.88.

CONCLUSIONS

Patients requiring IMV can be predicted early at birth with high accuracy. This may result in earlier initiation of IMV on patients who need it while reducing unnecessary intubations.