Machine-learning-based evaluation of the usefulness of lactate for predicting neonatal mortality in preterm infants.

Background: Unlike in adult and pediatric patients, the usefulness of lactate in preterm infants has not been thoroughly discussed. This study aimed to evaluate whether the lactate level in the first hours of life is an important factor associated with neonatal death in very-low-birth-weight (VLBW) preterm infants.

Methods: Electronic medical records from a level 4 neonatal intensive care unit in South Korea were reviewed to obtain perinatal and neonatal outcomes. Data on lactate levels of preterm infants in the first 12 h of life were collected. Neonatal mortality and morbidities were compared based on lactate levels. Subsequently, machine-learning models incorporating 20 independent variables, both with and without lactate, were compared for model performances and feature importance of lactate for predicting in-hospital mortality in the applicable models.

Results: One hundred and sixty-eight preterm infants were included. Death rates on days 7 and 30 of life (D30-mortality) were significantly higher in infants with high lactate levels (≥3rd interquartile range) than in those with lower levels (<3rd interquartile range). Though statistically insignificant, the overall in-hospital mortality was more than twice as high in the high lactate level group than in the lower lactate level group. Based on the machine learning results, Random Forest, Gradient Boosting, and LightGBM models all showed greater area under the curves when lactate was included. Lactate consistently ranked in the variables of top five feature importance, particularly showing the greatest value in the Gradient Boosting model.

Conclusion: Lactate levels during the early hours of life may be an important factor associated with in-hospital death of preterm VLBW infants. Based on the enhanced performance of the above-mentioned machine learning models, lactate levels in the early postnatal period may add to assessing the clinical status and predicting the hospital course in this population.