Predicting mortality risk for preterm infants using random forest.

Area of Science:

• Neonatal Medicine
• Computational Biology
• Clinical Informatics

Background:

• Mortality remains a significant risk for extremely preterm infants.
• Current prediction models for neonatal mortality do not account for the evolving clinical course.
• There is a need for improved prediction tools that integrate dynamic physiological data.

Purpose of the Study:

• To develop and evaluate a novel machine learning model for predicting mortality in extremely preterm infants.
• To compare the performance of a random forest model against conventional clinical prediction scores.
• To assess the utility of incorporating dynamic physiological data into mortality risk assessment.

Main Methods:

• Retrospective analysis of 275 infants born <32 weeks gestation.
• Conventional mortality risk quantification using CRIB-II score and logistic regression.
• Development of a random forest (RF) model trained on clinical factors and "worry" labels derived from physiological data within 6 hours of death.

Main Results:

• The RF model demonstrated superior performance compared to conventional methods.
• RF model achieved 88% sensitivity and 0.93 AUC.
• Conventional models (CRIB-II, logistic regression) had sensitivities of 71% and 80% with AUCs of 0.78 and 0.84, respectively.

Conclusions:

• A random forest model integrating clinical factors and physiological data shows enhanced accuracy in predicting mortality for extremely preterm infants.
• This approach offers a significant improvement over existing clinical prediction models.
• Dynamic physiological monitoring combined with machine learning holds promise for optimizing neonatal intensive care and improving outcomes.