Machine learning models for neurocognitive outcome prediction in preterm born infants

Area of Science:

• Neonatal Medicine
• Artificial Intelligence in Healthcare
• Developmental Pediatrics

Background:

• Neurocognitive outcome prediction in preterm infants is challenging.
• Existing methods often rely on complex predictors not suitable for routine clinical use.
• This study explored machine learning for improved prediction using accessible neonatal data.

Purpose of the Study:

• To investigate machine learning's potential to enhance neurocognitive outcome prediction in preterm infants.
• To compare machine learning models against conventional logistic regression.
• To utilize predictors readily available in neonatal settings.

Main Methods:

• Machine learning models (Random Forest, Support Vector Machine) were developed.
• Predictors from antenatal and neonatal periods were used for infants born <30 weeks gestation.
• Models were validated using internal cross-validation and compared to logistic regression.

Main Results:

• Random Forest and Support Vector Machine models achieved AUCs of 0.682 (2-year) and 0.695 (5-year).
• High negative predictive values (95% and 91%) were observed for identifying infants without adverse outcomes.
• Machine learning models significantly outperformed conventional logistic regression.

Conclusions:

• Machine learning offers a promising approach for early identification of preterm infants at low risk for adverse neurocognitive outcomes.
• While overall predictive performance was moderate, the high NPV is clinically valuable.
• Future research should explore integrating diverse routine clinical data, like vital sign time series, into predictive models.