Developing and validating an explainable digital mortality prediction tool for extremely preterm infants

Area of Science:

• Neonatal Medicine
• Medical Informatics
• Biostatistics

Background:

• Perinatal management of extremely preterm infants presents significant decision-making challenges.
• Mortality prediction tools are essential for supporting clinical decisions in neonatal care.
• Existing prediction models may not fully meet the needs for accuracy and utility in this high-risk population.

Purpose of the Study:

• To develop and internally validate an explainable online tool for predicting mortality in extremely preterm infants before neonatal discharge.
• To compare the performance of various machine learning approaches for mortality prediction in this cohort.
• To assess the tool's calibration, discrimination, and clinical utility compared to previously published models.

Main Methods:

• Utilized population-based electronic patient record data from 25,902 infants born between 23+0-27+6 weeks' gestation across 185 English and Welsh neonatal units (2010-2020).
• Developed and internally validated a mortality prediction tool using stepwise backward logistic regression, selected from nine machine learning algorithms.
• Externally validated the tool's performance in a multinational cohort of preterm infants.

Main Results:

• The developed tool demonstrated good discrimination (AUC 0.746) and calibration, with superior net benefit across probability thresholds (10%-70%).
• The tool outperformed previously published models in calibration and utility.
• Acceptable performance was confirmed in the external validation cohort, indicating generalizability.

Conclusions:

• An explainable, online mortality prediction tool for extremely preterm infants has been developed and validated.
• This tool shows promising utility for supporting high-risk perinatal decision-making.
• Further evaluation is recommended before widespread clinical adoption.