Infant death prediction using machine learning: A population-based retrospective study.

Area of Science:

• Perinatal Health
• Machine Learning in Healthcare
• Public Health Informatics

Background:

• Infant mortality rates in the U.S. remain a concern despite recent declines.
• National goals for reducing infant deaths have not yet been achieved.
• Predictive modeling offers a novel approach to understanding and mitigating infant mortality.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting infant death.
• To identify key predictors of infant mortality.
• To assess the performance of predictive models during pre-pandemic and pandemic periods.

Main Methods:

• A population-based retrospective study utilizing U.S. live birth data from 2016-2021.
• Thirty-three diverse factors were analyzed, encompassing birth facility, prenatal care, pregnancy history, labor, delivery, and newborn characteristics.
• XGBoost and four other machine learning models were compared for predictive accuracy.

Main Results:

• XGBoost outperformed other models, with an Area Under the Curve (AUC) of 0.98 for neonatal death prediction.
• Key predictors identified include gestational age, birth weight, 5-min APGAR score, and prenatal visits.
• A simplified four-predictor model achieved comparable performance (AUC: 0.91) to the full model (AUC: 0.93) and outperformed existing risk screening tools.

Conclusions:

• XGBoost-based models provide robust prognostic information for perinatal care.
• A simplified four-factor classification system is a practical tool for infant death risk prediction.
• These models can enhance perinatal education and counseling efforts.