Predictive Modeling for Perinatal Mortality in Resource-Limited Settings.

Area of Science:

• Global Health
• Perinatal Epidemiology
• Machine Learning in Medicine

Background:

• Fetal and neonatal deaths disproportionately affect low- and middle-income countries.
• Risk assessment tools are crucial for predicting adverse perinatal outcomes.
• Developing accurate prediction models is essential for targeted interventions.

Purpose of the Study:

• To develop and validate risk prediction models for intrapartum stillbirth and neonatal death.
• To evaluate the predictive accuracy of models using different sets of risk factors.
• To identify key predictors for intrapartum stillbirth and neonatal mortality.

Main Methods:

• A large prospective cohort study involving 502,648 pregnancies across South Asia, Africa, and Latin America.
• Sequential addition of risk factors in four scenarios: prenatal, predelivery, delivery/day 1, and postdelivery/day 2.
• Application of conventional and advanced machine learning techniques with data split into training, testing, and validation sets.

Main Results:

• Models using prenatal or predelivery data showed limited predictive accuracy (AUC ≤ 0.71) for both outcomes.
• Models incorporating delivery and postdelivery data demonstrated improved predictive accuracy for neonatal mortality (AUC > 0.80).
• Birth weight emerged as the most significant predictor for neonatal death in postdelivery scenarios, with AUCs of 0.78 and 0.76.

Conclusions:

• Risk prediction models incorporating delivery and postdelivery data are more effective for neonatal mortality.
• Birth weight is a critical factor in predicting neonatal mortality.
• Further refinement of these models can aid in reducing fetal and neonatal deaths globally.