Reducing Inequalities Using an Unbiased Machine Learning Approach to Identify Births with the Highest Risk of...

Area of Science:

• Public Health
• Machine Learning
• Health Informatics

Background:

• Neonatal mortality remains high in certain populations despite overall declines.
• Many neonatal deaths are preventable, necessitating precise risk identification methods.
• Existing methods lack the accuracy to effectively target high-risk births for intervention.

Purpose of the Study:

• To develop unbiased machine learning (ML) approaches for accurately identifying births at high risk of preventable neonatal death.
• To provide policymakers with tools to effectively target health interventions and reduce neonatal mortality and health inequalities.

Main Methods:

• Utilized administrative birth and death records from Brazil (2015-2017) covering nearly 8.8 million births.
• Trained and evaluated six ML algorithms to predict preventable neonatal deaths.
• Developed a novel policy-oriented metric and assessed model performance for fairness across disadvantaged populations.

Main Results:

• XGBoost demonstrated superior performance, with the top 5% of predicted high-risk births accounting for over 85% of preventable neonatal deaths.
• Risk predictions showed no statistical bias against disadvantaged populations based on race, education, marital status, or maternal age.
• These findings held true across different risk threshold levels.

Conclusions:

• Publicly available data and ML methods can accurately identify births at high risk of preventable mortality.
• The developed approach enables targeted, effective, and unbiased health interventions.
• This strategy can guide policymakers in reducing neonatal mortality and health disparities, with potential applicability in other developing countries.