Reducing inequalities using an unbiased machine learning approach to identify births with the highest risk of preventable neonatal deaths

TL;DR

This study uses machine learning to accurately identify high-risk births for preventable neonatal deaths in Brazil, helping policymakers target interventions without bias.

Contribution

The study introduces an unbiased machine learning approach to identify high-risk births for preventable neonatal deaths.

Findings

XGBoost identified the top 5% of high-risk births responsible for over 85% of preventable neonatal deaths.

Risk predictions showed no bias against disadvantaged populations based on race, education, marital status, or maternal age.

The approach can be adapted for use in other developing countries to reduce neonatal mortality and health inequalities.

Abstract

Despite contemporaneous declines in neonatal mortality, recent studies show the existence of left-behind populations that continue to have higher mortality rates than the national averages. Additionally, many of these deaths are from preventable causes. This reality creates the need for more precise methods to identify high-risk births, allowing policymakers to target them more effectively. This study fills this gap by developing unbiased machine-learning approaches to more accurately identify births with a high risk of neonatal deaths from preventable causes.

We link administrative databases from the Brazilian health ministry to obtain birth and death records in the country from 2015 to 2017. The final dataset comprises 8,797,968 births, of which 59,615 newborns died before reaching 28 days alive (neonatal deaths). These neonatal deaths are categorized into preventable deaths (42,290) and non-preventable deaths (17,325). Our analysis identifies the death risk of the former group, as they are amenable to policy interventions. We train six machine-learning algorithms, test their performance on unseen data, and evaluate them using a new policy-oriented metric. To avoid biased policy recommendations, we also investigate how our approach impacts disadvantaged populations.

XGBoost was the best-performing algorithm for our task, with the 5% of births identified as highest risk by the model accounting for over 85% of the observed deaths. Furthermore, the risk predictions exhibit no statistical differences in the proportion of actual preventable deaths from disadvantaged populations, defined by race, education, marital status, and maternal age. These results are similar for other threshold levels.

We show that, by using publicly available administrative data sets and ML methods, it is possible to identify the births with the highest risk of preventable deaths with a high degree of accuracy. This is useful for policymakers as they can target health interventions to those who need them the most and where they can be effective without producing bias against disadvantaged populations. Overall, our approach can guide policymakers in reducing neonatal mortality rates and their health inequalities. Finally, it can be adapted for use in other developing countries.

Figures