This project developed a service and decision-support platform for neonatal death prevention measures, using data from a Cidacs cohort. It offers three services: visualization of cohort data through human comparative visual analysis; neonatal death risk prediction based on machine learning models; and simulation of public policy impacts on neonatal death risk.
A new dataset, called BRNeoDeath, was created using information from the Mortality Information System (SIM) and the Live Birth Information System (SINASC). This process involved data analysis and correction, as SIM and SINASC records often have missing variables. The final dataset includes 47,326,125 samples covering the period from 2006 to 2016. For data visualization, demographic information from the Brazilian Census, spanning 2000 to 2016, was also incorporated.
To classify neonatal mortality risk, a subset of BRNeoDeath records specifically related to the city of São Paulo was used. An exploratory analysis characterized this subset and observed the distribution of features across the classes (alive or deceased). This subset was then subjected to experiments involving various transformations, such as aggregation and dummy variable creation, to enhance the representation of original features. Evaluation protocols were established to test machine learning models, including cross-validation and different training-test split strategies. The most promising machine learning methods—such as Multivariate Regression, Gradient Boosting, XGBoost, and SVM—were selected to classify samples based on their mortality risk. The best result on this initial dataset achieved an AUC of 96% using all available features.
To ensure the developed model could be effectively applied in real-world scenarios, SHAP values were used, providing a clear understanding of each feature’s importance in the machine learning model. Additionally, hypotheses regarding the use of pre- or postnatal features to build data-driven models were explored, examining the potential complementarity of these approaches. This methodology was then applied to BRNeoDeath to analyze model behavior in more complex situations, as mortality risks were evaluated across highly heterogeneous regions.
The project led to the creation of two proprietary methodologies that significantly enhance the current standards in this field. The first methodology is an AI-based algorithm designed to classify the risk of death in newborns within their first few days of life, utilizing information from the Live Birth System. The second methodology is an algorithm that predicts neonatal mortality rates across various health regions in Brazil.
The project adopts an innovative approach by utilizing specialized tools such as SaaS architecture within a Cloud Computing environment, greatly enhancing the analysis capabilities of the study. This setup allows for visualization and manipulation of data from large population segments, specifically leveraging information from the ‘Cohort of 100 Million Brazilians,’ facilitating the development of monitoring indicators.
The first 28 days of life are the most vulnerable for a child’s survival. Mortality during this period is influenced by biological, socioeconomic, and healthcare factors. Nearly 46% of deaths worldwide occur in children under five, concentrated in the first week of life. Major causes of neonatal death include congenital malformations, birth asphyxia, and perinatal infections, most of which are preventable.
In Brazil, between 1999 and 2013, despite a general decline in child mortality, 70% of deaths were caused by preventable factors. The need for proper prenatal and delivery care increased by about 33%. Addressing neonatal mortality is essential, as it continues to rise, and the necessary interventions differ from those for other causes. The high infant mortality rates in Brazil highlight the need for decision-support tools to assess risks and implement targeted preventive strategies.
One of the most significant outcomes of the project was the development of an online platform (plataformasami.com.br) that provides easy access to all the knowledge generated throughout the initiative. This resource aims to support the development of public health policies focused on maternal and child health.
The platform features various interactive visualizations that correlate indicators related to maternal and infant care. Additionally, it includes AI models that can statistically estimate the likelihood of a newborn dying within the first 28 days of life. Finally, the platform offers a future prediction service, providing forecasts of neonatal mortality rates up to four months in advance, broken down by health regions.
The project team plans to develop a simulation system to evaluate how specific public policy actions can impact various issues, such as neonatal mortality. To do this, it is essential to create machine learning models that accurately represent the impact of specific policies, like the “Bolsa Família’ program, on various factors related to neonatal death. This effort will require developing separate models for each variable of interest.