Evaluation of artificial intelligence models for predicting low birth weight using Brazilian real data

Low birth weight (LBW) is a significant global health concern, affecting millions of newborns and linked to increased risks of neonatal mortality, delayed development, and chronic diseases in adulthood. This work focuses on a subset of LBW called very low birth weight (VBLW), which exacerbates these risks. Utilizing data from the Sistema de Informações sobre Nascidos Vivos (SINASC), from the state of Pernambuco, Brazil, this work applies data analytics techniques to address issues found in the data set and then train machine learning models to predict VBLW.

By integrating feature selection (FS) with expert healthcare insights, data balancing methods, and hyperparameter optimization, the research aims to enhance model performance. Additionally, data slicing is employed to identify and discuss (possible) model biases.

In the data slicing analyse, particularly with respect to racial disparities in training data, revealed brown race a predominant group with an f1-score over 68% to all models used; however, for other racial groups the f1-score reached 68% at maximum. The work also highlights the challenges of model training due to the lack of significant attributes, such as detailed historical health data of the pregnant women, which limits the learning capabilities of the models.

Overall, the findings underscore the importance of comprehensive data and balanced model training to better generalize and accurately predict health outcomes in newborns.