On the usage of artificial intelligence for identifying main attributes and predicting neonatal sepsis.

During the neonatal period, newborns are more susceptible to developing conditions and diseases due to their fragility of the transition and adaptation to the extrauterine environment. Neonatal sepsis is one of the leading causes of morbidity and mortality in newborns, particularly among preterm and low birth weight infants, requiring early diagnosis to reduce complications and deaths. In this work, we evaluate the performance of artificial intelligence models in predicting neonatal sepsis and also identify the attributes that most contribute impact on models' learning and their relationship with the disease, using real data from the state of Pernambuco, Brazil. The six machine learning models evaluated were AdaBoost, CatBoost, Gradient Boosting, LightGBM, Random Forest and XGBoost. Performance metrics ranged from 0.7213 to 0.8548, with AdaBoost and LightGBM achieving the best results, reaching a sensitivity above 0.8197 and a specificity of 0.8397 in all three experiments. SHAPley Additive exPlanations (SHAP) analysis revealed strong relationships between sepsis and attributes such as intracranial hemorrhage, prematurity, CPAP use, TTN presence, and epicutaneous access, all of which were highly associated with sepsis cases. We conclude that the artificial intelligence models demonstrated promising results in predicting neonatal sepsis, highlighting critical clinical attributes associated with the disease and identifying the most relevant predictors.