Integrating Clinical Markers and Machine Learning for Accurate Prediction of Neonatal Sepsis

Abstract

Background: clinically suspected sepsis is a common occurrence in preterm and severely unwell newborns during their extended stay in the neonatal intensive care unit (NICU). This might be the first indicator of further negative outcomes. Consequently, our objective was to use data-driven learning techniques to forecast newborn in-hospital mortality using machine learning methods. The methodology included enrolling 1095 newborns who were admitted to a tertiary-level neonatal intensive care unit (NICU) in Taiwan from 2017 to 2020 for clinical suspicion of sepsis. Clinical symptoms, laboratory criteria, and the use of empiric antibiotics by doctors were used to define sepsis when it was clinically suspected. Patient demographics, clinical characteristics, laboratory data, and medicines were the variables used for analysis. Incorporating machine learning techniques, we employed DNN, k-nearest neighbours, support vector machine, random forest, and extreme gradient boost. We used the area under the receiver operating characteristic curve (AUC) to determine how well these models performed. Findings: 8.2% (90 newborns) died as a result of complications while in the hospital. The training set consisted of 765 patients (or 69.8% of the total) while the test set included 330 patients (or 30.2% of the total). Among the models that were evaluated for their ability to predict the outcome, DNN stood out with the highest area under the curve (0.923, 95% CI 0.953-0.893), the best accuracy (95.64%, 95% CI 96.76-94.52%), and the best values for Cohen's kappa (0.74, 95% CI 0.79-0.69) and Matthews correlation (0.75, 95% CI 0.80-0.70). Ventilator support needs upon suspicion of sepsis, feeding circumstances, and intravascular volume expansion were the three most important factors in the DNN significance matrix plot. Neither the training nor the test sets showed any significant difference in the model's performance. Clinicians may benefit from the machine learning algorithm's insights and improved advance communication with families after establishing the DNN model to forecast in-hospital mortality in newborns with clinically indicated sepsis.