Abstract:
Sepsis remains a significant cause of neonatal mortality and morbidity, especially in low and middle-income countries. Neonatal sepsis presents with nonspecific signs and symptoms that necessitate tests to confirm the diagnosis. Early and accurate diagnosis of this infection can improve clinical outcomes and decrease overuse of antibiotics. Current diagnostic methods in Rwanda rely on rule-based methods of physicians and conventional culture methods, which are time-consuming and may delay critical therapeutic decisions. This research project focuses on designing a machine learning model based on Artificial Neural Networks (ANN) for the early diagnosis of neonatal sepsis using data collected from neonates admitted to Kabgayi and Ruhengeri Level II Teaching Hospitals in Rwanda in the year 2023. This intervention is paramount because of the urgent need to improve neonatal survival rates, particularly in Sub-Saharan Africa. The model utilizes ANN architecture built upon clinical data extracted from Electronic Medical Records (EMRs) in Neonatal Intensive Care Unit (NICU) to determine the probability of neonatal sepsis among neonates admitted in NICU during the year 2023. The population used for this study consisted of a relatively balanced dataset of 1381 neonates admitted in NICU, with (47%) negative cases and (53%) positive cases; 966 of them were used for training the model, which was then validated on 315 neonates and then the remaining 100 were used for testing. Preprocessing steps were employed to handle missing values and extract categorical features, while sepsis criteria were defined to identify neonates who were at risk. The model undergoes 20 training epochs using Adam optimization and incorporates early stopping to prevent overfitting. Evaluation on a test set comprising 100 samples reveal a test accuracy of 85%, with a precision of 85.18%, recall of 86.79%, and F1 score of 85.98%. The Area Under the Curve of Receiver Operating Characteristics (ROC-AUC) was 84.88%. Of the tested samples, the model predicted 54 cases to have sepsis and 46 cases not to have sepsis. The classification report indicates a balanced classification performance between the two classes, with a weighted average F1-score of 85%, sensitivity of 86.79% and specificity of 82.98%. This model has the potential to be easily implemented as a decision support system once incorporated into EMR system in different NICUs.
