Development of a machine learning-based model for early diagnosis of neonatal sepsis in Rwanda. Case study: Kabgayi and Ruhengeri level II teaching hospitals

Neonatal sepsis is a major cause of death and morbidity in developing countries, particularly in Sub-Saharan Africa. It is a life-threatening condition that can cause multiple organ failure, so early detection is critical for better outcomes and more effective antibiotic use. Current diagnostic methods in Rwanda rely on physician judgment and traditional cultural techniques, which take time and can lead to delays in critical treatment decisions. However, there is limited used of automated, data-driven tools to support early diagnosis, especially at district hospitals where neonatal specialists and advanced laboratory diagnostics are often lacking. The study aimed to develop a machine learning based-model that could be used in Rwandan hospitals for early detection of neonatal sepsis. Using a retrospective cohort study, clinical data from all neonates born and hospitalized in 2023 at Kabgayi and Ruhengeri level II teaching hospitals were extracted from electronic medical records. A machine learning model based on the Artificial Neural Networks (ANN) architecture was developed to predict the risk of neonatal sepsis using clinical data extracted from Electronic Medical Records (EMRs) in neonatal units. Pre-processing steps were used to handle missing values and extract categorical features, and sepsis criteria were defined to identify at-risk neonates. The model underwent 20 training epochs with Adam optimization and included early stopping to prevent overfitting. When tested on a set of 100 samples, the model yielded a test accuracy of 85%, with 85.18% precision, 86.79% recall, and an F1 score of 85.98%. The area under the curve for Receiver Operating Characteristics (ROC-AUC) was 84.88%. The model predicted 54 cases of sepsis among the tested samples and 46 cases of non-sepsis. The classification report indicated a balanced classification performance between the two classes, with a sensitivity of 86.79%, and a specificity of 82.98%. These findings demonstrated the potential of machine learning-based approaches in improving the early diagnosis of neonatal sepsis in developing countries. However, the model should be retested in other hospitals before being considered for integration into the EMR system of neonatal units across the country.