Abstract
In South Africa, the publicly funded healthcare system which provides essential services at no cost to the majority of the population, is currently overburdened by the costs related to treating tuberculosis (TB), the human immunodeficiency virus (HIV), and the acquired immune deficiency syndrome (AIDS), as well as injuries, violence, and a spectrum of non-communicable diseases such as cancer, chronic respiratory
conditions, and diabetes. Additionally, the system is pressured by the high rates of maternal and child mortality, often attributed to malpractice. This strained healthcare infrastructure has had a negative effect on the functionality of hospital emergency care units (ECUs), further increasing the challenge of delivering urgent medical care.
This study investigates the current patient allocation processes in ECUs with a focus
on the Chris Hani Baragwanath Academic Hospital (CHBAH) located in Soweto, South Africa, and explores the potential of machine learning (ML) algorithms to address
identified deficiencies. The research highlights the existing triage system's challenges, such as misallocation and errors in the Triage Early Warning Score (TEWS)
calculations, using the South African Triage Scale (SATS). A total of 15 824 patient
records were collected from CHBAH to build ML models using the following algorithms:
Support Vector Machine (SVM), Random Forest (RF), Decision Tree (DT), AdaBoost-
SVM, AdaBoost-RF, AdaBoost-DT, and Stacking to correctly allocate patients to the relevant ECUs based on the patient’s records. By automating the decision-making process regarding patient allocation in ECUs, this study sought to mitigate the potential
of putting patients’ health at risk due to misallocation.
The designed ML models underwent rigorous testing, including cross-validation, and were assessed using two imputation methods and feature selection techniques, leading to the identification of RF and AdaBoost-RF as the most accurate classifiers, with over 99.53% accuracy, and exhibiting high precision (1.00), recall (1.00), Fmeasure (1.00), and ROC Area (1.00). Finally, the study confirms the suitability of these two models for patient allocation to the various ECUs ( medical, surgical, and trauma units) based on the above performance metrics. The findings suggest that ML can significantly enhance the allocation process in ECUs, thereby potentially
improving patient health outcomes and optimising resource utilisation.