Background: Machine learning in the healthcare sector represents a group of technologies in all aspects of medicine, and it appears promising, especially in emergency medicine. Hence, this study aims to utilize emergency department (ED) records to train machine learning algorithms and assess medical performance and outcomes.
Methods: This is a retrospective observational cohort study utilizing emergency patient records acquired from the Emergency Department of King Faisal Specialist Hospital & Research Centre in Riyadh City. Also, different machine learning models were evaluated, including regression, instance-based, regularization, tree-based, Bayesian, dimensionality reduction, and ensemble algorithms.
Results: A total of 149,513 emergency patient records were acquired. Due to many outliers and mislabeled data, clinical knowledge and a confident learning algorithm were used to preprocess the dataset. This resulted in only 84,970 patient records being kept. We observed that ensemble algorithms outperformed the others in all evaluation metrics, achieving an F-1 score and quadratic weighted kappa of 93.1% and 0.8623, respectively, in the case of CatBoost. In addition, the model never classified an emergent patient as nonurgent, nor did it classify a nonurgent ED patient as emergent. Optimizing the healthcare center workforce while ensuring that all critical patients are treated immediately is vital.
Conclusion: Machine learning-based triage models are feasible, highly accurate, and provide an in-depth assessment of the patient’s risk profile, which may not be found in routinely used emergency triage systems. A prospective study to evaluate the potential efficacy of machine learning-based triage models in predicting emergency visit outcomes needs to be conducted.
Key words: Machine learning, artificial intelligence, emergency medicine, triage.
|
