Submission ID 130924
| Issue/Objective | Globally, conditions necessitating emergency care contribute to 24-28 million deaths annually, accounting for 51% of total mortality and 42% of the global disease burden. The burden of emergency conditions is 4.4 times higher in low- and middle-income countries than in high-income countries. Ethiopia, for instance, has one of the highest mortality rates from emergency conditions, with 1,154 deaths per 100,000 people and 47,728 disability-adjusted life years (DALYs) per 100,000 people. The demand for emergency medical care has surged globally, particularly since the COVID-19 pandemic, outpacing the growth of hospital resources and resulting in severe overcrowding and boarding in emergency departments (EDs). ED crowding and boarding are a global healthcare crisis, significantly associated with increased mortality and morbidity, treatment delays, and reduced quality of care. In Ethiopia, national hospital guidelines recommend that patients be discharged within 24 hours of admission to the ED. However, in the study area, the average ED length of stay is three days, exacerbating the strain on emergency services and impacting patient outcomes. One of the most effective strategies to reduce ED overcrowding and boarding is triage, a system that classifies patients based on the urgency of their condition to prioritize care. While triage systems improve ED effectiveness, they remain vulnerable to mis-triage, which includes under-triage and over-triage. In the USA, mis-triage occurs in over 32% of ED encounters, with 10% of cases under-triaged and 90% over-triaged. In Ethiopia, 30.7% of patients are under-triaged, and 21.9% are over-triaged. These errors often stem from high workload and stress, inadequate training, non-standardized triage protocols, cognitive biases, reliance on initial impressions, limited triage tools, and resource constraints. This study, hence, aimed to develop and evaluate an explainable machine learning (XML) model designed to optimize ED by minimizing mis-triage and reducing boarding and crowding. We also identify the most important predictors at the population and triage levels, investigate the order, direction, and effects of these predictors across triage levels, and quantify the minimum information required to make a triage decision. |
|---|---|
| Methodology/Approach | An experimental research design was employed, and data were collected from the EDs of three hospitals in Ethiopia. Machine learning algorithms, namely Logistic Regression (LR), Gaussian Naive Bayes (GNB), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGB), and CatBoost, were tested, and SHAP (Shapley Additive exPlanations) was used to explain the model and conduct post hoc analysis. |
| Results | LR, GNB, DT, RF, XGBoost, and CatBoost achieved an F1-score of 85.29%, 84.21%, 92.55%, 92.38%, 94.18 %, and 94.45%, respectively. The CatBoost algorithm scored the highest performance with an F1-score of 94.45%, 20% higher than the previously reported triage model. The top five important features are modified early warning score (MEWS), mobility level, chief complaint, non-trauma, and systolic blood pressure. Results revealed that the order, direction, and strength of features' effects vary across triage levels, and context-specific and historical data are less relevant for making triage decisions. For example, MEWS is the most influential feature by a substantial margin, contributing significantly to predictions across all triage categories. Its consistent relevance highlights MEWS as a universal indicator of patient acuity, as it incorporates vital signs and the level of consciousness. Mobility level ranks as the second most important predictor, with a particularly strong influence in predicting red and green triage levels. This suggests that impaired mobility may be a strong signal of critical illness, while preserved mobility is more indicative of lower acuity presentations. The chief complaint is also highly influential, especially in predicting higher acuity levels, such as red and orange, underscoring the role of specific presenting symptoms in triage assessment. Non-trauma status contributes moderately across several categories but does not appear to be among the most critical features. Systolic blood pressure is more influential in predicting red triage cases, likely due to its role in identifying hemodynamic instability. Conversely, oxygen saturation (SPO2) has a greater impact on predicting green triage levels, reflecting its utility in confirming physiological stability in low-acuity patients. According to our analysis, triage nurses primarily need information on MEWS, chief complaint, non-trauma, mobility level, consciousness level, oxygen saturation, systolic blood pressure, heart rate, respiratory rate, temperature, age, and mode of arrival to make accurate triage decisions. Contextual (e.g., address, path to ED) and historical data (e.g., pre-hospital care, prior illness) were less predictive, highlighting the model's generalizability in the absence of complete data and practical utility across diverse settings. |
| Discussion/Conclusion | LR, GNB, DT, RF, XGBoost, and CatBoost achieved an F1-score of 85.29%, 84.21%, 92.55%, 92.38%, 94.18 %, and 94.45%, respectively. The CatBoost algorithm scored the highest performance with an F1-score of 94.45%, 20% higher than the previously reported triage model. The top five important features are modified early warning score (MEWS), mobility level, chief complaint, non-trauma, and systolic blood pressure. Results revealed that the order, direction, and strength of features' effects vary across triage levels, and context-specific and historical data are less relevant for making triage decisions. For example, MEWS is the most influential feature by a substantial margin, contributing significantly to predictions across all triage categories. Its consistent relevance highlights MEWS as a universal indicator of patient acuity, as it incorporates vital signs and the level of consciousness. Mobility level ranks as the second most important predictor, with a particularly strong influence in predicting red and green triage levels. This suggests that impaired mobility may be a strong signal of critical illness, while preserved mobility is more indicative of lower acuity presentations. The chief complaint is also highly influential, especially in predicting higher acuity levels, such as red and orange, underscoring the role of specific presenting symptoms in triage assessment. According to our analysis, triage nurses primarily need information on MEWS, chief complaint, non-trauma, mobility level, consciousness level, oxygen saturation, systolic blood pressure, heart rate, respiratory rate, temperature, age, and mode of arrival to make accurate triage decisions. Contextual (e.g., address, path to ED) and historical data (e.g., pre-hospital care, prior illness) are less predictive, highlighting the model's generalizability across diverse settings. We demonstrate that accurate, explainable, and efficient triage is feasible, offering a promising, scalable ED decision-support tool to minimize mis-triage and reduce ED boarding and crowding. This work also enables safer, equitable, and efficient care in resource-constrained settings. In the future, we plan to measure its impact in the real world. |
| Presenters and Affiliations | Berihun Tsega Alemayehu University of Gondar Abel Digafe Ketema Hawassa University Gebrie Mekonen Haile Addis Ababa University |
