In this study, we compared the ETAT guidelines commonly used in Kenyan public hospitals and the newly developed Smart Triage model based on their ability to classify children presenting to the ED into emergency, priority, and non-urgent categories. The major finding based on the selected model thresholds, was a shift in participants classification from non-urgent to priority to emergency when moving from ETAT to the ST-only model. The magnitude of this shift was larger when independent triggers were added. The emergency category consistently had the highest proportion of children who were admitted or died across all triage systems.

The purpose of paediatric triage is to distinguish children who require urgent intervention by the clinical team from those who can safely wait in the queue when they present to the emergency department. As expected in any triage system, the number of emergency cases should be fewer than priority and non-urgent cases. This was the case in our study, where the proportion of children who were classified as emergency was lower than those assigned priority and non-urgent categories by all triage systems. Assignment of fewer patients in the high acuity category is particularly important to minimize pressure on the already overwhelmed ED. Additionally, the emergency classification expedites clinical review and intervention for children in need of urgent care, thereby improving patient outcomes and ensuring the appropriate use of meager resources.

The Smart Triage models caused a shift in the distribution of children across the three triage categories, with more children being assigned priority categories by the ST-only model and the ST model with independent triggers. Additionally, all ST models classified more children into the emergency category compared to ETAT. This was intentionally done with the selection of risk thresholds used to classify patients into these categories, to maximize the models’ sensitivity and specificity, and to reduce misclassification during triage (19). Furthermore, as the Smart Triage algorithm is a continuous probability, the proportion of cases in the emergency and priority groups determined by the algorithm can be adjusted to meet the local context and resources.

Since there is no absolute measure of acuity, admission and mortality have typically been employed in research to assess how effectively a triage system performs (24). Similarly, we used these outcomes as proxies for illness severity. An ideal triage system should categorize children in need of admission or at a high risk of mortality in the high-acuity category. Our study findings showed that ETAT and Smart Triage models accurately identified critically ill children, with most admissions and deaths assigned to the emergency category. In addition, incorporation of independent triggers improved the Smart Triage model’s ability to classify mortality cases. The ST model with independent triggers and the recalibrated ST model with independent triggers assigned all mortality cases to either emergency or priority category compared to ETAT and ST-only model which had some mortality cases assigned to the non-urgent category.

Smart Triage offers a major advantage over ETAT in that it can robustly combine a range of continuous and categorical variables in a prediction algorithm that would not be possible for even the most highly trained clinician. The algorithm can also be tuned to optimize the trade-offs clinicians perform in clinical practice. In the current implementation, a high specificity was selected to avoid missing critically ill children, especially those in the early stages of critical illness that have not reached the criteria for priority or emergency status using ETAT. This is reflected in the increased number of children in the priority and emergency categories. However, this must be balanced by including too many children in the emergency categories and slowing down access to care for critically ill children with obvious danger signs.

The incorporation of independent emergency and priority triggers into the Smart Triage models improved the categorization of admissions and mortality cases, as evidenced by the higher agreement between the Smart Triage models with independent triggers and the ETAT classification. Most of these triggers are used in ETAT to classify patients into emergency and priority categories and can explain the reason for the improvement in the categorization of admission and mortality cases. Therefore, this suggests that the inclusion of independent triggers as an additional criterion in paediatric triage systems that rely on prediction models is essential to avoid missing critically ill children. The prediction algorithm is not necessary in those children with a single symptom or sign that indicates a critically ill child. A similar argument has been proposed when comparing early warning scores to a single extreme observation (25).