https://orcid.org/0000-0002-8420-9156
Figures
Abstract
Studies comparing emergency department (ED) patient prioritization rules often use single averages, which can hide important clinical trade-offs. This paper presents and demonstrates a three-part evaluation framework designed for clear, multi-faceted comparisons of prioritization policies. The framework includes: (1) statistics that account for extreme outcomes, (2) profiles showing how well time targets are met, and (3) analysis based on stakeholder priorities. We illustrate the framework in a unified discrete-event simulation of a 30-bed mixed-acuity ED to show how conclusions can change across tails, thresholds, and stakeholder preferences; the numerical results are for illustration only and are not recommendations for any specific hospital. Our main contribution is the method itself: a consistent and repeatable way to reveal different but complementary information, helping decision-makers match policies to their local goals, limits, and risk tolerance. Before implementation, future work should apply this framework using data from specific hospitals and gathering input from their stakeholders.
Citation: DeHollander A, Karwan M, Casucci S (2025) Transparent comparisons of Emergency-Department prioritization policies: integrating tail risk, target attainment, and utility analysis. PLoS One 20(12): e0326722. https://doi.org/10.1371/journal.pone.0326722
Editor: Inge Roggen, Universitair Kinderziekenhuis Koningin Fabiola: Hopital Universitaire des Enfants Reine Fabiola, BELGIUM
Received: June 4, 2025; Accepted: December 10, 2025; Published: December 31, 2025
Copyright: © 2025 DeHollander et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: We made the data and code publicly available on the following GitHub repository: https://github.com/adamdeho/ER-Evaluation-Techniques.git.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction and literature review
1.1 Background and significance
Emergency department (ED) crowding is a persistent threat to timely, high-quality care. As of 2016, over 90% of EDs reported regular crowding, and the COVID-19 pandemic has worsened the situation [1]. Prolonged length of stay (LOS), boarding, and repeated “left-without-being-seen” (LWBS) events have each been linked to excess mortality, lower patient satisfaction, and staff burnout [2,3]. Long wait times not only delay time-sensitive interventions for high-acuity patients but also deteriorate the overall quality of care [4–7]. Even low-acuity patients may experience prolonged discomfort or clinical deterioration if neglected. Beyond clinical outcomes, crowding constrains a hospital’s ability to respond to new emergencies and impairs operational efficiency, with downstream effects on financial performance [8]. While root causes span hospital-wide capacity constraints, one tool ED managers can control is the patient prioritization strategy—that is, the rule that determines which patient is served next once a resource becomes available. Over the past two decades, researchers have proposed a rich catalog of such rules, ranging from simple first-come-first-served (FCFS) queues to dynamic algorithms that blend acuity, projected workload, and downstream bed availability [9].
Despite this methodological progress, evidence on which strategy works best is inconclusive. Primary studies differ widely in (i) the key-performance indicators (KPIs) they report (e.g., average LOS versus 90th-percentile LOS), (ii) whether they examine distribution tails, and (iii) the degree to which stakeholder preferences are incorporated. Since evaluation methods are inconsistent, the same strategy can look good in one study but bad in another. This makes it hard to apply findings broadly and adopt new methods [10]. The absence of a systematic evaluation framework therefore represents both a scientific gap and a practical barrier to evidence-based ED operations.
This paper addresses that gap by introducing a unified framework built on three complementary evaluation techniques. The first technique involves computing KPI summary statistics with explicit tail analysis to detect hidden extremes in performance distributions. The second constructs threshold-based performance curves that reveal sensitivity to time-target selection, making it easier to interpret operational trade-offs. The third incorporates stakeholder-informed utility functions that translate multidimensional clinical objectives into a single, interpretable scalar score.
Using nine prioritization rules—including widely studied approaches (e.g., FCFS, Accumulating Priority Queue [APQ]) and several novel strategies—we demonstrate that each rule reveals distinct trade-offs that are obscured by average-only performance metrics. We show that a strategy may appear optimal under one evaluation criterion but perform poorly under another. Our aim is not to identify a single “best” rule, but to equip ED decision-makers with a transparent framework for comparing strategies within their specific operational contexts.
1.2 Literature review
1.2.1 Patient-flow challenges in crowded EDs.
Systematic reviews consistently find that crowding worsens clinical outcomes and elevates operational costs [3,4,8]. Among the most frequently reported indicators are median LOS and time-to-physician [3,10]. However, these studies also emphasize that serious incidents like ambulance diversions are more often caused by the severe operational strains measured by extreme tail events (e.g., 99th-percentile LOS) than by the typical performance reflected in averages.
1.2.2 Patient-prioritization strategies.
A wide array of interventions have been proposed to alleviate ED crowding, including the use of telehealth solutions [4], educational initiatives for patients, and process-improvement frameworks like Six Sigma [11]. This review, however, centers on patient prioritization strategies—rules that dictate the sequencing of patients when a treatment resource becomes available. The most basic of these is the first-come, first-served approach, while more advanced methods utilize algorithmic or heuristic logic to improve performance metrics. As shown in Table 1, prior researchers have explored a diverse spectrum of prioritization methods, spanning from static mechanisms such as structured priority queues [12–14] to adaptive systems that make real-time decisions [15–17]. Some authors have pursued optimization-based formulations [18,19], although these often struggle with the incorporation of uncertainty. Other researchers have turned to data-driven techniques like machine learning to better model and respond to stochastic dynamics [20,21]. Additional investigations have examined more traditional strategies, including revised triage procedures [9,22] and evaluations of heuristic decision-making by frontline clinicians [23,24]. Comparative studies often use discrete-event simulation because it allows them to test ‘what-if’ scenarios for each patient without affecting actual hospital operations [25].
1.2.3 Existing evaluation practices.
Although numerous studies investigate patient prioritization strategies, no standard approach exists for evaluating their effectiveness. Instead, three methodological traditions have emerged in the literature.
The first and most common approach involves reporting single-moment KPIs, such as mean or median LOS, wait time, or throughput [25–28]. This practice is widespread in both simulation and empirical studies. However, because LOS distributions usually have a long ‘tail’ of very long stays, focusing only on the average can hide rare but severe delays.
A second approach assesses target-achievement rates, often in alignment with regulatory benchmarks—such as the proportion of patients discharged within a 4-hour window [29]. While such metrics are straightforward to interpret and align with policy goals, they depend heavily on the chosen threshold, which may be arbitrary and insensitive to broader performance variation.
The third approach, found in a smaller body of literature, employs utility-based multicriteria scoring to synthesize performance across several KPIs. These studies use explicit utility functions—linear, exponential, or Chebyshev—to represent stakeholder preferences [16,30]. Although this method enhances transparency, most implementations do not test the robustness of results to changes in utility-parameter values, limiting their prescriptive reliability (Table 2).
1.2.4 Lack of generalizability.
Early evidence suggested that simply introducing a structured triage scale would shorten waits and lower mortality [31]. Yet subsequent observational work uncovered the opposite effect: Sax et al. noted that widespread assignment to mid-acuity patients (i.e., ESI Level 3) created a “mid-acuity log-jam,” lengthening throughput for all but the sickest patients because beds were occupied by patients whose severity had been overestimated [32]. These conflicting findings highlight that the same triage rule can either alleviate or exacerbate crowding.
Commentaries have questioned whether current validation methods transfer across jurisdictions. Twomey et al. argued that techniques developed in well-resourced settings “may not be appropriate and repeatable in developing countries,” and highlighted conceptual problems in declaring any single metric the gold standard for validity [33]. One hospital might benchmark triage accuracy against ICU admission; another might use expert consensus. These differences make it difficult to compare hospitals and generalize findings.
The same pattern emerges in the rapidly growing AI/ML triage literature. El Arab and Al Moosa found that most machine-learning studies were single-center and lacked external validation, with selection bias and overfitting as recurrent threats [34]. When Ryu et al. trained a gradient-boosted triage score at one hospital and deployed it at sister sites, the AUC for predicting admission ranged from 0.93 to 0.71 across locations served by the same health system [35]. Broader reviews of data-driven admission predictors echo the call for “rigorous external evaluation before clinical use” [36]. Meanwhile, Ingielewicz et al. surveyed traditional scales and concluded that “no existing triage system clearly outperforms others in every aspect,” effectively dispelling the notion of a universal best-in-class tool [37]. From a resource-constrained perspective, Siddiqui et al. stated that “the need and practical applicability of any triage is dictated by the hospital system and setting” [38].
Adding to these problems, most studies only compare a new rule to the current one, instead of to other advanced rules, and they often use just one metric [20,22]. As we later demonstrate, the ranking of nine common prioritization strategies changes when analysts shift from average LOS to tail-sensitive or utility-based criteria within the same 30-bed ED. If evaluation choice alone can flip conclusions in one configuration, then extrapolating results across hospitals with different capacity, acuity mix, or stakeholder priorities is doubly precarious.
1.2.5 Gaps in the literature.
To our knowledge, no consensus exists on which KPIs constitute the minimal reporting set when analyzing a patient prioritization strategy. Recent umbrella reviews explicitly call for “standardized, multidimensional evaluation frameworks” to enable meta-analysis and real-world translation [10].
1.3 Contributions and organization
We synthesize and demonstrate a tripartite evaluation framework that (i) requires analysts to quantify tail behavior; (ii) exposes threshold dependence; and (iii) embeds explicit stakeholder utilities. The main contribution is the framework itself, not the specific numerical results. We illustrate its use in one unified DES setting to show how evaluation choice alone can reverse apparent rankings across strategies and cohorts. We do not advocate any specific rule in this paper; real-world selection requires site-specific validation and stakeholder preference elicitation, which we identify as future work.
The remainder of the study is organized as follows. Section 2 details the proposed methodology, including cohort‐specific metrics, threshold‐attainment curves, and utility analyses; Section 3 provides an illustrative application; Section 4 discusses managerial implications, limitations, and avenues for future research; and Section 5 concludes by summarizing the key contributions, offering practical recommendations for aligning prioritization rules with clinical objectives, and highlighting the study’s broader significance. Supplementary details appear in the Appendices in S1 File: Appendix A describes each prioritization strategy; Appendix B defines the stakeholder utility functions; Appendix C provides full simulation‐model parameters; and Appendix D presents the context-specific formulation of the area under the curve (AUC) that is used in our threshold‐attainment analysis in Section 2.3.
2. Materials and methods
2.1 Study design and objectives
The general objective of this study is to develop and demonstrate a multi-faceted framework for the comparative evaluation of ED patient-prioritization strategies. The framework is designed to address the inconsistencies highlighted in Section 1.2.4 by providing a transparent and standardized approach to assessment.
Within this overarching aim, the specific objectives are fourfold. First, we implement a discrete-event simulation (DES) of patient flow calibrated to a 30-bed, mixed-acuity ED to provide a model to illustrate the use of our evaluation framework. Second, we apply this model to nine distinct prioritization policies, spanning established rules, novel strategies, and composite approaches. Third, we evaluate each policy using three complementary techniques: distributional tail-risk statistics, threshold-attainment profiles, and stakeholder-informed utility analysis. Finally, we illustrate how the apparent preference among strategies shifts across these lenses, thereby motivating the need for standardized, multi-criteria reporting.
The numerical results presented are illustrative only; their purpose is to demonstrate the mechanics and insights of the framework rather than to advocate any particular prioritization strategy. To maintain clarity, we restrict attention to LOS—a KPI that is both widely reported in the literature and directly relevant to clinicians, administrators, and patients. The framework itself can be used with any key performance indicator and can be replicated with other indicators such as door-to-doctor time (DTDT), LWBS rates, or mortality. To illustrate this generality, Appendix E provides a brief DTDT example, showing that the same evaluation techniques apply seamlessly to other ED metrics. Expanding the analysis to multiple KPIs would significantly lengthen the manuscript without adding to its methodological contribution.
For similar reasons, we exclude the highest-acuity (i.e., ESI-1) arrivals from subsequent analyses, since these patients always receive immediate treatment regardless of prioritization rules. Throughout the paper, the term ESI refers to the Emergency Severity Index, a five-level triage system used to categorize patients in the emergency department based on acuity and anticipated resource needs. This scale ranges from level 1 for patients requiring immediate life-saving intervention to level 5 for stable patients who require no resources upon examination.
2.1.1 Patient prioritization strategies.
To demonstrate our evaluation framework, we apply it to nine distinct patient prioritization strategies. A comprehensive technical description of each strategy is available in Appendix A; this section provides a high-level summary. The strategies are grouped into three categories: established baselines from the literature, novel rules developed for this study, and composite strategies that integrate the novel approaches.
- Established Baseline Strategies
- First-Come-First-Served (FCFS): patients are selected based strictly on their arrival order, irrespective of their acuity level.
- Acuity-Based FCFS: prioritizes patients with higher acuity, using the FCFS rule to resolve ties within the same acuity level.
- Accumulating Priority Queue (APQ): integrates both patient acuity and their current length of stay (LOS). A patient’s priority score is calculated by multiplying their LOS by a predefined acuity weight.
- Novel Base and Add-on Strategies
- Additive Accumulating Priority Queue (AAPQ): a novel base rule. It defines a patient’s score as the sum of their acuity weight and a small, scaled LOS term. This design primarily orders patients by acuity while using their current LOS as a tiebreaker, which makes it intuitive and easy to implement.
- Low Workload Physician (LWP): a conditional “add-on” rule designed to balance physician workload in real-time. If a physician has significantly fewer active patients than their peers, this rule assigns them the highest-priority patient from the queue, temporarily overriding the default prioritization logic.
- Partial Fast Track (PFT): another add-on that designates one physician to preferentially treat low-acuity patients (i.e., ESI 4 and 5) on an FCFS basis whenever such patients are available in the queue. The other physicians continue to serve all patients according to the base rule, ensuring flexibility.
- Composite Strategies
- AAPQ-LWP and AAPQ-PFT: layer the LWP and PFT add-on rules over the AAPQ base rule, respectively.
- AAPQ-LWP-PFT: sequentially applies the logic of PFT, then LWP, before defaulting to the AAPQ base rule if neither of the add-on conditions is met.
2.1.2 Notation.
Finally, Table 3 summarizes the notation we use throughout this manuscript.
2.2 Technique 1: KPI summary statistics and tail analysis
This technique reflects standard practice by summarizing the distribution of each KPI. It is intentionally simple and does not incorporate advanced methodological tools. We include it because many studies provide limited distributional insight, typically reporting only the mean and occasionally the standard deviation. In contrast, our approach reports both central-tendency and right-tail metrics, as the extreme upper tail of the distribution often represents the worst outcomes, which pose the greatest risks to patients and hospital operations but are rarely discussed in the literature.
Assuming that lower KPI values indicate better performance, we recommend reporting a comprehensive set of statistics for each strategy. These include the sample size, which ensures transparency about the number of observations and supports the assessment of statistical power; the mean, which reflects the average outcome and facilitates comparison of overall performance; and the median, which serves as a robust measure of central tendency that is not overly influenced by outliers. Additionally, reporting the minimum and maximum values allows for identification of the best and worst observed outcomes, respectively—offering insight into exceptional performance as well as potential failures.
To quantify tail risk explicitly, we also recommend including the 75th, 90th, 95th, and 99th percentiles of each KPI distribution. These percentiles reveal how frequently patients experience extremely long waits or lengths of stay, offering a granular view of performance in high-risk scenarios. This practice is consistent with prior guidelines in emergency department analytics [3].
Each of these metrics can—and should—be reported separately for relevant cohorts. In Section 3 (Illustrative Application of the Framework), for example, we will present these statistics for all patients combined, as well as stratified by Low and Mid acuity groups, which will reveal differing performance profiles when disaggregated. At a minimum, reports should be stratified by acuity level; additional cohort definitions might include arrival‐time window, required resource type, boarding status, or other clinically meaningful categories.
2.3 Technique 2: Threshold-based performance
While summary statistics and tail percentiles (Technique 1) reveal overall distributional properties, stakeholders often specify explicit time targets—for example, the proportion of patients discharged within four hours. Technique 2 addresses this by quantifying, for each strategy and cohort, the fraction of patients whose KPI falls below a clinically meaningful threshold. By evaluating this performance over a range of thresholds, analysts can visualize and compare how sensitive each prioritization rule is to the choice of time target.
We begin by defining the indicator function
Let denote a patient cohort (e.g., all patients or those in a given acuity group), and let
be the KPI value for patient
(e.g., length of stay, waiting time, door-to-doctor time, boarding indicator, or satisfaction score). For a specified threshold
, we define the threshold-attainment function
which represents the proportion of patients in cohort whose KPI does not exceed
.
In practice, we compute over a grid of thresholds
. Plotting
against
yields a threshold-attainment curve, with higher curves indicating faster achievement of the target. Separate curves are generated for each strategy
and cohort
, enabling direct visual comparison. Confidence bands (e.g., bootstrap 95% intervals) may be overlaid to assess statistical significance.
For our illustrative analysis, we compute and
where
corresponds to the low- and mid-acuity groups, respectively, and
is each patient’s length of stay. By comparing these curves across the nine prioritization rules, one can readily identify how many patients meet a four‐hour discharge target and observe how performance changes under more stringent (e.g., three‐hour) or more lenient (e.g., five‐hour) thresholds.
Threshold‐attainment analysis offers several key advantages. First, plotting curves over a continuous range of facilitates a sensitivity analysis, since intersections of curves reveal threshold ranges in which one strategy outperforms another. Second, curves directly encode clinical relevance, allowing stakeholders to read off the proportion of patients meeting an institution’s policy‐driven time target without recomputing separate summary statistics. Third, when it is impractical to display full curves—such as in print‐constrained venues—selecting a few representative percentile targets (for instance, the time to 75 percent, 90 percent, or 95 percent attainment) and tabulating those values can achieve space efficiency.
Besides looking at the curves, we can calculate summary metrics from to make comparisons easier. The AUC over the interval
is defined by
or more generally
where is a weight function (e.g., linear or exponential decay) emphasizing early, middle, or late thresholds and
represents the upper bound of interest (such as 12 hours for LOS). To standardize comparisons across KPIs, we report the standardized AUC as
, which ranges from 0 to 1. Appendix D provides the full details of our AUC calculations. AUC is particularly beneficial because it consolidates performance across the entire range of clinically relevant thresholds into a single measure, enabling straightforward quantitative comparison of different strategies across all thresholds.
A second concise metric is the speed to attainment, defined by
which answers the question, “How long until of patients meet the KPI target?” For example, one might report that “Strategy A reaches 90 percent discharge 25 minutes faster than Strategy B for mid‐acuity patients.” Because
is a stakeholder‐defined parameter, it can be varied to reflect different operational goals.
Researchers can tabulate AUC or values for key percentiles—such as 50 percent, 75 percent, 90 percent, and 95 percent—to present succinct comparisons without plotting every curve. All threshold‐based metrics—full curves, AUC, and
—should be reported separately for each cohort of interest. In Section 3, we will present these metrics for low‐acuity and mid‐acuity groups, demonstrating how strategy rankings may shift when performance is disaggregated. At a minimum, stratification by acuity level is required; additional analyses may consider other cohorts (for example, arrival‐time windows, required resource types, or boarding status) to uncover subgroup‐specific trade‐offs.
2.4 Stakeholder-informed utility functions
Techniques 1 and 2 show what each rule achieves in terms of KPI distributions and threshold attainment, but they don’t show how much stakeholders value those outcomes. Technique 3 remedies this by mapping multi-dimensional performance into a single, stakeholder-aligned score, , constructed to increase as outcomes improve. While the precise functional form—whether linear, quadratic, exponential, Chebyshev, or otherwise—is specified in Appendix B, the following paragraphs describe how utility-based visualizations enhance decision support and how we apply them in our illustrative application of the framwork. The numerical utility parameters used in our analysis are illustrative placeholders; they were not elicited from ED stakeholders and serve only to demonstrate Technique 3.
First, one can plot
for each strategy where and
are two cohorts. In our illustrative analysis, we use
. By tracking utility across the same grid of LOS thresholds used in Technique 2, these curves reveal trade-offs in a single view: a rule that underperforms at short thresholds may nonetheless deliver high stakeholder value at longer thresholds, indicating favorable tail performance despite mediocre central metrics.
Second, for any fixed threshold , a scatter plot for a specific threshold is an easy-to-understand visual. We plot
on the horizontal and vertical axes, respectively, and overlay utility contours . Each strategy appears as a point colored by its utility level (e.g., green for high, red for low). Note that both axes remain raw KPI percentages; utility influences only the contour lines and point colors. This visualization combines the interpretability of familiar KPI percentages with a direct encoding of stakeholder preferences.
Finally, when multiple utility formulations or hyperparameter settings are under consideration, a cross-utility comparison can be performed by plotting against
for each strategy at selected thresholds. Near-linear alignment along a positively sloped line indicates that strategic ordering is robust to the choice of utility specification; divergence suggests that different stakeholder weighting induces materially different recommendations, warranting closer examination.
These utility‐driven plots enrich our evaluation framework by embedding explicit stakeholder trade‐offs into familiar performance metrics, highlighting threshold dependence—since the utility curves mirror LOS threshold curves while summarizing value rather than raw attainment—and testing robustness through cross‐utility comparisons that guard against overconfidence in any single preference model.
3. Illustrative application of the framework
All results in this section are illustrative—they reflect a single discrete-event-simulation calibration for a 30-bed mixed-acuity ED and utility parameters chosen for demonstration only. They should not be interpreted as prescriptive guidance for any specific hospital.
3.1 KPI summary statistics and tail analysis
Table 4 presents LOS summary statistics for the combined low- and mid-acuity cohort. Across the nine prioritization strategies, the average LOS is very similar—ranging from 193 to 200 minutes. However, the longest stays (the right-hand tail) vary a lot, with the 99th percentile spanning 59 minutes.
When stratified by acuity level, sharper contrasts emerge. For low-acuity patients, the mean LOS range triples to 22 minutes (170–192 min) and the 99th-percentile spread widens to 102 minutes. In contrast, mid-acuity patients exhibit a mean range of 74 minutes (189–263 min) and a 99th-percentile spread of 116 minutes.
Acuity stratification also reverses strategy rankings. For low-acuity patients, FCFS outperforms all other rules on mean, median, and percentile metrics, whereas AAPQ-LWP-PFT yields the highest mean LOS. In the mid-acuity cohort, FCFS performs worst in mean, median, and upper-tail percentiles, while AAPQ-LWP-PFT ranks best on those metrics.
However, rankings vary by KPI. Among low-acuity patients, AAPQ-LWP-PFT produces the worst mean (albeit by a narrow margin) yet its right-tail performance remains near average. For mid-acuity patients, AAPQ-LWP excels on mean, median, and 75th-percentile metrics but performs poorly in more extreme tail measures, such as the 95th and 99th percentiles.
These findings highlight two key points: (1) averaging across different patient groups hides important differences, and (2) metrics focused on extreme outcomes often tell a different story than metrics based on the average.
3.2 Threshold-based performance
Fig 1 plots the percentage of low-acuity patients discharged within minutes, denoted
. Several patterns emerge. For short thresholds (
minutes), Acuity-Based FCFS outperforms other strategies by approximately 10 percentage points; however, its relative performance deteriorates at higher thresholds (
minutes). Beyond this, the choice of threshold does not meaningfully affect the relative ordering of most strategies.
This figure plots , the percentage of low-acuity patients discharged within
minutes, for each prioritization strategy over a 12-hour window. The early-time advantage of Acuity-Based FCFS is evident for
min—outperforming alternatives by roughly 10 percentage points—while its performance converges or declines relative to other rules at longer thresholds.
Analogous trends appear for mid-acuity patients in Fig 2, where the curves similarly suggest minimal divergence between strategies across thresholds. In this case, these illustrative results suggest that the choice of time target is unlikely to change the final decision. Notably, Fig 2 also reinforces the pattern observed in Section 3.1: FCFS underperforms during the initial 5 hours but surpasses other strategies at later thresholds.
This figure presents , the share of mid-acuity patients discharged by time
, across all nine rules. Although curves remain tightly clustered overall, FCFS underperforms during the first five hours yet surpasses many strategies at later thresholds, mirroring the tail-behavior patterns identified in Section 3.1.
We now present the AUC metrics, formally defined in Appendix D, in Table 5. In addition to the standard AUC computed over a 12-hour window, we introduce three complementary variants: (1) a half-range AUC limited to the first 6 hours, (2) a weighted AUC emphasizing earlier thresholds via a linearly decreasing weight function, and (3) a weighted AUC emphasizing later thresholds via a linearly increasing weight function. These variants enable us to evaluate whether the choice of AUC definition alters performance conclusions.
Overall, the conclusions remain stable across AUC definitions. One exception is Acuity-Based FCFS, which appears more favorable under the variant emphasizing early thresholds—consistent with Fig 1, where it dominates in the initial portion of the curve.
Next, we sort strategies from best to worst based on each metric. This confirms earlier trends: FCFS ranks highest for low-acuity patients but lowest for mid-acuity patients, whereas AAPQ-LWP-PFT achieves the opposite pattern, performing best in the mid-acuity group but worst in the low-acuity group.
Finally, we quantify each strategy’s responsiveness in Table 6 by reporting the minimum LOS threshold required to serve 50%, 75%, 90%, and 95% of patients. These metrics respectively correspond to the time needed to reach the median, upper-quartile, near-complete, and very-high service-level benchmarks. Together, they provide a granular view of how quickly each strategy meets increasingly stringent performance goals. We find that the choice of threshold meaningfully affects the relative ranking of some strategies but not others. For instance, in the low-acuity cohort, AAPQ-PFT reaches 50% of patients within 185 minutes—midway between FCFS (170 min) and Acuity-Based FCFS (200 min). However, to reach 95%, AAPQ-PFT requires 315 minutes, which is far closer to the worst-performing Acuity-Based FCFS (325 min) than to the best-performing FCFS (260 min). In contrast, strategies like AAPQ-LWP-PFT in the mid-acuity cohort demonstrate consistently strong performance across all thresholds, making them less sensitive to the choice of benchmark.
3.3 Stakeholder-informed utility analysis
We evaluate two utility functions—an elliptical form, , and a linear form,
(definitions in Appendix B)—by first plotting their values against the LOS threshold
. Figs 3 and 4 display the curves
and
, respectively, allowing us to observe how strategy rankings evolve as
increases. Because both curves preserve the same ordering up to approximately five hours, the choice of threshold has minimal impact on the relative performance of the rules—whereas frequent crossings would mandate consideration of multiple thresholds to capture divergent conclusions.
This plot shows the elliptical utility for each strategy as a function of the discharge-time threshold
. The mid-acuity weighting hyperparameter is λ = 1.5 and the overshooting value is δ = 0.3. Consistent vertical ordering up to approximately five hours indicates that strategy rankings under
are robust to the choice of
within this range. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
This figure depicts the linear utility plotted against
. The low-acuity downweighing parameter is set to α = 0.6. The near–parallel curves for most rules confirm that, under
, conclusions about relative strategy performance remain largely unaffected by threshold selection. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
We then fix a threshold and generate scatter plots of
overlaid with utility contours
, using color to encode utility level. In Figs 5 and 6, which employ the elliptical utility
, we set
and 7h. At five hours, FCFS achieves the highest low-acuity discharge rate but ranks poorly overall because its mid-acuity rate trails the next-worst strategy (APQ) by roughly 12 percentage points, resulting in low utility. The other strategies differ only marginally on low-acuity performance, with mid-acuity rates spanning about nine points between APQ and AAPQ-LWP-PFT. At seven hours, FCFS’s utility improves dramatically—rising from worst to second best—illustrating how deeper tail performance can overturn short-threshold conclusions. Figs 7 and 8 repeat this analysis with the linear utility
, confirming that FCFS’s strong long-threshold performance persists despite its weaker results at shorter thresholds.
The mid-acuity weighting hyperparameter is λ = 1.5, the overshooting value is δ = 0.3, and LOS goals are 100% of low-acuity patients and 90% of mid-acuity patients. Scatter plot of for each strategy, overlaid with elliptical-utility contours
. At this threshold, FCFS maximizes low-acuity discharges but scores low overall due to substantially poorer mid-acuity performance, illustrating the trade-off captured by
. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
The mid-acuity weighting hyperparameter is λ = 1.5, the overshooting value is δ = 0.3, and LOS goal is set to 100% for both low and mid-acuity patients. Scatter plot of with
contours. At seven hours, FCFS moves from the worst to the second-best position, demonstrating how right-tail evaluation alters strategy rankings under the elliptical utility. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
The low-acuity downweighing parameter is set to α = 0.6 and LOS goals are 100% of low-acuity patients and 90% of mid-acuity patients. Scatter of with linear-utility contours
, the linear utility penalizes strategies with imbalanced performance, reaffirming FCFS’s disadvantage at shorter thresholds. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
The low-acuity downweighing parameter is set to α = 0.6 and LOS goal is set to 100% for both low and mid-acuity patients. Scatter of with
contours. The shift in FCFS’s relative position—from low to high utility—mirrors the pattern seen under the elliptical utility, underscoring consistency across utility forms.
Finally, Fig 9 compares and
at
by plotting each strategy’s pair of utility values. Since the points fall almost perfectly along the line, the choice of utility function and its parameters has little effect on the strategy rankings; a more scattered pattern would have revealed sensitivity to the utility specification.
Bivariate plot of each strategy’s utility values at
, with a fitted regression line. The mid-acuity weighting hyperparameter is λ = 1.5, the overshooting value is δ = 0.3, the low-acuity downweighing parameter is set to α = 0.6, and LOS goals are 50% of low-acuity patients and 50% of mid-acuity patients. The near-monotonic alignment indicates minimal sensitivity of strategy rankings to the choice of utility function or its hyperparameters. Utility parameters are placeholders used solely to demonstrate visualization and interpretation; no prescriptive conclusion should be drawn from these settings without stakeholder elicitation.
3.4 Strategy strengths and weaknesses
Table 7 synthesizes each strategy’s principal strengths and weaknesses, with supporting tables or figures indicated in parentheses. We reiterate that the aim of this analysis is to demonstrate our evaluation framework and to illustrate how it can uncover strategy‐specific trade-offs—not to endorse any particular rule. These findings should not be used as implementation guidance, since they are based on a single simulation scenario and on illustrative utility parameters rather than stakeholder-calibrated values.
4. Discussion
4.1 Synthesis of principal findings
This section interprets the illustrative outputs to show how the framework reveals complementary insights across tails, thresholds, and utilities. The goal is not to prescribe a specific strategy but to show how a standardized, multi-lens evaluation changes what appears “best” depending on risk tolerance and targets.
The illustrative example presented in Section 3 demonstrates that the choice of evaluation lens—summary statistics, threshold-attainment curves, or stakeholder-weighted utilities—can shape how ED prioritization strategies are ranked. Even within a single 30-bed simulated ED, a rule that looks benign under a LOS comparison can exhibit sizeable right-tail liabilities, while another rule that excels on extreme percentiles may fall behind when stakeholder utilities emphasize shorter thresholds. These differences highlight our main argument: no single metric can fully capture the effects of prioritization policies, and conclusions should never be based on just one number.
4.2 Why separating patient groups is essential
Averaging performance across different patient urgency levels hides the very differences that clinicians and managers need to see. When low- and mid-acuity LOS are pooled, the observed spread between strategies narrows to just 7 minutes, suggesting no meaningful distinction. However, when evaluated separately, the spread expands to 22 minutes for low-acuity patients and 74 minutes for mid-acuity patients—highlighting substantial and actionable differences in strategy performance.
By diluting the average LOS in this fashion, an analyst might incorrectly infer that any of the nine rules is “good enough,” missing the variation for each group of patients. Operationally, such masking can divert resources toward the wrong bottleneck or delay recognition of inequitable wait times among patient groups.
4.3 Strategic trade-offs across patient groups
Ideal policies would compress the entire KPI distributions for every cohort, yet the data reveal persistent trade-offs. FCFS and AAPQ-LWP-PFT epitomize this tension: the former minimizes low-acuity mean and upper-tail LOS, whereas the latter delivers the best mid-acuity central and tail performance at the expense of longer waits for low-acuity cases.
These findings reinforce the need for hospitals to state explicit priorities—whether reducing crowding for mid-acuity patients, accelerating flow for discharged fast-track patients, or balancing both objectives through mixed strategies or dynamic rules.
4.4 Statistical significance versus clinical relevance
Table 2 shows that the worst low-acuity mean LOS (192 min) is twenty-two minutes longer than the best (170 min). While statistically significant, administrators must decide whether a twenty-minute average difference is operationally meaningful when the clinically acceptable window for discharge may span several hours. Hence, statistical tests alone cannot substitute for clinical judgment; practical significance must be interpreted in context.
4.5 Actionability of threshold-attainment metrics
Decision makers consistently ask, “How quickly can we reach a given service target?” Table 6 answers this in language executives and frontline clinicians immediately understand. For instance, APQ discharges half of low-acuity patients twenty minutes sooner than Acuity-Based FCFS, and it reaches the 90 percent benchmark thirty minutes earlier. Such statements translate directly into staffing discussions, fast-track lane designs, and patient-communication scripts.
4.6 Insights from strengths–weaknesses mapping
The qualitative synthesis in Table 7 distills dozens of numerical comparisons into a concise narrative of strategic profiles. Strikingly, all identified strengths and weaknesses stem from a single KPI—LOS—in one virtual ED. If this level of heterogeneity arises from one performance dimension, the variation across additional KPIs such as left-without-being-seen rates, bed-blocking time, or door-to-doctor-time is likely far greater. The implication is sobering: past studies that benchmarked policies on a single mean LOS or four-hour target may have drawn overly general conclusions, exacerbating translation gaps when rules are transplanted to new hospitals.
4.7 Central role of sensitivity analysis
Throughout Section 3, sensitivity analysis served as a safeguard against misleading conclusions. For instance, Figs 5 and 6 displayed the Elliptical utility function evaluated at five- and seven-hour thresholds, respectively. Despite the modest two-hour difference, the resulting strategy rankings shifted noticeably. FCFS, which appeared to perform worst under the five-hour threshold, ranked among the best when evaluated at seven hours. Relying on a single threshold would have given an interpretation that depended on that specific threshold and could have been misleading. In contrast, other metrics, such as AUC-to-time-to-threshold, demonstrated robustness to parameter choices, producing consistent results across variations. Nonetheless, systematically conducting sensitivity analyses remains helpful to guard against occasional but consequential parameter-driven artifacts. Recognizing these dependencies discourages overconfidence in model generalizability and helps ensure that operational insights remain valid when applied to alternate settings.
4.8 Recommended steps for strategy evaluation
To ensure that patient‐prioritization studies are both rigorous and transparent, we offer the following guidelines for researchers. These principles will help one apply each evaluation technique consistently, highlight strategy trade‐offs, and guard against overconfident claims of “best” performance.
Technique 1 should be employed across all KPIs and cohorts by reporting sample size, mean, median, minimum, maximum, and key upper‐tail percentiles (75th, 90th, 95th, 99th) for each clinically relevant subgroup (e.g., acuity level, arrival window, resource requirement, boarding status). Technique 2 (threshold‐attainment) ought to accompany every KPI, with concise tabular summaries of the time required to reach a stakeholder‐defined attainment level (such as 90 percent) for each cohort, thus providing both clinical relevance and ease of interpretation. Technique 3 (stakeholder‐informed utility) should then translate the most critical KPI(s) into a single utility score that incorporates stakeholder preferences—thereby facilitating equitable, real‐world comparisons. Throughout, results should be presented as context‐dependent trade‐offs rather than absolute winners, with each strategy’s strengths and weaknesses clearly articulated across techniques and cohorts; universal superiority should be asserted only when an exhaustive range of KPIs, cohorts, evaluation methods, and operational scenarios has been rigorously tested and validated.
We conclude by reflecting on how often each evaluation technique was cited in our analysis summary in Section 3.4 (Table 7). We counted the number of times Techniques 1–3 were referenced when comparing each strategy’s strengths and limitations. Technique 1 (detailed summary statistics) accounted for only 26% of all citations—yet most researchers report only a less‐detailed subset of these metrics, implying much information about upper‐tail behavior and cohort‐specific performance is routinely omitted. Technique 2 (threshold attainment) and Technique 3 (stakeholder‐informed utility) comprised roughly 29% and 45% of references, respectively. While citation frequency does not directly imply importance, it offers a useful proxy for how readily each approach reveals key trade‐offs—underscoring the need to apply all three in concert rather than relying on any single method.
4.9 Limitations and directions for future research
We reiterate that the purpose of this study is to introduce and demonstrate a suite of evaluation techniques—not to endorse any particular prioritization strategy. There are several limitations, so the example results should be interpreted with caution. The simulation was calibrated to a single ED with 30 beds and moderate patient volume. As such, different acuity distributions, arrival patterns, boarding durations, or staffing constraints in other settings may yield materially different rankings. Moreover, our analysis centered on LOS, supplemented only by a brief DTDT example in Appendix E, and did not extend to other outcomes such as LWBS, patient safety, or staff workload. The utility parameters used in this study were intended for illustrative purposes only; they were not elicited from stakeholders or validated through multicenter analysis.
This paper introduces three evaluation techniques designed to promote more rigorous and transparent comparisons of patient prioritization strategies. While these techniques provide a structured foundation, they are not comprehensive. Future work should aim to develop additional methods that illuminate distinct strengths and weaknesses of each strategy. Although quantitative tools such as AUC can be useful, their benefits are limited if they lack interpretability or practical relevance. When mathematical measures are employed, visualizations should be designed to clearly convey real-world implications—as exemplified by the utility function plots in Figs 5–8.
Our evaluation framework was applied to a single ED configuration, making it well suited for guiding local implementation decisions. However, many researchers aim to develop strategies that generalize across multiple ED settings. The most direct approach to support such generalization is to apply the evaluation techniques across a diverse set of simulated ED environments—ideally spanning 4–8 configurations that vary by bed count, staffing model, and arrival volume. While this approach provides valuable insights, it is resource-intensive: even a single-configuration analysis across multiple KPIs can span dozens of pages. A promising direction for future research is to identify principled methods for aggregating results across configurations. One such approach, based on covariance ovals, is introduced in Appendix B.5. Care must be taken to preserve between-scenario variability, as aggregation can obscure meaningful differences. For example, in Section 3.1, combining low- and mid-acuity cohorts reduced the mean LOS difference across strategies from 22 and 74 minutes, respectively, to only 7 minutes—thereby masking clinically relevant distinctions.
Although the proposed techniques allow rigorous numerical comparisons of patient-prioritization strategies, implementation feasibility must be assessed before any analysis begins. A reinforcement-learning (RL) policy (e.g., see Lee and Lee [27]), for example, may offer superior theoretical performance, yet its practical requirements can be prohibitive. Many RL formulations presume real-time knowledge of each patient’s location, resource needs, and even uncertain variables such as future arrivals and service times. Collecting this information in simulation is straightforward; acquiring it in a working ED would demand either extensive manual data entry—adding workload for clinical staff—or continuous computer-vision monitoring, which entails new infrastructure, software, and privacy concerns. A strategy that looks good in a simulation but creates major operational or ethical problems in a real hospital is not truly beneficial. Researchers should therefore verify that any recommended policy can be deployed with existing data streams and minimal additional burden on personnel; otherwise, theoretical optimality will remain purely academic.
Finally, we briefly reflect on the performance of our novel strategies—PFT, LWP, AAPQ, and their combinations—within the illustrative analysis. While we emphasize that these illustrative results are not intended as definitive endorsements, it is noteworthy that our proposed strategies consistently outperformed traditional benchmarks such as Acuity-Based FCFS and APQ across multiple evaluation criteria. As such, these strategies represent a secondary contribution to this work: they appear promising and merit further investigation. In addition to their favorable quantitative performance, they were deliberately designed for ease of implementation and interpretability, making them attractive candidates for real-world adoption.
5. Conclusion
This study improves the study of ED operations by presenting and demonstrating a three-part evaluation framework—tail-sensitive summary statistics, threshold-attainment profiles, and stakeholder-informed utility analysis—for transparent comparison of patient-prioritization strategies. Applying the framework to nine rules within a common discrete-event simulation revealed that strategic rankings are highly sensitive to the chosen evaluative lens and to cohort aggregation: analyses confined to overall means understated clinically meaningful differences that emerged once low- and mid-acuity patients were considered separately, and strategies that appeared dominant under central-tendency metrics were often eclipsed when extreme percentiles or utility-weighted outcomes were examined. These findings confirm that no single KPI, time target, or composite score adequately captures the multidimensional consequences of queue-management decisions, and they underscore the practical necessity of reporting a minimum set of distributional, threshold-based, and preference-aligned measures.
By framing results as trade-offs rather than pronouncing universal “winners,” the proposed framework equips hospital leaders to align prioritization rules with explicit local objectives—whether accelerating flow for mid-acuity patients, protecting low-acuity throughput, or balancing both via mixed or adaptive policies. Because the three techniques use common statistics and simple graphs, the method is easy for different stakeholders to understand while still being rigorous enough for researchers and quality-improvement teams. Moreover, the framework is KPI-agnostic and extensible to additional outcomes such as door-to-doctor time, boarding duration, patient safety indicators, or staff workload, inviting comprehensive performance audits without presupposing any specific metric hierarchy.
Several limitations temper the generalizability of the illustrative results. The simulation reflected a single, moderately busy, 30-bed ED; different capacity profiles, arrival patterns, or boarding pressures may yield alternative strategic orderings. Only length of stay was modeled, utility parameters were illustrative rather than elicited, and no external validation across multiple health systems was undertaken. Future research should therefore replicate the framework across diverse ED configurations, incorporate a broader KPI portfolio, elicit context-specific utilities, and test whether adaptive or hybrid strategies can dominate static rules when assessed under the full triad of metrics. Methodological work is also needed to synthesize results across multiple scenarios without obscuring cross-site heterogeneity—an aggregation challenge analogous to cohort masking within a single ED.
Notwithstanding these caveats, the present contribution provides a framework that reconciles statistical robustness with managerial interpretability, furnishing researchers and practitioners with a common language for evidence-based policy design. Widespread adoption of this evaluation standard promises to accelerate meta-analysis, clarify when and where novel prioritization algorithms add value, and ultimately promote safer, timelier, and more equitable emergency care.
References
- 1. Moskop JC, Geiderman JM, Marshall KD, McGreevy J, Derse AR, Bookman K, et al. Another look at the persistent moral problem of emergency department crowding. Ann Emerg Med. 2019;74(3):357–64. pmid:30579619
- 2. Mostafa R, El-Atawi K. Strategies to measure and improve emergency department performance: a review. Cureus. 2024.
- 3. Vanbrabant L, Braekers K, Ramaekers K, Van Nieuwenhuyse I. Simulation of emergency department operations: a comprehensive review of KPIs and operational improvements. Comput Ind Eng. 2019;131:356–81.
- 4. Sharifi Kia A, Rafizadeh M, Shahmoradi L. Telemedicine in the emergency department: an overview of systematic reviews. J Public Health. 2022.
- 5. Cha WC, Shin SD, Cho JS, Song KJ, Singer AJ, Kwak YH. The association between crowding and mortality in admitted pediatric patients from mixed adult-pediatric emergency departments in Korea. Pediatr Emerg Care. 2011;27(12):1136–41. pmid:22134231
- 6. Guttmann A, Schull MJ, Vermeulen MJ, Stukel TA. Association between waiting times and short term mortality and hospital admission after departure from emergency department: population based cohort study from Ontario, Canada. BMJ. 2011;342:d2983. pmid:21632665
- 7. Hong KJ, Shin SD, Song KJ, Cha WC, Cho JS. Association between ED crowding and delay in resuscitation effort. Am J Emerg Med. 2013;31(3):509–15. pmid:23159432
- 8. Savioli G, Ceresa IF, Gri N, Bavestrello Piccini G, Longhitano Y, Zanza C, et al. Emergency department overcrowding: understanding the factors to find corresponding solutions. J Pers Med. 2022;12(2):279. pmid:35207769
- 9. Ashour OM, Okudan Kremer GE. Dynamic patient grouping and prioritization: a new approach to emergency department flow improvement. Health Care Manag Sci. 2016;19(2):192–205. pmid:25487711
- 10. Samadbeik M, Staib A, Boyle J, Khanna S, Bosley E, Bodnar D, et al. Patient flow in emergency departments: a comprehensive umbrella review of solutions and challenges across the health system. BMC Health Serv Res. 2024;24(1):274. pmid:38443894
- 11. Furterer SL. Applying Lean Six Sigma methods to reduce length of stay in a hospital’s emergency department. Qual Eng. 2018;30(3):389–404.
- 12. Bilodeau B, Stanford DA. High-priority expected waiting times in the delayed accumulating priority queue with applications to health care KPIs. INFOR: Inf Syst Oper Res. 2022;60(3):285–314.
- 13. Cildoz M, Ibarra A, Mallor F. Accumulating priority queues versus pure priority queues for managing patients in emergency departments. Oper Res Health Care. 2019;23:100224.
- 14. Ferrand YB, Magazine MJ, Rao US, Glass TF. Managing responsiveness in the emergency department: comparing dynamic priority queue with fast track. J Oper Manag. 2018;58–59(1):15–26.
- 15.
Daknou A, et al. A dynamic patient scheduling at the emergency department in hospitals. IEEE; 2010.
- 16. Burdett RL, Kozan E. An integrated approach for scheduling health care activities in a hospital. Eur J Oper Res. 2018;264(2):756–73.
- 17. de Queiroz TA, et al. Dynamic scheduling of patients in emergency departments. Eur J Oper Res. 2023;310(1):100–16.
- 18. Harzi M, Condotta J-F, Nouaouri I, Krichen S. Scheduling patients in emergency department by considering material resources. Procedia Comput Sci. 2017;112:713–22.
- 19. Li W, Sun Z, Hong LJ. Who is next: patient prioritization under emergency department blocking. Oper Res. 2021.
- 20.
Furian N, et al. Machine learning-based patient selection in an emergency department. arXiv:2206.03752 [Preprint]. 2022.
- 21. Alipour-Vaezi M, Aghsami A, Jolai F. Prioritizing and queueing the emergency departments’ patients using a novel data-driven decision-making methodology, a real case study. Expert Syst Appl. 2022;195:116568. pmid:35125674
- 22. Zhang A, Zhu X, Lu Q, Zhang R. Impact of prioritization on the outpatient queuing system in the emergency department with limited medical resources. Symmetry. 2019;11(6):796.
- 23. Ding Y, Park E, Nagarajan M, Grafstein E. Patient prioritization in emergency department triage systems: an empirical study of the Canadian Triage and Acuity Scale (CTAS). M&SOM. 2019;21(4):723–41.
- 24. Li N, Stanford DA. Multi-server accumulating priority queues with heterogeneous servers. Eur J Oper Res. 2016;252(3):866–78.
- 25. Doudareva E, Carter M. Discrete event simulation for emergency department modelling: a systematic review of validation methods. Oper Res Health Care. 2022;33:100340.
- 26. Hou J, Zhao X. Using a priority queuing approach to improve emergency department performance. JMA. 2019;7(1):28–43.
- 27. Lee S, Lee YH. Improving emergency department efficiency by patient scheduling using deep reinforcement learning. Healthcare. 2020.
- 28. Luscombe R, Kozan E. Dynamic resource allocation to improve emergency department efficiency in real time. Eur J Oper Res. 2016;255(2):593–603.
- 29. Wlamyr P-A, Luna-Pereira HO, Caicedo-Rolon AJ. Discrete event simulation in the optimization of emergency departments. J Lang Linguist Stud. 2022;18(4).
- 30. Claudio D, Okudan GE. Utility function-based patient prioritisation in the emergency department. EJIE. 2010;4(1):59.
- 31. Mitchell R, Fang W, Tee QW, O’Reilly G, Romero L, Mitchell R, et al. Systematic review: what is the impact of triage implementation on clinical outcomes and process measures in low- and middle-income country emergency departments? Acad Emerg Med. 2024;31(2):164–82. pmid:37803524
- 32. Sax DR, Warton EM, Mark DG, Vinson DR, Kene MV, Ballard DW, et al. Evaluation of the emergency severity index in US emergency departments for the rate of mistriage. JAMA Netw Open. 2023;6(3):e233404. pmid:36930151
- 33. Twomey M, Wallis LA, Myers JE. Limitations in validating emergency department triage scales. Emerg Med J. 2007;24(7):477–9. pmid:17582037
- 34. El Arab RA, Al Moosa OA. The role of AI in emergency department triage: an integrative systematic review. Intensive Crit Care Nurs. 2025;89:104058. pmid:40306071
- 35. Ryu AJ, Romero-Brufau S, Qian R, Heaton HA, Nestler DM, Ayanian S, et al. Assessing the generalizability of a clinical machine learning model across multiple emergency departments. Mayo Clin Proc Innov Qual Outcomes. 2022;6(3):193–9. pmid:35517246
- 36. Shin HA, Kang H, Choi M. Triage data-driven prediction models for hospital admission of emergency department patients: a systematic review. Healthc Inform Res. 2025;31(1):23–36. pmid:39973034
- 37. Ingielewicz A, Rychlik P, Sieminski M. Drinking from the Holy Grail-does a perfect triage system exist? And where to look for it? J Pers Med. 2024;14(6):590. pmid:38929811
- 38. Siddiqui E, Daniyal M, Khan MAR, Siddiqui S ul I, Siddiqui Z ul I, Farooqi W. Triage for low income countries: is ESI truly the way forward? Int J Community Med Public Health. 2018;5(11):5003.