This secondary economic analysis used data from a previously reported preintervention and postintervention study of the AI triage CDS across 3 EDs within a single health system [8]. Implementation occurred in staggered fashion between May 19, 2021, and April 4, 2023. The analysis included 170,723 ED visits during 180-day preintervention and postintervention periods at each site, excluding 2.2% (3925/174,648) of visits in which patients left prior to disposition. Cohort demographics and patient outcomes have been previously reported [8].

The study received approval from the Yale University Institutional Review Board (2000036009) and is reported in accordance with the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) statement (Checklist 1) [23].

The expected operating margin rate of 5.8% was adjusted from a national report by Wilson and Cutler [19] to account for payer type and admission rate across the study sites.

The AI triage CDS [8,9,24-26] applies site-specific machine learning algorithms to ED triage data routinely collected on patient arrival to predict risk of 3 patient outcomes: critical care (defined as in-hospital mortality or intensive care unit admission), emergency surgery (any surgery performed in a dedicated operating room suite within 12 hours of ED disposition), and hospital admission. Predictors include demographics (age and sex), arrival mode, vital signs (heart rate, respiratory rate, oxygen saturation, temperature, and systolic blood pressure), chief complaint, and active medical problems.

Predicted probabilities are translated to recommended acuity levels (1‐5, with lower values indicating higher acuity). Probabilities for critical care and emergency surgery meeting specific thresholds trigger recommendations for highest-risk levels (1 or 2), while probability of hospitalization guides recommendations for lower risk levels (3-5) [8]. The AI triage CDS output appears within seconds in the ED nurse workflow, displaying the recommended triage level with individualized explanations of the recommendation generated using Shapley Additive Explanations values transformed to natural language for each patient. Nurses retain triage decision-making autonomy and may assign acuity levels aligned with or divergent from the AI triage CDS recommendations.

Triage acuity levels dictate early course of care for patients and shape ED operational efficiency [27-29]. High-acuity patients are prioritized for immediate evaluation, while low-acuity patients are commonly diverted to high-efficiency workstreams, such as fast-track or vertical care areas (not requiring an ED bed), where they can be treated quickly. Midacuity patients typically follow traditional ED workflows with longer wait times [27].

Postintervention, low-acuity visit assignments (levels 4‐5) increased by a relative 48.2% (from 19,934/83,404, 23.9% to 32,300/91,244, 35.4%), enabling more patients to benefit from streamlined fast-track pathways. Conversely, midacuity (level 3) visits decreased 18.7% (from 40,701/83,404, 48.8% to 36,223/91,244, 39.7%), reducing the proportion of patients in the ambiguous category that typically experiences prolonged wait times [8]. Median ED LOS for all patients decreased 6.1% (19/311.0 min; from 311 to 292 min), with 5% (15.6/311.0 minutes) attributed to the intervention after adjustment for confounders [8]. These efficiency gains created capacity for an additional 9.6% (7795/81,466) of visits [8].

The economic model measured total ED revenue (before the intervention and after the intervention) and provided 2 approaches for estimating costs and corresponding operating margin. Revenue was defined as the payment received by the hospital for health care services; this excludes fees for physician-specific services. Costs were defined as expenses incurred by the hospital to provide health care services [20]. Operating margin was defined as revenue minus cost, with the corresponding operating margin rate shown in Textbox 1 [30]. Operating margin indicates profitability of core patient care operations [30,31]. All economic model inputs and assumptions may be adjusted to reflect local hospital characteristics and projections. A spreadsheet (Microsoft Excel) for static calculations and a script (Python version 3.8; Python Software Foundation) to enable additional sensitivity analyses is publicly available for download [22]. Table S1 in Multimedia Appendix 1 provides a guide to support custom model inputs based on published data that have been inflation-adjusted to current value (2025) using the Bureau of Labor Statistics Consumer Price Index for Medical Care [32]. A formal economics analysis plan was not developed prior to this secondary evaluation; the economic model structure, assumptions, and analytic approach are described in this section, and the full model is publicly available.

Hospital revenue data for the cohort were collected from billing tables housed in the electronic health record system. Revenue for ED discharged patients was attributed wholly to the ED, while revenue for hospital admitted patients was calculated as a fraction of the revenue realized from their entire hospital stay [19]. This fraction was calculated as 24 hours divided by total hospital LOS (hours) for hospital stays exceeding 24 hours, and 12 hours divided by total LOS for stays under 24 hours. To mitigate the influence of outliers common in revenue data, revenue per ED visit was capped at the 95th percentile (US $7241). This approach purposefully yields conservative estimates of total revenue and operating margin by excluding potential high-revenue outliers, which follow high-acuity emergency pathways independent of triage methods. For patients with missing revenue estimates (5447/170,723, 3.19%), values were imputed using the mean total revenue within strata defined by admission status (discharge or admit) and primary payer type (Medicare, Medicaid, private insurance, and self-pay).

Baseline preintervention cost was estimated by presuming a 5.8% operating margin rate on observed revenue. This rate was derived for our study cohort based on payer-specific margin rates reported by Wilson and Cutler [19], weighted by the distribution of revenue across payer and admission groups in our preintervention data (Textbox 1). Thus, the total preintervention cost was calculated as total revenue multiplied by (1–0.058). The resulting preintervention total costs and average cost per ED visit served as the reference for postintervention comparisons. Postintervention operating margin was calculated as revenue minus cost.

Two distinct cost frameworks were applied to postintervention data. The hospital management perspective accounts for time and resources, allocating 70% of total costs as fixed for standby ED care, 20% as variable costs dependent on patient volume, and 10% as modifiable clinician labor that fluctuates with care hours (Textbox 1). This allocation is consistent with prior ED cost analyses [12-14] and reflects that most ED costs—including standby capacity labor, facility, equipment, and overhead—accrue independent of patient throughput, while supplies and pharmaceuticals vary with volume, and a smaller portion of labor can be flexed through mechanisms such as contract labor [18], overtime, or flexible staffing. Though exact ratios may vary by institution, the model allows adjustment of these assumptions. Alternatively, the public policy perspective applies a consistent cost-per-visit approach based on hospital charges and cost-to-charge ratios, holding average cost per ED visit constant regardless of time and resources consumed (ie, care hours).

The economic model was applied to preintervention and postintervention data using measured revenue and both cost frameworks (hospital and public policy) to estimate costs and operating margin changes relative to baseline (before intervention). Sensitivity analyses were performed on these results to both conceptualize and generalize how changes in ED efficiency (ED LOS and patient throughput) translate to financial metrics under 2 common scenarios: (1) scenario 1, capacity-constrained EDs, where efficiency gains create new capacity for additional patient volume, applicable where demand for ED services exceeds supply; and (2) scenario 2, volume-stable EDs, where efficiency gains do not increase patient volume, applicable where demand for ED services is being met.

For both scenarios, ED LOS changes were modeled from the baseline observed median of 311 minutes to −10% (decreased efficiency, LOS increase to 342 min) to +10% (increased efficiency, LOS decrease to 280 min). Preintervention data (volume, LOS, and revenue) served as baseline comparators. Revenue, costs, and operating margin were estimated across this parameter space.

Break-even analyses were conducted to determine maximum AI tool cost per ED visit to achieve break-even, the point at which tool costs would exactly offset operating margin gains. Break-even estimates per visit were calculated as the difference in margin from the preintervention period to the postintervention period divided by total postintervention visits. This represents the per visit cost an efficiency-enhancing intervention could incur without negating the observed operating margin improvement.

The study demonstrates a financial impact assessment of an AI intervention designed to drive ED efficiency. The assessment illustrates how financial impact varied substantially based on cost-modeling framework. From a hospital management perspective, accounting for fixed and variable cost structures, the AI triage CDS implementation was associated with significant gains in operating margin, amounting to US $12.6 million (158% increase) in our multisite implementation. Conversely, public policy-level cost approaches substantially underestimated economic value, showing only US $0.9 million (11% increase) in margin gain despite identical revenue increases. The large difference in break-even thresholds between perspectives (US $66.02 vs US $4.69 per visit) has direct implications for implementation decisions and highlights the critical need for appropriate economic frameworks when evaluating ED efficiency interventions for a hospital or the public.

The increased financial value under the hospital management perspective reflects ED cost structure: approximately 70% of costs remained fixed (standby capacity), while only 20% varied with volume (supplies and pharmaceuticals) and 10% represented modifiable labor. This meant US $12.6 million of the US $15.4 million revenue increase translated directly to operating margin improvement. The cost structure assumptions (70% fixed, 20% variable, and 10% modifiable) align with prior ED cost analyses [12-14] and reflect the fundamental economics of emergency care delivery. Modifiable labor costs have become increasingly relevant given that contract nursing labor now represents 40% (2022) [18] of hospital-wide nursing expenses, with even higher use in EDs [17]. Our sensitivity analyses demonstrate that even modest reductions in modifiable costs through improved ED efficiency can generate substantial financial value in volume-stable ED settings.

While previous economic evaluations of ED interventions have documented efficiency gains [6,7,12,19,21], they typically do not explicitly distinguish between hospital management and policy-level cost perspectives. Our economic model highlights how cost-modeling approach (perspective) fundamentally shapes economic value assessment. Recent studies have highlighted the promise of AI tools for improving ED triage accuracy and patient flow [8,9,24], but comprehensive economic analyses comparing cost perspectives have been limited.

Hospital decision-makers should evaluate AI efficiency tools using cost frameworks that reflect operational economics, not policy-level per-visit approaches. In capacity-constrained EDs, even modest efficiency gains can yield substantial operating margin improvements when additional patient volume can be accommodated (Figure 1). In volume-stable ED settings, efficiency gains provide opportunities to reduce modifiable labor costs (eg, contract labor and overtime) where applicable. Although the patient perspective was not explicitly measured in these analyses, the operational efficiency gains (reduced LOS) underlying the modeled financial value are patient-relevant outcomes previously associated with reduced morbidity and improved satisfaction [3,8,9,27]. Efficiency gains that improve operating margin can also reduce patient wait times and crowding-related harms. Policymakers should recognize that traditional cost-per-visit approaches substantially underestimate value to hospitals and may inadvertently discourage adoption of beneficial technologies; reimbursement and incentive policies should account for hospitals’ actual cost structures and local market dynamics. Furthermore, developers of AI tools addressing ED crowding should adopt setting-specific value propositions that reflect differential impact based on ED capacity constraints and revenue-generating opportunity from new patient volume.

The economic model developed in this study is designed for broad applicability beyond the specific AI triage CDS implementation studied. We provide the model publicly [22] with guidance for adapting inputs to local contexts (Table S1 and user guide in Multimedia Appendix 1). Our baseline financial metrics align with other large-scale reports of ED financials (Table S1 in Multimedia Appendix 1), with observed differences attributable to cohort selection, payer mix, admission rates (acuity), and accounting methods for ED ancillary and physician-based services. While our cost allocation assumptions (70% fixed, 20% variable, and 10% modifiable labor) are a synthesis of published analyses, institutions can adjust these parameters to reflect local cost structures. Similarly, baseline operating margins, revenue per visit, and efficiency projections can be customized using institutional data or published benchmarks adjusted for inflation.

The 2-scenario approach (capacity-constrained vs volume-stable) allows decision-makers to model their specific operational context. Most US EDs operate in capacity-constrained environments where demand exceeds supply [1,2], making scenario 1 widely applicable. However, EDs in lower-demand settings or those with excess capacity may better align with the volume-stable scenario 2, where value derives primarily from modifiable cost reductions rather than volume increases. The model accommodates both scenarios and the full spectrum between them.

Overall, to support practical application of the economic model, Table S1 in Multimedia Appendix 1 provides published ED financial benchmarks with a user guide for selecting inputs appropriate to local cohort, payer mix, and ED type, and Table S2 in Multimedia Appendix 1 reports the underlying numerical values across the modeled efficiency range (−10% to +10% LOS change), allowing users to read off expected operating-margin and break-even outcomes at specific efficiency assumptions without rerunning the model. Furthermore, EDs assessing their position on the capacity-constrained–to–volume-stable spectrum may consider a combination of indicators—such as inpatient boarding burden, wait times, ambulance diversion, and rates of patients leaving prior to disposition—interpreted alongside local administrative knowledge, recognizing that no single metric reliably distinguishes the 2 scenarios. The −10% to +10% efficiency range modeled in our sensitivity analyses is intended to encompass the magnitude of LOS and throughput gains that may be associated with AI-based ED interventions in isolation.

This analysis relies on data from a single health system, though revenue estimates align with published national data. The economic estimates also depend on the attribution of volume and LOS changes to the AI triage CDS established in the parent intervention study [8]; to the extent that residual unmeasured confounding contributes to those observed changes, the financial estimates here would be commensurately affected. The analysis focused on ED-specific costs and does not capture downstream hospital financial impacts. A formal patient perspective was not modeled; future economic work should incorporate equity considerations and out-of-pocket cost effects to complement the hospital management and public policy perspectives presented here. The break-even analyses assume costs and benefits remain stable over time and do not account for potential learning curves, technology updates, or changing market conditions. How efficiency tools vary in value across different hospital ownership structures (for-profit, not-for-profit, and public), market competitiveness levels, and regulatory environments were not studied. Specifically, the model parameters reflect the US health care payment system; application to publicly funded or mixed systems (eg, Europe) would require substitution of locally appropriate values. Finally, while we demonstrate substantial financial value under realistic assumptions, individual hospital experience will vary based on local factors, including payer mix, labor markets, competitive dynamics, and operational practices. Future research should validate this framework across diverse health systems and settings with varying levels of capacity constraint, cost structures, ownership types, and market conditions.

This study demonstrates a pragmatic economic model for assessing ED efficiency interventions, applied to a real-world AI triage CDS implementation. The analysis illustrates how financial impacts differ substantially based on the cost-modeling framework. Traditional policy-level approaches substantially underestimate economic value from a hospital management perspective by not accounting for fixed and variable hospital cost structures. The model, publicly available for local adaptation, provides health care decision-makers with a practical tool that accounts for actual hospital cost behavior. As health care systems increasingly consider AI tools to improve efficiency and address ED overcrowding, economic evaluation frameworks that reflect operational realities are essential for informed adoption decisions.