Disparities in emergency department outcomes by race during the 2022–2023 COVID-influenza twindemic

Abstract

Background

The 2022–2023 “twindemic” of COVID-19 and influenza has worsened emergency department (ED) overcrowding, revealing racial/ethnic disparities in emergency care. This study assessed racial differences in ED operational outcomes—specifically, discharge ED length of stay (LOS), boarding time among admitted patients, and left without being seen (LWBS) rates—and characterized disparities concentrated in the upper tails of delay distributions.

Methods

This retrospective observational study included adult ED encounters with respiratory presentations and/or laboratory-confirmed COVID-19 or influenza across 10 EDs in Michigan between September 2022 and December 2023. The primary outcomes were ED LOS at discharge, boarding of admitted patients, and LWBS. Models adjusted for covariates and ED site (fixed effects) with site-clustered standard errors; for time outcomes, log-minutes (back-transformed). Tail burden was defined as exceeding the pooled 75th percentile and the upper Tukey fence. Multiple imputations (m = 20) were used to address missing data.

Results

Of the 6,476 encounters, the unadjusted median LOS to discharge was 193.0 min (mins) (Black), 177.0 min (White), and 172.0 min (‘Other’). Adjusted differences from White were − 1.8 min (95% CI, − 9.5 to 6.2) for Black and + 1.9 min (− 6.5 to 10.6) for ‘Other’; the boarding differences were + 33.8 min (− 16.5 to 91.8) and + 58.1 min (− 26.1 to 164.4), respectively. The LWBS rate was 5.7% (Black) vs. 2.3% (White); the adjusted odds ratio was 0.98 (0.57–1.68) and 0.79 (0.64–1.00) after imputation (AUC 0.81).

Conclusion

Median-based metrics obscured disparities in the upper tail of the institutions’ ED delays, specifically boarding time. Tail-aware metrics could augment standard dashboards by flagging uneven exposure to prolonged delays to drive equity-minded operational improvements.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12873-026-01508-9.

Introduction

Emergency departments (EDs) were under tremendous pressure from the “twindemic” wave of COVID-19 and the influenza surge during the 2022–2023 season, which exacerbated inequities in healthcare access and quality for racial/ethnic minority groups. Prior studies show that non-Hispanic Black, Hispanic, and Indigenous ED patients experience longer waits, less equitable triage, and higher leaving without being seen (LWBS) rates than non-Hispanic White patients [1, 2]. Structural racism, socioeconomic barriers, and implicit biases cause disparities by influencing clinical decision-making processes and resource allocation [3]. During the COVID-19 pandemic, existing disparities were exacerbated by overcrowding, staffing shortages, and multi-clinical demands [4, 5]. We focused on COVID-19 and influenza as high-volume, high-strain clinical contexts characterized by diagnostic overlap and resource saturation, conditions under which operational inequities are most likely to emerge.

The COVID-19 pandemic has disproportionately affected racial and ethnic minorities owing to structural factors, such as work-related exposure and crowded housing, which increased the rate of infection and mortality among these groups [6, 7]. Racial and ethnic minority populations during the 2009 H1N1 influenza pandemic reported by Quinn, Kumar [8] had restricted access to care and poorer health outcomes owing to structural issues. In 2022, EDs faced simultaneous outbreaks of COVID-19 and influenza, leading to overlapping symptoms that disrupted triage and strained diagnostic resources [9]. This interplay of factors has the potential to increase disparities, particularly during periods of high demand when diagnostic uncertainty and system strain are greatest [10]. These disparities may be influenced by factors such as hospital crowding, payer mix, staffing variability, implicit bias, and community-level access barriers.

Equitable healthcare access relies on ED performance metrics, specifically length of stay (LOS), boarding time, and LWBS rates. Black and Hispanic patients spend longer pre- and post-discharge LOS in the ED due to delayed diagnostics and inequitable resource allocation. These differences remained significant after adjusting for illness severity and insurance type [11]. Boarding time, measured as the duration between admission and ED departure, reflects system-level impediments to hospital bed availability and care coordination. The rates of LWBS are higher in racial/ethnic minority groups, and this relationship is augmented in resource-constrained hospitals. LWBS rates increase in hospitals that predominantly serve low-income patients [12].

Although awareness of healthcare disparities has grown in recent years, limited research has explored the role of race in shaping ED care delivery during twindemics. Existing studies have examined individual viruses, such as COVID-19, or data collected before the start of the pandemic, without accounting for the cumulative effects of concurrent epidemics [13, 14]. This study explores these gaps by analyzing disparities in LOS, boarding time, and LWBS rates across racial groups during the COVID-19/Influenza pandemic and characterizes disparities concentrated in the upper tails of delay distributions, revealing systemic inequities in overlapping public health emergencies.

Methods

Study design and setting

This retrospective observational study analyzed patient presentations at a tertiary health system with 10 EDs across four counties in southeastern and south-central Michigan from September 1, 2022, to December 18, 2023. The health system comprises five hospital-based EDs (HEDs) integrated within affiliated hospitals, offering a full continuum of emergency and specialty care, and collectively serving approximately 300,000 patient visits each year. Complementing these are five freestanding EDs (FEDs) that deliver advanced emergency care on campuses without inpatient facilities and manage approximately 140,000 visits annually. When necessary, patients from FEDs are transferred to HEDs within the system or to external hospitals for further treatment, as clinically indicated. The main differences between these two types of EDs are acuity, admission rates, and boarding capacity. Models include ED-site fixed effects with site-clustered standard errors to account for heterogeneity in practice within sites.

All 10 EDs serve racially/ethnically diverse populations. Generally, the HEDs in this study reported higher patient volume and acuity, significantly longer ED LOS, reduced throughput due to inpatient holding burdens, and higher LWBS rates than the FEDs. The study period coincided with the peak activity months of both COVID-19 (Omicron subvariants) and influenza (H3N2 strain), according to the CDC’s Respiratory Virus Surveillance Network [15].

Study population and cohort flow

All ED encounters during the study period involving individuals aged 18 years or older were included. Encounters were identified using a prespecified algorithm incorporating respiratory chief complaints and laboratory-confirmed COVID-19 and/or influenza. Ascertainment of acute respiratory cases (and/or ED-encounter laboratory confirmation of SARS-CoV-2/influenza) was developed in the data warehouse a priori and executed prior to analysis using a prespecified algorithm (chief-complaint lexicon; triage-note SNOMED groupers; International Classification of Diseases-10 J-chapter/viral pneumonia codes; ED-encounter PCR/antigen results). A CONSORT-style diagram details the adult denominator, sequential exclusions, and the final analytic N by race and admission status (Fig. 1).

Fig. 1.

Study flow chart of respiratory disease patient cohorts (COVID-19 and influenza patients)

Conceptual framework

Consistent with the framework of Boyd et al. [16], we conceptualized race as a sociopolitical marker shaped by systems of structural racism rather than a biological determinant. We evaluated racialized patterns in care delivery as manifestations of systemic inequities embedded within the healthcare infrastructure, staffing, and decision-making. This approach is consistent with the National Institutes of Health (NIH) and JAMA guidance on minimizing racial essentialism in research interpretation and presentation [17, 18].

Chart-review standards and data quality assurance

We adhered to recommended practices for medical record review in emergency medicine (i.e., prespecifying variables and rules; training abstractors with a written manual and test cases; conducting blinded dual abstraction on a 5% random sample and reporting Cohen’s κ for the “respiratory symptoms present” flag; adjudicating discrepant results). Key references include: [19–21].

Data sources

De-identified electronic health (EHR) data, including clinical, operational, and demographic variables, and standardized timestamps, were extracted in near real time from a centralized, harmonized data warehouse across sites. Variables included:

1. Patient demographics: Patient demographic information including self-reported race with categories of White, Black, and ‘Other’ (including American Indian/Alaska Native, Asian, Native Hawaiian or other Pacific Islander, multiracial, Middle Eastern, Hispanic, and patients who declined or did not know their race), age, sex, mode of arrival (emergency medical services [EMS] and non-EMS [walk-in, personal car, taxi, and public transportation]), and insurance type (private, Medicaid, Medicare, and uninsured).

2. Clinical data: Clinical data featured triage acuity levels based on Emergency Severity Index scores (ESI) from 1 to 5, final diagnoses verified by PCR tests for COVID-19 and influenza, and comorbidity ratings using the Charlson Comorbidity Index.

3. Operational metrics: Operational metrics included door-to-discharge time (LOS), boarding time (between admission decision and ED departure), and LWBS status, which served as our study outcomes.

Racial data were collected from patients at registration, following NIH standards for health disparities research [17]. Prior research has validated the operational metrics derived from EHR data [22].

Variables and definition

Primary predictor

The primary predictor for this analysis was race. The study followed NIH guidelines from 2022 to classify race using White as the baseline group [17]. Race was not used as a biological construct, but as a proxy for exposure to racism, in line with guidance from the AMA and Boyd, Lindo [16]. Because cell sizes for many racial and ethnic groups were small, the analysis collapsed them into White, Black, and ‘Other’ categories. We recognize that this heterogeneous ‘Other’ category aggregates diverse populations and may mask important differences.

Outcomes

These outcomes reflect the institutional processes through which structural racism can manifest via prolonged delays, crowding, and premature departure resulting from unmet needs. The study outcomes included: (1) ED LOS: A validated metric of ED efficiency is door-to-discharge time (minutes) for patients who leave the hospital from the ED [23]; (2) Boarding Time: The American College of Emergency Physicians [24] defines boarding time as the duration from admission decision to ED departure, measured in hours; and (3) LWBS: The National Hospital Ambulatory Medical Care Survey (NHAMCS) criteria identifies patients who walked out before seeing a doctor or departed against medical advice [25].

The primary outcomes were (1) discharged ED LOS (minutes [mins]), (2) boarding time among admitted patients (mins from admission decision to ED departure), and (3) LWBS (yes/no). Secondary outcomes were admitted ED LOS, subgroup contrasts by triage acuity (ESI), and tail-burden percentages for LOS and boarding defined as > 75th percentile (P75) and > Q3 + 1.5×Interquartile range (IQR) (upper Tukey fence).

Covariates

Patient-level covariates included age, sex (male/female), mode of arrival (EMS, yes/no), insurance status (Private/Medicare/Medicaid/Unknown), triage acuity (ESI 1–5), Charlson Comorbidity Index (CCI, 1–12), and ED site (fixed effects for each site). The CCI was computed from ICD-10 codes using the Quan adaptation with a 365-day look-back across the health system [26].

Statistical analysis

Continuous variables were summarized as medians (interquartile ranges, IQR), and categorical variables as counts (percentages, %). Groups were compared using the Kruskal-Wallis test and chi-squared (χ²) test. As the time outcomes were right-skewed, we log-transformed them. The coefficients were exponentiated and multiplied by 60 (to convert to minutes) at the geometric mean baseline to report the adjusted differences in minutes. Primary models were adjusted for prespecified covariates (age, sex, insurance status, mode of arrival, triage acuity, and CCI), including ED-site fixed effects, and site-level cluster-robust standard errors were used. For LWBS, we reported adjusted odds ratios (aORs) and 95% CIs, as well as model discrimination (area under the receiver operating characteristic curve; AUC).

We prespecified three primary outcomes and applied no multiplicity correction. For secondary outcomes, we controlled for the false discovery rate (Benjamini–Hochberg, q = 0.05). To provide context for robustness to unmeasured confounding, we reported the E-value, defined as the minimum association (on a risk ratio scale) that an unmeasured confounder would need to have both the exposure and outcome (conditional on the measured covariates) to explain the observed association.

Missing values for race/ethnicity and select covariates were imputed using multivariate imputation by chained equations (MICE; m = 20) under a missing-at-random assumption. We examined convergence diagnostics and Monte Carlo error and reported both complete-case and imputed analyses for the primary endpoints. All tests were two-sided with α = 0.05. Analyses were performed in Python (Python Software Foundation, Wilmington, DE, USA) version 3.12 for statistical evaluation.

Sensitivity analysis

Sensitivity analyses tested (i) “tail-burden” differences in LOS and boarding (risk of > 75th percentile, > upper Tukey fence Q3 + 1.5×IQR) via risk ratios, 95% CIs, and E-values and (ii) robustness to outcome missingness by comparing available-case estimates to those obtained after multiple imputation (MICE).

Ethical considerations and community accountability

To reduce stigmatization, we (1) framed race as a social exposure to structural racism rather than a biological factor, (2) avoided pathologizing language, and (3) reported absolute and relative differences with context [18]. Data were de-identified in accordance with the Health Insurance Portability and Accountability Act (HIPAA). While this retrospective analysis did not involve direct patient participation, the findings will be shared with health system equity leadership and community partners to inform institutional policies aimed at racial equity in ED operations.

Results

Baseline characteristics of the study population

A total of 6,476 adult ED respiratory encounters were included in the analysis (White, 3,138 [48.5%]; Black, 2,560 [39.5%]; ‘Other,’ 778 [12.0%]; Table 1), of which 96% were COVID-19 positive and 4% were influenza positive. Compared with White patients, Black patients were less likely to have arrived via EMS (10.3% vs. 20.5%), more likely to be insured by Medicaid (42.0% vs. 25.1%), and less likely to have Medicare (25.2% vs. 45.5%). Triage acuity was similar between groups (ESI 3 ≈ 58% in each group), but a higher proportion of White patients had a CCI ≥ 3 (19.7% vs. 12.2% for Black; 8.4% for ‘Other’).

Table 1.

Baseline characteristics of the study population by race

Characteristics	Race
	Overall (n = 6,476)	Black (n = 2,560)	Other (n = 778)	White (n = 3,138)
Age, median (IQR)	49.0 [32.0, 68.0]	42.0 [30.0, 59.0]	40.0 [27.2, 58.0]	58.0 [36.0, 76.0]
Sex, n (%)
Female	3,903 (60.3)	1,603 (62.6)	457 (58.7)	1,843 (58.7)
Male	2,573 (39.7)	957 (37.4)	321 (41.3)	1,295 (41.3)
Mode of arrival
EMS	958 (14.8)	263 (10.3)	53 (6.8)	642 (20.5)
Non-EMS	5,518 (85.2)	2,297 (89.7)	725 (93.2)	2,496 (79.5)
Insurance status
Medicaid	2,279 (35.2)	1,076 (42.0)	414 (53.2)	789 (25.1)
Medicare	2,228 (34.4)	645 (25.2)	154 (19.8)	1,429 (45.5)
Private	1,733 (26.8)	739 (28.9)	172 (22.1)	822 (26.2)
Uninsured	236 (3.6)	100 (3.9)	38 (4.9)	98 (3.1)
ESI acuity
1: Life-threatening	43 (0.7)	5 (0.2)	4 (0.5)	34 (1.1)
2: High-risk	1,302 (20.1)	546 (21.3)	130 (16.7)	626 (19.9)
3: Urgent	3,776 (58.3)	1,447 (56.5)	488 (62.7)	1,841 (58.7)
4: Less urgent	1,246 (19.2)	503 (19.6)	147 (18.9)	596 (19.0)
5: Non-urgent	86 (1.3)	44 (1.7)	6 (0.8)	36 (1.1)
CCI score
0	4557 (70.4)	1909 (74.6)	625 (80.3)	2023 (64.5)
1	609 (9.4)	205 (8.0)	65 (8.4)	339 (10.8)
2	189 (2.9)	59 (2.3)	11 (1.4)	119 (3.8)
>=3	995 (15.4)	313 (12.2)	65 (8.4)	617 (19.7)

Adult ED encounters with respiratory presentations during the COVID-19/influenza period

Abbreviations: EMS, emergency medical services; ESI, emergency severity index; CCI, Charlson comorbidity index

Note: CCI computed from ICD-10 (Quan adaptation) with 365-day look-back. Percentages use non-missing denominators. Data are presented as mean and standard deviation (SD), median and interquartile ranges (IQR), or n (%)

Primary outcomes

Discharged ED length of stay

Unadjusted median discharged LOS was 193.0 min (IQR 124.0–300.0) among Black patients, 177.0 min (116.0–275.0) among White patients, and 172.0 min (114.2–264.0) among ‘Other’ patients (Table 2). Adjusted differences (vs. White) were − 1.8 min (95% CI: −9.5 to 6.2) among Black patients and + 1.9 min (− 6.5 to 10.6) among ‘Other’ patients, after adjustment for age, sex, insurance, EMS arrival, ESI, CCI, and ED site. The distributions are presented in Fig. 2.

Table 2.

Primary outcomes by race: unadjusted and adjusted estimates

Outcome	Race	Unadjusted median (IQR), mins	OR vs. White (95% CI)	Adjusted difference vs. White, mins (95% CI)	Effect ratio†	aOR vs. White (95% CI)	AUC
Discharged ED LOS	White	177.0 [116.0, 275.0]	—	—	—	—	—
	Black	193.0 [124.0, 300.0]	—	-1.80 (-9.45 to 6.18)	0.99	—	—
	Other	172.0 [114.2, 264.0]	—	1.87 (-6.45 to 10.59)	1.00	—	—
Boarding (admitted)	White	283.0 [153.0, 641.8]	—	—	—	—	—
	Black	461.0 [263.0, 855.0]	—	33.81 (-16.47 to 91.76)	1.10	—	—
	Other	426.0 [238.0, 766.0]	—	58.09 (-26.08 to 164.4)	1.17	—	—
LWBS — N (%)	White	72 (2.3)	—	—	—	—	0.81
	Black	147 (5.7)	2.59 (1.94, 3.46)	—	—	0.98 (0.57, 1.68)
	Other	29 (3.7)	1.65 (1.06, 2.56)	—	—	0.99 (0.58, 1.68)

Abbreviations: ED LOS, emergency department length of stay; LWBS, left without being seen; AUC, area under the curve; IQR, interquartile range; mins, minutes; OR, odds ratio; aOR, adjusted odds ratio

^†Effect ratio is exp (β) from the log-minutes model (ratio of means on the geometric scale)

Note. The primary outcomes were discharge ED LOS, time-boarded hospital LOS, and LWBS. Adjusted models control for age, sex, insurance, EMS arrival, ESI, CCI, and ED-site fixed effects with site-clustered SEs. Time outcomes were modeled on log minutes and back-transformed to minutes at the geometric mean baseline

Fig. 2.

Discharged ED length of stay by race. Violin plots illustrate the distribution of LOS (in minutes). Solid vertical segments denote the interquartile range (IQR), with dots at the median. Dashed horizontal line marks the overall 75th percentile; dotted line marks the overall upper whisker (Q3 + 1.5×IQR). Numeric median and IQR values are annotated for each race

Boarding among admitted patients

Unadjusted boarding medians were more than twice as long for Black patients compared to White patients (461.0 vs. 283.0 min). Adjusted differences were + 33.8 min (95% CI: −16.5 to 91.8) among Black versus White patients and + 58.1 min (− 26.1 to 164.4) among ‘Other’ versus White patients (Table 2; Fig. 3).

Fig. 3.

Boarding time among admitted patients by race. Violin plots as above. Tail-burden statistics (% above P75 and % above upper whisker) are reported in Table 4

Left without being seen (LWBS)

LWBS occurred in 5.7% of Black patients, 3.7% of ‘Other’ patients, and 2.3% of White patients (Table 2). Unadjusted ORs were 2.59 (95% CI 1.94–3.46) for Black vs. White and 1.65 (95% CI 1.06–2.56) for ‘Other’ vs. White. In adjusted models, aORs were 0.98 (95% CI 0.57–1.68) for Black versus White and 0.99 (95% CI 0.58–1.68) for ‘Other’ versus White; model discrimination for LWBS was 0.81 (Fig. 4).

Fig. 4.

Left-without-being-seen (LWBS) model performance. Receiver operating characteristic (ROC) curve for the multivariable LWBS model; AUC shown with 95% CI (bootstrap). Covariates: age, sex, insurance, EMS, ESI, CCI; ED-site fixed effects; cluster-robust SEs by site

Secondary outcomes

Admitted ED length of stay and ESI stratified analysis

For admitted ED LOS (arrival-to-ED departure among admitted), the unadjusted gap was substantial (median 856.0 min Black vs. 559.5 min White), but the adjusted difference was + 50.8 min (95% CI − 15.1 to 123.2) for Black vs. White and + 69.5 min (− 46.5 to 206.8) for ‘Other’ vs. White (Table 3, Supplementary Fig. S1). ESI-stratified analyses of discharged LOS (Table S1) yielded effect estimates with varying directions (e.g., ESI 2: +15.9 min, ratio 1.05; ESI 5: −12.7 min, ratio 0.88), and none of the comparisons remained significant after FDR adjustment (all q ≥ 0.32).

Table 3.

Secondary outcome by race: unadjusted and adjusted estimates

Outcome	Race	Unadjusted median (IQR), mins	OR vs. White (95% CI)	Adjusted difference vs. White, mins (95% CI)	Effect ratio†
Admitted ED LOS	White	559.5 [390.0, 965.0]	—	—	—
	Black	856.0 [552.2, 1288.5]	—	50.75 (-15.14 to 123.22)	1.08
	Other	663.0 [480.0, 1100.0]	—	69.46 (-46.50 to 206.78)	1.10

Abbreviations: ED LOS, emergency department length of stay; IQR, interquartile range; mins, minutes; OR, odds ratio; aOR, adjusted odds ratio

^†Effect ratio is exp (β) from the log-minutes model (ratio of means on the geometric scale)

Note. One of the secondary outcomes was admitted ED LOS. Adjusted models control for age, sex, insurance, EMS arrival, ESI, CCI, and ED-site fixed effects with site-clustered SEs. Time outcomes were modeled on log minutes and back-transformed to minutes at the geometric mean baseline

Tail-burden analyses

Racial disparities focused on the upper tails of the distribution (Table 4). For discharged LOS, Black vs. White risk ratios (RR) were 1.23 (95% CI 1.11–1.36) for > P75 and 2.80 (2.14–3.68) for > upper Tukey fence (Q3 + 1.5×IQR); corresponding E-values were 1.76 and 5.05, respectively. For boarding, Black vs. White RRs were 1.39 (1.08–1.79) for > P75 and 2.49 (1.22–5.10) for > upper-whisker (E-values 2.12 and 4.42). In a complementary description of admitted ED LOS, 34.2% of Black patients and 22.0% of White patients exceeded the pooled P75 (1087 min), and 4.7% of Black patients and 2.6% of White patients exceeded the upper whisker (2101 min) (Supplementary Table S2).

Table 4.

“Tail-burden” of ED delays by race

Metric	Race	Threshold type	Threshold (mins)	RR vs. White (95% CI)	E-value
Discharged ED LOS	Black	> P75	283.0	1.23 (1.11–1.36)	1.76
	Black	> Upper whisker	529.0	2.80 (2.14–3.68)	5.05
	Other	> P75	283.0	0.95 (2.14–3.68)	1.31
	Other	> Upper whisker	529.0	1.30 (2.14–3.68)	1.93
Boarding (admitted)	Black	> P75	745.0	1.39 (1.08–1.79)	2.12
	Black	> Upper whisker	1601.5	2.49 (1.22–5.10)	4.42
	Other	> P75	745.0	1.19 (0.74–1.92)	1.67
	Other	> Upper whisker	1601.5	2.68 (0.88–8.20)	4.81

Abbreviation: P75, > 75th percentile; Upper whisker, > Q3 + 1.5×IQR (upper Tukey fence); RR, risk ratio; min, minutes

Notes: P75 and upper-whisker thresholds calculated from the pooled distribution for each outcome. RR computed with continuity correction if needed

Sensitivity analysis

The results of the multiple imputation (m = 20) analyses were generally similar to those of the complete-case analyses (Supplementary Table S3). The adjusted difference between Black and White races for discharged LOS was + 4.4 min (95% CI, − 5.6 to 14.9) after imputation, versus − 1.8 min in complete cases. For boarding, the imputed difference was + 45.2 min (0.06 to 96.2), versus + 33.8 min for the complete-case analysis. The association of LWBS with race attenuated toward the null with imputation (aOR 0.79, 95% CI 0.64–1.00, versus 0.98 complete-case).

Discussion

In this multi-site cohort of 6,476 adult ED encounters during the COVID-19/influenza “twindemic,” we observed that although central-tendency differences in discharged ED LOS by race were small after adjustment, marked, clinically meaningful disparities emerged in the upper tails of the distributions, especially for boarding among admitted patients. Black patients were 23% more likely than White patients to exceed the pooled 75th percentile of discharged LOS and nearly three times as likely to exceed the upper Tukey fence; analogous tail-burden gaps were present for boarding. In contrast, the adjusted association between Black race and LWBS attenuated toward the null with multiple imputation. These patterns are more consistent with intermittent but severe delays during periods of system stress than with small, uniform shifts in the average care time.

Our results contribute to a large body of literature showing racialized differences in ED care that are reflected in longer wait times, prolonged throughput, and higher LWBS rates for Black and ‘Other’ minoritized patients after adjustment for acuity and insurance status [2, 27–29]. In high-pressure ED settings during peak pandemic surges, crowding, stretched staffing resources, and constrained diagnostic capacity increased the stakes around throughput. Although our analysis cannot assess individual decision-making, structural factors and implicit biases in bed allocation or consult availability may contribute to differential exposure to prolonged delays and warrant further investigation. Prior work has cautioned that these conditions exacerbate (rather than create) inequities [10]. In this study, we provide a tail-focused lens for this discussion; medians and IQRs may have seemed comparable across groups, but the likelihood of extreme delay was not. This finding is consistent with prior work that highlights the uneven distribution of overcrowding, which undermines the association between hospital-level processes and racial inequities in flow and boarding [30–32].

Our ESI-stratified secondary analyses indicated that the disparities in ED flow were not restricted to a single acuity level; point estimates differed by ESI category, and none remained significant after FDR correction. This finding is consistent with the system-level (rather than purely patient-level) determinants. The E-value for the largest tail effects (e.g., ≥ 5 for the discharged LOS upper whisker) suggests that an unmeasured confounder would need to be strongly associated with both race and extreme delay (conditional on measured covariates) to explain the observed disparities [33].

Structural racism is an analytical framework for centering race as a social determinant throughout our healthcare systems and policies. Our results may be shaped by a variety of drivers along the care continuum that differentially cluster in the right tail. Bed management and boarding practices (including remote/digital bed assignment and informal patient prioritization rules) could introduce “virtual” or algorithmic bias when inpatient capacity is constrained. Small variations in downstream consult and imaging queues can be amplified by crowding pressures and shift encounters into the tail, and differences in ED coverage and care coordination (e.g., insurance mix, EMS arrival mode, outpatient access) could differentially slow handoffs and disposition while also mapping onto axes of structural disadvantage [12, 34, 35]. Implicit bias can also operate at the micro level to introduce small but pervasive differences in triage reassessment, symptom interpretation, and the prioritization of non-visible care tasks, which can lengthen stays within overlapping distributions [36, 37]. The attenuation of the LWBS association after multiple imputations suggests confounding by missingness and site-level practice heterogeneity with the EHR data [38], but the larger unadjusted LWBS gaps are also policy-relevant as sentinel markers of access and experience inequities in the ED setting [39].

The county-level socioeconomic context also plausibly explains our extreme-tail delays, as the 10 EDs are distributed across very different median household income counties (e.g., lower in Wayne vs. higher in Oakland), suggesting heterogeneity in payer mix, primary-care availability, transportation access, and hospital bed capacity [40–42]. In lower-income catchments, higher proportions of Medicaid/uninsured patients and fewer outpatient alternatives can fuel consult/diagnostic demand and delay inpatient placement, leading to more encounters exceeding the pooled P75 and upper-whisker benchmarks, even when medians are similar [43]. Such community factors mediate ED-level workflows (bed assignment, consult acceptance, patient handoffs) and implicit/structural bias to produce episodic, high-magnitude delays disparately impacting Black patients, in line with tail-burden findings for discharged LOS and boarding [35, 44]. County income may not perfectly capture individual-level socioeconomic status, but could enhance equity dashboards (when added to race-stratified tail measures) to steer targeted resources (surge beds, hospitalist coverage, case management) to facilities and neighborhoods that are most at risk for extreme delay [45, 46].

Clinically and operationally, focusing on means or medians can obscure inequity residing in tails; thus, EDs should (i) supplement routine dashboards with tail-aware equity metrics (e.g., % of encounters > P75 and > Q3 + 1.5×IQR, by race/ethnicity and insurance); (ii) conduct boarding equity audits with real-time escalation triggers to track time-to-bed, consult-to-decision, and ED-to-inpatient handoff times and detect widening of race-stratified tail gaps; and (iii) implement equity-by-design changes (transparent auditable bed-assignment rules, standardized consult SLAs, implicit-bias–aware triage/reassessment, and partnerships with clinical and community organizations to extend outpatient follow-up and community respiratory care as needed to decompress EDs during surges). These recommendations are tailored to evidence that crowding and boarding impact Black and other minoritized patients disproportionately and can be driven by hospital-level processes and implicit bias, as well as patient mix [2, 47, 48], and are consistent with contemporary guidance on the responsible use and reporting of race and ethnicity in clinical operations [18].

This study used a 10-ED integrated system with a single EHR and a prespecified respiratory encounter ascertainment. We adhered to EM chart-review standards [19–21] and site fixed effects with cluster-robust SEs to maximize the isolation of within-system inequities. We report the central and tail metrics, FDR for multiple secondary contrasts, and multiple imputation to improve the robustness against missing data. This study has several limitations. First, its retrospective design is vulnerable to misclassification of symptom codes and operational timestamps. In particular, the absence of reliable door-to-provider times across sites may have introduced non-differential measurement error, which would be expected to bias estimates toward the null. Although standardized rules and sensitivity analyses were applied, residual error cannot be excluded. Second, despite adjustment for multiple patient- and site-level factors, residual confounding by unmeasured structural determinants—including inpatient bed availability, staffing ratios, consult response times, primary care attachment, and social needs—likely remains and may underlie the extreme delays captured by tail-based metrics. Even with site-fixed effects, differences in patient mix and acuity between HEDs and FEDs may have contributed to observed tail delays. Third, the absence of encounter-level crowding measures limits adjustment for dynamic operational pressures, and we did not assess whether facilities serving higher proportions of Black patients experienced systematically longer delays, an important target for future equity audits. Fourth, as a single-system study during a time of increased respiratory viral burden, our findings may not generalize to other settings or non-surge periods and could be affected by selection bias due to differential testing practices. Finally, our multiple imputation assumed that data were missing at random; imputation produced imprecise estimates for low-frequency outcomes (e.g., LWBS), and aggregating smaller racial and ethnic groups may have masked heterogeneity within groups. Tail thresholds were determined based on pooled distributions; alternative definitions may affect effect sizes, but are unlikely to change qualitative conclusions.

Conclusion

Racial disparities in ED care during the COVID-19/influenza twindemic were driven by a concentration of inequities in the upper tail of the institutions’ ED delays. The median time masked disparity. Tail-aware metrics identify system-level, racially skewed bottlenecks that are remediable through bed management reform, standardized consultation workflows, and equity-by-design operations. Optimizing these processes is essential for delivering timely and equitable emergency care.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ACEP

American College of Emergency Physicians

AIC/ICD

International Classification of Diseases

AMA

American Medical Association

AOR/aOR

Adjusted Odds Ratio

AUC

Area Under the Curve (Receiver Operating Characteristic)

CCI

Charlson Comorbidity Index

CDC

Centers for Disease Control and Prevention

CI

Confidence Interval

COVID-19

Coronavirus Disease 2019

CONSORT

Consolidated Standards of Reporting Trials

Δ(Delta)

Difference

ED

Emergency Department

EMS

Emergency Medical Services

ESI

Emergency Severity Index (triage acuity score)

E-value

Strength-of-confounding metric used in sensitivity analysis

FED

Freestanding Emergency Department

FDR

False Discovery Rate

HED

Hospital-Based Emergency Department

HIPAA

Health Insurance Portability and Accountability Act

H1N1

Influenza A subtype H1N1

ICD-10

International Classification of Diseases, 10th Revision

IQR

Interquartile Range

LOS

Length of Stay

LWBS

Left Without Being Seen

m

Number of imputed datasets (in MICE)

MICE

Multivariate Imputation by Chained Equations

NIH

National Institutes of Health

OR

Odds Ratio

PCR

Polymerase Chain Reaction

P75

75th Percentile Threshold

Q3

Third Quartile (75th percentile)

q-value

False discovery rate–adjusted p-value

RR

Risk Ratio

ROC

Receiver Operating Characteristic

RSV (contextual)

Respiratory Syncytial Virus

SD

Standard Deviation

SE

Standard Error

SES

Socioeconomic Status

SLAs

Service-Level Agreements (consult workflows)

SNOMED

Systematized Nomenclature of Medicine

Tukey fence

Upper whisker threshold (Q3 + 1.5×IQR)

TWINDEMIC

Concurrent COVID-19 and influenza surges

US

United States

URI (contextual)

Upper Respiratory Infection (referenced indirectly)

WHO

World Health Organization

Author contributions

J. E. and J. M. conceived and designed the study methodology. J. E., E-E. E., and J. M. supervised the data collection. E-E. E., C. D., and K. E. managed the data, including quality control. J. L., I. T., and S. M. provided statistical advice on study design and analyzed the data; J. M. and S. G. chaired the data oversight committee. J. E. and S. M. drafted the manuscript, and all authors read, contributed substantially, reviewed, and approved the final manuscript.

Funding

No funding or grants were received for this study.

Data availability

The electronic health record data supporting this study were obtained from the participating health system’s centralized data warehouse. Restrictions apply to the availability of these data, which were accessed under an institutional license and used with approval from the health system and the Institutional Review Board. As such, the data are not publicly available. De-identified data may be made available upon reasonable request and with permission from the health system’s data governance committee and Institutional Review Board approval.

Declarations

Ethics approval and consent to participate

This minimal-risk study was approved by Henry Ford Health System Institutional Review Board, Detroit, Michigan, USA, which waived the requirement of informed consent. All procedures were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments. No identifiable patient information was accessed, and all data were handled in compliance with institutional policies and applicable privacy regulations.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.