Defining Incidental Versus Non-incidental COVID-19 Hospitalizations

Background Rates of COVID-19 hospitalization are an important measure of the health system burden of severe COVID-19 disease and have been closely followed throughout the pandemic. The highly transmittable, but often less severe, Omicron COVID-19 variant has led to an increase in hospitalizations with incidental COVID-19 diagnoses where COVID-19 is not the primary reason for admission. There is a strong public health need for a measure that is implementable at low cost with standard electronic health record (EHR) datasets that can separate these incidental hospitalizations from non-incidental hospitalizations where COVID-19 is the primary cause or an important contributor. Two crude metrics are in common use. The first uses in-hospital administration of dexamethasone as a marker of non-incidental COVID-19 hospitalizations. The second, used by the United States (US) CDC, relies on a limited set of COVID-19-related diagnoses (i.e., respiratory failure, pneumonia). Both measures likely undercount non-incidental COVID-19 hospitalizations. We therefore developed an improved EHR-based measure that is better able to capture the full range of COVID-19 hospitalizations. Methods We conducted a retrospective study of ED visit data from a national emergency medicine group from April 2020 to August 2023. We assessed the CDC approach, the dexamethasone-based measure, and alternative approaches that rely on co-diagnoses likely to be related to COVID-19, to determine the proportion of non-incidental COVID-19 hospitalizations. Results Of the 153,325 patients diagnosed with COVID-19 at 112 general EDs in 17 US states, and admitted or transferred, our preferred measure classified 108,243 (70.6%) as non-incidental, compared to 71,066 (46.3%) using the dexamethasone measure and 77,399 (50.5%) using the CDC measure. Conclusions Identifying non-incidental COVID-19 hospitalizations using ED administration of dexamethasone or the CDC measure provides substantially lower estimates than our preferred measure.


Introduction Background
Since the start of the Coronavirus disease -2019 (COVID-19) pandemic, national and local governments and healthcare organizations have struggled to monitor the pandemic's severity and the degree of burden on healthcare resources [1].In an era of widespread at-home testing [2], infection counts have become increasingly unreliable and are often not reported [3][4][5][6][7].Therefore, COVID-19 hospitalizations remain an important measure of the COVID-19 burden.However, starting in December 2021, the dominance of the highly infectious but often less severe Omicron variant led to increasing numbers of hospitalized patients who tested positive for COVID-19 but were admitted primarily for other reasons [8,9].This rise in hospitalizations of patients with incidental COVID-19 infections blurs the signal provided by hospitalization counts for the COVID-19 burden on the population and hospitals.

Importance
Despite the importance of hospitalizations as a measure of the public health burden of COVID-19, there is no good measure that can reliably distinguish between cases where COVID-19 is the primary reason for hospitalization (non-incidental admission) and cases where COVID-19 is incidental to the hospitalization.As we discuss below, the two principal measures in current use undercount non-incidental hospitalizations, especially during the Omicron period that began in December 2021.
One approach, developed by the Commonwealth of Massachusetts, uses the in-hospital administration of dexamethasone as a proxy for non-incidental COVID-19 hospitalizations [10][11].Dexamethasone is a steroid medication indicated for patients with COVID-19-associated acute respiratory failure with hypoxemia [12].Yet in the Omicron era, patients hospitalized with non-incidental COVID-19 are less likely to have severe respiratory failure with hypoxemia [13,14].In our experience as ED clinicians, many COVID-19-positive patients, especially older patients, are admitted to hospitals (or transferred for admission) with COVID-19 symptoms and complications but without an indication for dexamethasone.Similarly, some patients may have a contraindication to receiving dexamethasone, such as an allergy.Therefore, relying solely on dexamethasone to count non-incidental COVID-19 hospitalizations likely lacks sensitivity as a measure of which patients with COVID-19 are hospitalized primarily for their COVID-19 infection.The dexamethasone proxy also has limited specificity because dexamethasone may be prescribed for conditions not due to COVID-19 infection.For example, dexamethasone may be administered for allergic reactions, pretreatment for IV contrast administration to allergic patients, or adrenal insufficiency.
The second method in common use to identify non-incidental COVID-19 hospitalizations relies on a limited number of co-diagnoses identified by the CDC as being related to COVID-19 [15,16].The CDC has reported several iterations of these highly related co-diagnoses [16].During the early pandemic, the CDC developed a COVID-19 hospitalization dashboard which indicated how often COVID-19-related discharge diagnoses included co-diagnoses for acute respiratory distress syndrome, acute respiratory failure, pneumonia, sepsis, and acute renal failure/kidney injury [15,16].Since these diagnoses were available at the conception of our study, we modeled our CDC definition on these diagnoses.Nevertheless, all iterations of diagnoses used by the CDC to identify non-incidental COVID-19 hospitalizations rely heavily on a co-diagnosis of acute respiratory failure [15,16].Thus, the definition is highly correlated with the dexamethasone measure and has similar shortcomings.Therefore, the CDC approach also likely undercounts non-incidental COVID-19 hospitalizations, especially during the Omicron period.

Study objective
We sought to develop an improved method for measuring the proportion of COVID-19-positive patients admitted/transferred from the ED for non-incidental COVID-19, which could be applied using diagnoses that should be available either from the ED visit alone (as we use it), from the hospitalization alone, or from both together.Conceptually, these and other COVID-19-related diagnoses (e.g., acute metabolic encephalopathy) would often also be present at the time of admission to the ED [17].We hypothesized that relying on a set of co-diagnoses that are strongly related to COVID-19 infection, but a broader set than the CDC has used, can provide a reasonable way to identify non-incidental COVID-19 hospitalizations.Conversely, we did not view dexamethasone administration as adding materially to the information already available for a patient with COVID-19 and a diagnosis of respiratory failure with hypoxemia, the indication for which dexamethasone is recommended.

Practical consideration: need for a measure that can be simply and cheaply used at scale
Prior studies using manual chart review have estimated non-incidental COVID-19 admissions to be between 55-88% of all hospitalizations of COVID-19-positive patients pre-Omicron and 69% after the onset of Omicron [18][19][20][21][22].However, manual chart reviews are impracticable at scale.The public health need is for a simpler measure that can be implemented relying on information already captured in standard EHRs.We therefore present an approach to measuring the approximate incidence of non-incidental COVID-19 hospitalizations using COVID-19-related co-diagnose codes.

Study design, setting, and patients
We conducted a retrospective study of billing and visit data from a national emergency medicine physician group, in which we studied patients who visited the ED, what proportion were diagnosed with COVID-19, of those, what proportion were either admitted or transferred to another facility, and the co-diagnoses for these patients.Transfers were counted together with admissions because admission to another hospital is the most common reason for transfer from the ED to another facility [23].Below we refer to admitted or transferred patients simply as admitted.The dataset has been described previously [24,25].Briefly, charts are reviewed by billing and coding specialists shortly after the ED visit.These specialists have ongoing training and undergo regular quality and compliance audits to ensure accurate coding of diagnoses using International Classification of Disease, Tenth Revision (ICD-10) codes.Visit information, including primary and secondary diagnoses, procedures, and medications administered during the ED stay are stored in a deidentified research dataset.Our final study dataset included all visits to 112 non-pediatric, non-freestanding EDs (in 17 states) with data for the full study period from April 1, 2020, to August 31, 2023.
We collected ED visit dates, medication orders for dexamethasone, primary ED diagnoses, secondary ED diagnoses (up to two), and ED disposition (i.e., admitted, transferred, discharged, left without being seen) for each encounter in the study period.Since dexamethasone administration is used by others to measure non-incidental COVID-19 hospitalizations [10][11], visits with missing medication data (2.6% of all visits) were excluded.We selected patients with ED-diagnosed COVID-19 using the Agency for Healthcare Research and Quality Clinical Classification Software Refined (CCSR) code INF012 [26].We examined the proportions of admitted COVID-19 patients hospitalized specifically for symptoms or complications of COVID-19 based on various definitions, including the dexamethasone definition, the CDC definition, and the definition we developed and how these proportions varied over the study period.The Allegheny Health Network Institutional Review Board approved secondary analyses of this de-identified research dataset.

Variables
We compared several approaches for determining which patients were admitted for non-incidental COVID-19 versus those admitted with an incidental COVID-19 infection.We considered a dexamethasone-only approach, using dexamethasone administration to identify non-incidental COVID-19 hospitalizations [11].We also considered the CDC approach, classifying COVID-19 hospitalizations with concomitant acute renal failure/acute kidney injury, acute respiratory distress syndrome, acute respiratory failure, pneumonia, and sepsis as non-incidental [15].However, both the dexamethasone-based and CDC definitions will, based on our clinical experience, miss admissions for conditions that are often consequences of, or associated with, a COVID-19 infection, for example, those with acute metabolic encephalopathy or acute pulmonary embolism [17,27,28].
We therefore sought to develop a more sensitive definition based on co-diagnoses that are likely to indicate a non-incidental admission.As a basis for choosing a set of co-diagnoses that are likely to indicate a nonincidental admission, we reviewed all Agency for Healthcare Research and Quality (AHRQ) Clinical Classifications Software Refined (CCSR) codes (N=543, excluding the INF012 CCSR code for COVID-19) and divided them into categories signifying the strength of their association with COVID-19 [26].Two boardcertified ED-physician authors (JO and DN) independently marked each CCSR code "Maybe Related" or "Not Related" with a third ED-physician author (JP) resolving disagreements.Then, three authors (JO, NR, and DN) independently marked all the "Maybe Related" CCSR codes as "High Likelihood," "Medium Likelihood," or "Low Likelihood."Agreement between two of the three raters decided the final classification, except for a small number of codes (N=4) for which all three raters disagreed, which we considered "Medium Likelihood."Classifications and details on interrater reliability are presented in the appendix.We classified 23 CCSR codes as High Likelihood, 34 as Medium Likelihood, and 14 as Low Likelihood.The remaining 472 CCSR codes were considered Not Related.
We then used these categories to identify COVID-19-related diagnoses and define approaches to measuring non-incidental COVID-19 admissions.The approaches included: 1) an upper bound approach including all high and medium likelihood diagnoses, 2) a diagnosis-based approach including all high likelihood diagnoses, 3) a mixed, high-likelihood diagnosis or dexamethasone approach, including all patients with high likelihood diagnoses or with dexamethasone ordered in the ED, 4) a CDC-based approach adopted on the CDC's list of COVID-19 associated diagnoses [15], which corresponds to the CCSR categories for pneumonia, septicemia, respiratory failure, or acute renal failure, and 5) a dexamethasone-only approach where dexamethasone was ordered in the ED.ED-diagnosed COVID-19 patients admitted without any secondary diagnoses were counted as non-incidental COVID-19 admissions for all approaches except dexamethasone-only.
Recall, however, that dexamethasone for the treatment of COVID-19 is indicated when acute respiratory failure with hypoxemia is present [12].The diagnosis-based approach, which includes all diagnosis variants of respiratory failure (Table 1), should capture all of these cases.Therefore, there should be little difference between approach 2 (high-likelihood diagnoses only) and approach 3 (high-likelihood or dexamethasone).Moreover, any admissions captured by approach 3 but not approach 2 would have to involve cases where dexamethasone was administered in the absence of the indication for its use for COVID-19 patients.Thus, these incremental cases can be considered false positives, where dexamethasone was administered for an indication other than acute respiratory failure with hypoxemia.

Upper bound
The patient had a diagnosis categorized as Maybe COVID-Related -High or Medium Likelihood.

Diagnosis-based or dexamethasone
The patient had a diagnosis categorized as Maybe Related -High Likelihood OR a dexamethasone order.

Diagnosis-based
The patient had a diagnosis categorized as Maybe Related -High Likelihood.

CDC-based
The patient had a diagnosis under the following CCSR codes: RSP002, INF002, RSP012, or GEN002.

Dexamethasone-Only
The patient had a dexamethasone order.

Outcomes
The principal study outcomes were the proportion of all admitted COVID-19 ED patients having nonincidental COVID-19 classified using the defined approaches.Since the severity of COVID-19 illness and the dominant virus variants have changed over time [7,13,14], the proportions of non-incidental admissions were plotted by month over the study period to examine trends.

Analysis
Categorical and continuous variables are presented as counts with percentages and means with standard deviations (SD).Since the Omicron variant of COVID-19 often causes less severe respiratory illness [13,14], we studied the five approaches separately before and after the onset of Omicron.The pre-Omicron (April

Results
Of the 12,569,528 ED visits to 112 facilities in 17 US states during the study period, 323,971 visits were excluded for missing medication data (Figure 1, Table 2).Of the remaining visits, 468,754 visits included a positive COVID-19 diagnosis; of these, 153,325 were admitted (Figure 1).The mean age in years of admitted COVID-19 positive patients during the study period was 64 (SD 18.6), including 62 (17.8) pre-Omicron and 67 (19.5) after the onset of Omicron (Table 3).Co-diagnoses of pneumonia decreased from 39.5% to 17.9% and respiratory failure from 20.1% to 14.7% between the pre-Omicron and Omicron periods (Table 3).
Self-pay 11,832 The proportion of ED-diagnosed COVID-19 patients admitted decreased over the study period (Figure 2).The proportion of patients admitted with non-incidental COVID-19 was roughly flat during the pre-Omicron period for all approaches except the dexamethasone-only approach and fell with all approaches at the onset of the Omicron period and then flattened out again (Figure 3).The proportion of non-incidental COVID-19 admissions varied greatly between measurement approaches (Figure 3 and Table 4).The proportion of patients admitted for COVID-19 over the study period was as high as 89.0% using the upper bound approach (averaged over the sample period) or as low as 50.5% with the dexamethasone-only approach (Figure 3 and Table 4).In 2020, before dexamethasone became the standard of care for COVID-19associated acute respiratory failure with hypoxemia, the CDC-based approach estimated more nonincidental COVID-19 hospitalizations than the dexamethasone-only approach.However, in 2021 and 2022, the two approaches yielded similar estimates (Figure 3 and Table 4).The number of probable false positive cases captured by the dexamethasone-only definition but not by the high-likelihood diagnosis approach was estimated to be 4,773 of 97,323 (4.9%) for the pre-Omicron period, 3,439 of 55,940 (6.1%) for the Omicron period, and 8,212 of 153,263 (5.4%) over the study period (Table 4).The use of dexamethasone decreased over the study period across multiple subgroups.For example, the percentage of patients receiving dexamethasone in the ED decreased between the pre-Omicron and Omicron periods from 65.7% to 57.0% for those diagnosed with pneumonia, 49.1% to 33.9% with sepsis, 74.7% to 67.8% with respiratory failure, and 42.0% to 27.1% with acute renal failure (Table 5).The lower dexamethasone-only proportions, relative to the co-diagnosis approaches, are unlikely to be due to differences between the ED-based counts we used and the in-hospital counts used by Massachusetts.From January 10, 2022, to September 27, 2022, levels and time trends were similar for our approach versus Massachusetts' publicly reported data (Figure 4) [30].

Pre-Omicron Omicron Overall
Received

Discussion
In our multisite study, the diagnosis-based approach developed by our study team to distinguish nonincidental COVID-19 hospitalizations from incidental ones estimated many more non-incidental COVID-19 hospitalizations than the CDC-based approach or the dexamethasone approach (Figure 3 and Table 4), and a proportion of non-incidental admissions similar to estimates based on chart reviews, as discussed below.
Our clinical assessment is that the diagnosis-based approach yields a more accurate estimate of nonincidental COVID-19 hospitalizations than the CDC-based or dexamethasone approaches.In particular, in our experience, patients hospitalized for COVID-19 can have various non-pulmonary sequelae, and many do not meet the clinical indications for dexamethasone administration.Therefore, the CDC-based and dexamethasone approaches, which yield similar results, are likely to substantially undercount the number of non-incidental COVID-19 hospitalizations (Table 4), especially given the reduced incidence and often reduced severity of pneumonia and respiratory failure after the onset of Omicron (Table 4 and Table 5) [13,14].This undercounting would underestimate the actual COVID-19 related burden on the population, and on hospitals and health systems.
However, the upper bound approach may overcount hospitalizations.This potential overcounting is illustrated by the less drastic drop in non-incidental hospitalizations at the onset of Omicron (Figure 3).The more gradual decline with the upper bound approach compared to the other approaches can be explained by more COVID-19 infected patients being hospitalized with less-related co-diagnoses (Table 1).
We also showed that while the dexamethasone approach undercounts non-incidental hospitalizations, it also counts a substantial number of false positives.We estimated that 5.4% of admitted COVID-19 ED patients could be falsely classified as non-incidental by the dexamethasone approach during the study period (Table 4).This finding was expected given that dexamethasone has many indications other than respiratory failure with hypoxemia due to COVID-19 infection (e.g., allergic reactions).
Additionally, despite having similar estimates and trends (Figure 3 and Table 4), we observed evidence that the dexamethasone and CDC-based approaches likely count different patients.For example, in the study period, only a fraction of patients with pneumonia (63.9%), sepsis (42.9%), respiratory failure (72.7%), and acute renal failure (35.6%) received dexamethasone during the study period (Table 5).If the dexamethasone and CDC-based approaches counted the same patients, these percentages would more closely approximate 100% given the similar estimates both approaches provide for non-incidental COVID-19 hospitalizations.
Our results align with prior studies estimating the proportion of non-incidental COVID-  [18,[20][21][22].These results align with our diagnosis-based approach, which estimates non-incidental COVID-19 hospitalizations at 70.6% on average over the sample period although generally below 60% from April 2022 through the end of the sample period in August 2023 (Figure 3 and Table 4).Our approach was also sensitive enough to capture an apparent uptick in the proportion of non-incidental admissions from around 55% to 60% in April 2023, coinciding with the expiration of the US public health emergency authorizations.Nevertheless, our diagnosis-based approach can be implemented at scale without labor-intensive chart review.
We also observed a similar proportion of non-incidental COVID-19 hospitalizations using our ED-based dexamethasone approach as was reported by Massachusetts using hospitalization data in the Omicron period (Figure 4) [30].This suggests that the difference between measuring ED-based administration of dexamethasone (our study) and hospital-based administration (Massachusetts reporting), does not explain the lower rates of non-incidental COVID-19 hospitalizations we found using the dexamethasone approach.
Systematic undercounting of non-incidental COVID-19 hospitalizations can have important public health implications.First, COVID-19 hospitalization estimates are used to make public health and policy decisions impacting large segments of the population.These health and policy decisions, such as implementing mask and vaccine mandates, may have economic and educational consequences (e.g., reduced employment rates, reduced child test scores, etc.) [31].In addition, the estimates may change individuals' behaviors by affecting their perception of the COVID-19 risk they face.Therefore, since dexamethasone has a limited indication for acute respiratory failure with hypoxemia and dexamethasone use has been decreasing with the Omicron variant (Table 5), we believe that better methods to estimate the proportions of incidental and nonincidental COVID-19 hospitalizations are available.The CDC approach performed similarly to the dexamethasone approach.To correct the likely undercounting from both approaches, we propose the diagnosis-based approach as an alternative with greater face validity (Table 1, Figure 3).

Limitations
Our study was retrospective, observational, and limited to variables that were systematically collected by the emergency medicine group, such as ICD-10 diagnoses and medication orders.Using different variables than the ones we chose, or a manual chart review, may result in different estimates of non-incidental COVID-19 hospitalizations.Future research might examine integrating more clinical information (e.g., vital signs) into the definitions identifying non-incidental COVID-19 infections.Furthermore, there is no gold standard criteria for comparison to classify incidental versus non-incidental COVID-19 hospitalizations.For example, the CDC has had multiple iterations of its diagnosis-based definition [16].Also, we did not capture ED diagnoses beyond the first three diagnoses (primary plus two secondary).We also used CCSR codes, rather than assessing relevant co-diagnoses directly from ICD-10 codes.Lastly, the interrater reliability between raters in choosing high-, medium-, and low-likelihood co-diagnoses was fair.Therefore, a different set of raters, or non-ED clinicians, may have classified the CCSR codes differently.Finally, our results were limited to ED visits during the study period and may not generalize to new variants, which may have different disease manifestations.

Conclusions
We propose an approach to measuring which hospitalizations of patients with ED-diagnosed COVID-19 infections were non-incidental based on co-diagnoses that are highly related to COVID-19 infection.This approach counts substantially more non-incidental hospitalizations than the dexamethasone or CDC-based approaches and comports better with our clinical experience as ED physicians who must decide which COVID-19-positive patients to admit and with the available evidence from chart-review-based studies.
Given the high interrater reliability between "Maybe Related" and "Not Related", we also assessed interrater reliability for only "High Likelihood" and "High/Medium Likelihood" diagnoses between the three raters (JO, NR, and DN), assuming they agreed on which diagnoses were "Not Related" (Table 6).For determining which diagnoses were "High Likelihood" or not between the three raters, Fleiss' Kappa would be 0.546 (z =

FIGURE 1 :
FIGURE 1: Study flow diagram depicting the study population and exclusions.

FIGURE 2 :
FIGURE 2: Total percent of COVID-19 infections admitted/transferred. Line plots of the percentage of ED patients with COVID-19 infection who were admitted/transferred over the study period (denominator being all COVID-19-infected ED patients), including all infected patients and those in the diagnosis-based definition.

FIGURE 4 :
FIGURE 4: Percent of hospitalized COVID-19 patients receiving dexamethasone.Line graph of the percent of COVID-19 hospitalizations receiving dexamethasone over time comparing Massachusetts (MA) in-hospital data to US Acute Care Solutions (USACS) national ED data.

TABLE 3 : Demographics for ED patients admitted/transferred with a COVID-19 diagnosis.
SD, standard deviation a The Pre-Omicron period is from April 1, 2020, to December 18, 2021.The Omicron period is from December 19, 2021, to August 31, 2023 (end of available study data).The delineation between periods is based on metadata associated with sequences available on Global Initiative on Sharing All Influenza Data (GISAID), and accessible at doi.org/10.55876/gis8.220330me.bDiagnoses under CCSR codes: RSP002, INF002, RSP012, and GEN002.

Omicron a Omicron a Total
Line plot of the percentage of admissions/transfers for COVID-19 over the study period using different definitions of non-incidental COVID-19.2024Nikolla et al.Cureus 16(3): e56546.DOI 10.7759/cureus.565467 of 27Pre-
a b Since the diagnosis-based approach includes acute hypoxemic respiratory failure with hypoxemia and other respiratory failure diagnoses, the additional cases captured by the diagnosis-based or dexamethasone approach are likely false positive cases with dexamethasone administered for indications other than COVID-19.

TABLE 6 : Clinical Classifications Software Refined (CCSR) codes used to classify diagnoses as related to COVID-19 or not.
[26]ical Classifications Software Refined (CCSR) codes group diagnosis codes into similar clinically meaningful categories[26].