key: cord-0722496-ttl3xfx5
authors: Wang, Alfred Z.; Ehrman, Robert; Bucca, Antonino; Croft, Alexander; Glober, Nancy; Holt, Daniel; Lardaro, Thomas; Musey, Paul; Peterson, Kelli; Schaffer, Jason; Trigonis, Russell; Hunter, Benton R.
title: Can we predict which COVID‐19 patients will need transfer to intensive care within 24 hours of floor admission?
date: 2021-04-04
journal: Acad Emerg Med
DOI: 10.1111/acem.14245
sha: bc1189be1f63a924abb8ef744ec6ba6eb8f074f6
doc_id: 722496
cord_uid: ttl3xfx5

BACKGROUND: Patients with COVID‐19 can present to the emergency department (ED) at any point during the spectrum of illness, making it difficult to predict what level of care the patient will ultimately require. Admission to a ward bed, which is subsequently upgraded within hours to an intensive care unit (ICU) bed, represents an inability to appropriately predict the patient's course of illness. Predicting which patients will require ICU care within 24 hours would allow admissions to be managed more appropriately. METHODS: This was a retrospective study of adults admitted to a large health care system, including 14 hospitals across the state of Indiana. Included patients were aged ≥ 18 years, were admitted to the hospital from the ED, and had a positive polymerase chain reaction (PCR) test for COVID‐19. Patients directly admitted to the ICU or in whom the PCR test was obtained > 3 days after hospital admission were excluded. Extracted data points included demographics, comorbidities, ED vital signs, laboratory values, chest imaging results, and level of care on admission. The primary outcome was a combination of either death or transfer to ICU within 24 hours of admission to the hospital. Data analysis was performed by logistic regression modeling to determine a multivariable model of variables that could predict the primary outcome. RESULTS: Of the 542 included patients, 46 (10%) required transfer to ICU within 24 hours of admission. The final composite model, adjusted for age and admission location, included history of heart failure and initial oxygen saturation of <93% plus either white blood cell count > 6.4 or glomerular filtration rate < 46. The odds ratio (OR) for decompensation within 24 hours was 5.17 (95% confidence interval [CI] = 2.17 to 12.31) when all criteria were present. For patients without the above criteria, the OR for ICU transfer was 0.20 (95% CI = 0.09 to 0.45). CONCLUSIONS: Although our model did not perform well enough to stand alone as a decision guide, it highlights certain clinical features that are associated with increased risk of decompensation.

SARS-CoV-2 is a novel coronavirus first identified in Wuhan, China, in November 2019, which has quickly spread globally, with the United States accounting for nearly one-quarter of all cases. [1] [2] [3] As of the writing of this manuscript, cases have exploded exponentially in the United States after a brief period of stagnated growth. 4 Worldwide, the SARS-CoV-2 pandemic has killed hundreds of thousands of patients, with reported mortality ranging from 0.4% to 7%. 5 Those who are elderly or comorbid have the highest risk of death. 6, 7 While most patients have mild illness at onset, some are completely asymptomatic, and others eventually manifest severe symptoms requiring intensive care unit (ICU) hospitalization. [6] [7] [8] [9] Factors such as rapid disease progression, variability in decisions by inpatient and emergency department (ED) providers, and ICU bed availabilities can all complicate the process of predicting what level of care will be required for these patients. However, admission to a non-ICU bed, which is subsequently upgraded within hours to an ICU level of care, can put undue strain on the inpatient teams, who have to admit the patient: spending substantial time gathering information and writing orders, only to have another (ICU) team have to repeat the entire process again just several hours later. Similarly, admission to an ICU bed, which is then downgraded to a medical bed within 24 hours, may be problematic especially when there are bed shortages. In addition, placing a COVID-19 patient into a room that they quickly leave requires an extensive decontamination process and ultimately costs precious availability of an inpatient bed.

Predicting which patients are going to require ICU or ventilator support within 24 hours would allow more appropriate allocation of resources from the onset of admission, improving patient care and eliminating repetitive work and freeing up space and providers to care for the many other patients who need it during this pandemic. The primary objective of this study was to determine clinical variables associated with need for an upgrade to ICU care within 24 hours of admission to a non-ICU floor.

This retrospective electronic medical record (EMR) review was approved as exempt research by the local institutional review board (Indiana University). came first. A note indicating an ICU transfer that did not subsequently occur was not counted as an event. Secondary outcomes were death within 24 hours, death prior to hospital discharge, intubation within 24 hours, and intubation at any time during hospitalization.

Data are described using means (with standard deviation), median (with interquartile range), or proportions (with 95% confidence interval

[CI]), where appropriate; normality assumption was checked using the Shapiro-Wilk test. Given that limiting analysis to patients with complete data (complete case analysis) can lead to bias in study results, 11 multiple imputation (MI) was performed. Variables where missingness was ≤30% were imputed under the assumption that they were missing at random (MAR). Data were determined to have an arbitrary missingness pattern and, therefore, the fully conditional specification approach was used, with linear regression used to impute continuous variables and logistic regression (LR) used for categorical variables.

Cut-points for continuous predictor variables were determined using Youden's J statistic; to meet the distributional assumptions of the imputation model, right-skewed continuous data were log-transformed prior to imputation and then back-transformed prior to determination of the optimal cut-point. Auxiliary variables for the imputation model were selected where correlation (Pearson's r) with imputed variables was ≥0.4, or where aggregate values (OR = proportions) were significantly different between those with complete versus missing data on bivariate analysis (e.g., significantly different age between those with versus without missing values for imputed variable X). The number of imputations was set to the maximum percentage of missing data (m = 30), with 100 burn-in iterations before the first imputation step and 25 iterations between successive steps, which achieved >95% relative efficiency for all imputed variables. Convergence of the imputation models was assessed by visual inspection of trace plots. In the final model, imputed variables (number imputed, percent missing) were troponin (n = 116, 21.40%), procalcitonin (n = 162, 29.89%), total leukocyte count (n = 2, 0.37%), lymphocyte count (n = 22, 4.06%), and glomerular filtration rate (GFR) (n = 9, 1.66%), plus cut-points for each.

Auxiliary variables included aspartate aminotransferase, age, respiratory rate, initial ED oxygen saturation, CO 2 , death or intubation during hospitalization, obesity, history of HF, ischemic heart disease, diabetes mellitus, or COPD; the dependent variable for the primary outcome (ICU transfer within 24 hours) was also included.

After completion of the imputation model, LR was used to assess univariate association between clinical and laboratory variables and the primary outcome; those with a p-value of <0.2 were retained for further consideration in a multivariable (MV) model. and C) for use as a clinical decision aid, with final selection of components and performance performed as previously described. Finally, age (given the importance attributed to this factor by clinicians when making admission decisions) and disposition location (our data set included patients admitted to both the floor and the PCU and thus adjustment accounts for potential differences in odds of ultimately needing ICU level care between these groups) were added as covariates to the composite variable model to assess its independent association with the primary outcome. That is, the association of the composite variable with the primary outcome, regardless of patient age or location of disposition from the ED.

For the imputation models, to test the MAR assumption, 10 addition MI models, with 30 imputations each, were created under the assumption of missing not at random. The first five multiplied the continuous variables by a scale factor of 0.5 to 0.9, in steps of 0.1.

The next five were created using only one class of completely observed categorical variables (heart failure = yes, COPD = yes, diabetes mellitus = no, in-hospital death = no, in-hospital intubation = yes).

LR models, with the same variables as used in the main analyses, were then constructed, with pooled effects analyzed as previously 

Of 751 patents with PCR-confirmed COVID-19, 542 were initially admitted, 86 of whom were admitted directly to the ICU and were excluded from this study. Among the 456 included patients, the average age was 62.8% and 50.2% were female. Table 1 Table 3 ). AUC for this model was 0.84 (SE = 0.03, 95% CI = 0.78 to 0.89). No significant interactions were found among final variables or other clinically plausible (i.e., "by meaning") scenarios and thus none were included in the final model.

We derived a composite outcome variable using factors from the final MV model that would be available to EPs at the time of disposition location decision (Model 2a/2b in Table 3 ). GFR and WBC count were dichotomized at a cut-point determined by Youden's J-statistic (46 and 6.4, respectively). Initial ED oxygen saturation was dichotomized at 93%, which was felt to be more clinically useful than the Youden's cut-point of 82%, and remained a statistically significant discriminator of the primary outcome.

We ultimately derived a set of criteria and evaluated the utility of the instrument to identify either the highest risk or the lowest risk patients. For the composite of history of HF, plus initial oxygen saturation of <93%, plus either WBC count > 6.4 or GFR <46 (Model 2a in Table 3 Table 3 ). We additionally assessed whether patients without the high-risk criteria could safely be considered "low risk" (Model 2b in Table 3 ). Table 3 ). While a model of this type is informative, application at the bedside can be difficult, and therefore we created a dichotomous decision aid model (Model 2 in Table 3 ). Disposition location was excluded from this model since this information is not available to the ED clinician. However, because our data were compiled after admission (to detect occurrence of the primary outcome)

we created a final model that adjusted for disposition location to understand the independent association of our decision aid with ICU transfer (Model 3 in Table 3 ). Age was also included as a covariate in this model "by meaning" as it often influences disposition decisions by ED clinicians.

We chose to adjust for age rather than including it in the decision aid to prevent the loss of signal associated with dichotomizing a continuous variable. Other risk factors associated with increased odds of the primary outcome but not retained due to significance included bilateral findings on chest radiography, initial and last documented ED respiratory rate, and requiring supplemental oxygen upon ED presentation. Interestingly, our study differs from prior literature that link comorbidities such as type 2 diabetes, coronary artery disease, or obesity with increased illness severity. 12, 17, 18 We found that these risk factors (specifically, hypertension, hyperlipidemia, COPD, smoking history, obesity, coronary artery disease and length of disease) were not significant for predicting who would need critical care within 24 hours. Notably, these factors have previously been shown to be related to final disease severity such as mortality, but in our study were not helpful in predicting 24-hour decompensation.

Our final composite (dichotomous) decision aid to identify "highrisk" patients consisted of history of HF, initial oxygen saturation of <93%, WBC count > 6.4, or GFR < 46 and was associated with an OR of 5.43 predicting ICU transfer, with a high specificity of 0.98 and low sensitivity of 0.11. Although this rule was highly specific, very few patients met the criteria for high risk and there was a high occurrence of false positive making its clinical utility doubtful.

We also assessed the ability of the instrument to identify those at lowest risk: those patients with no history of HF, initial oxygen 

There were several important limitations in our study. The most prominent limitation in our study is that the best-fit model we could design appears to have limited clinical utility. We initially strived to find a specific model that could help determine which patients were at high risk of needing an ICU bed within 24 hours of admission. Our model (2a/3a in Table 3 ) was highly specific but had such low sensitivity and identified so few patients as high risk that it would have a limited role at the bedside.

We reversed the criteria to try to identify low-risk patients (2b/3b in Table 3 ) for decompensation. The utility of this version was more promising, with higher sensitivity and moderate specificity and a negative predictive value of 0.96. However, like many clinical decision rules, both versions neglect clinical gestalt. 19 Furthermore, similar to many other COVID-19-specific decision rules, our model had different "high-risk" variables from other models published. For example, the quick COVID-19 severity index found a correlation with respiratory rate while the COVID-GRAM critical illness risk score includes such variables as cancer history and direct bilirubin. 16, 20 These models (including our own) may have different clinical/laboratory variables because of inherent differences between patient populations as well as statistical methodology. Because of these limitations, we suggest that when using these models, clinicians also add their clinical judgment when making disposition decisions. It seems likely that most of the patients who met the primary outcome had a legitimate need for ICU care, because the majority were intubated within 24 hours of arrival. Finally, these data were also collected from a single health care system in one state, which may limit generalizability.

Our model of history of heart failure, initial oxygen saturation at a cutoff of 93%, and either white blood cell count at a cutoff of 6.4 or glomerular filtration rate at a cutoff of 46 can assist in predicting which COVID-19 patients initially thought to not require intensive care unit level care are either particularly high or low risk for decompensating and requiring intensive care unit admission within the first 24 hours. However, its application does require further validation and it did not perform well enough to stand alone as a decision guide.

Outbreak of coronavirus disease 2019

Clinical characteristics of coronavirus disease 2019 in China

Coronavirus covid-19 has killed more people than SARS and MERS combined, despite lower case fatality rate

COVID-19) Dashboard. World Health Organization

The many estimates of the COVID-19 case fatality rate

Risk factors associated with acute respiratory distress syndrome and death in patients with coronavirus disease 2019 pneumonia in Wuhan, China

Clinical course and risk factors for mortality of adult inpatients with COVID-19 in Wuhan, China: a retrospective cohort study

Clinical characteristics and morbidity associated with coronavirus disease 2019 in a series of patients in metropolitan Detroit

A systematic review of asymptomatic infections with COVID-19

Research electronic data capture (REDCap)-a metadata-driven methodology and workflow process for providing translational research informatics support

The estimation and use of predictions for the assessment of model performance using large samples with multiply imputed data

Risk factors of critical & mortal COVID-19 cases: a systematic literature review and meta-analysis

Covid-19: risk factors for severe disease and death

A novel scoring system for prediction of disease severity in COVID-19

Paris-region area

Development and validation of the quick COVID-19 severity index: a prognostic tool for early clinical decompensation

Prevalence of comorbidities and its effects in patients infected with SARS-CoV-2: a systematic review and meta-analysis

Obesity in patients younger than 60 years is a risk factor for COVID-19 hospital admission

Structured clinical decision aids are seldom compared with subjective physician judgment, and are seldom superior

Development and validation of a clinical risk score to predict the occurrence of critical illness in hospitalized patients with COVID-19

Sensitivity of chest CT for COVID-19: comparison to RT-PCR

Diagnosis of the coronavirus disease (COVID-19): rRT-PCR or CT?

The authors have no potential conflicts to disclose.