Characteristics and Outcomes of US Patients Hospitalized With COVID-19

Ithan D. Peltan, MD, MSc; Ellen Caldwell, MS; Andrew J. Admon, MD, MPH, MSc; Engi F. Attia, MD, MPH; Stephanie J. Gundel, RD; Kusum S. Mathews, MD, MPH, MSCR; Alexander Nagrebetsky, MD, MSc; Sarina K. Sahetya, MD, MHS; Christine Ulysse, MS; Samuel M. Brown, MD, MS; Steven Y. Chang, MD, PhD; Andrew J. Goodwin, MD, MSc; Aluko A. Hope, MD, MSCE; Theodore J. Iwashyna, MD, PhD; Nicholas J. Johnson, MD; Michael J. Lanspa, MD; Lynne D. Richardson, MD; Kelly C. Vranas, MD, MCR; Derek C. Angus, MD, MPH; Rebecca M. Baron, MD; Benjamin A. Haaland, PhD; Douglas L. Hayden, PhD; B. Taylor Thompson, MD; Todd W. Rice, MD, MSc; Catherine L. Hough, MD, MSc


Am J Crit Care. 2022;31(2):146-157. 

In This Article


Design and Setting

We conducted a retrospective, multicenter cohort study of adult patients admitted to US hospitals with laboratory-confirmed SARS-CoV-2 infection and symptomatic COVID-19. Participating hospitals were members of the National Heart, Lung, and Blood Institute Prevention and Early Treatment of Acute Lung Injury (PETAL) Network and included 57 geographically diverse US hospitals organized within 12 clinical centers (Figure 1). The PETAL Network central institutional review board at Vanderbilt University and the institutional review boards at each participating hospital approved the study or determined that the study was exempt from review.

Figure 1.

Map of contributing hospitals with associated county-level COVID-19 incidence during cohort eligibility. Choropleth map illustrates spatial variation in county-level COVID-19 incidence rate (cases per 10 000 residents) during the third week of March 2020 (see Supplement 1). Dots represent contributing hospitals. For closely adjacent hospitals, a single dot indicates the location of multiple hospitals and is labeled with the number of contributing sites represented.

The 57 geographically diverse hospitals participating in this retrospective cohort study were members of the NHLBI-sponsored PETAL Network.


Patients aged 18 years or older admitted to a study hospital from March 1 to April 1, 2020, were eligible for inclusion if they had a positive polymerase chain reaction test result for SARS-CoV-2 during their admission or within the preceding 14 days and exhibited clinical or radiological evidence of acute infection (including fever, cough, dyspnea, hypoxemia, or bilateral airspace opacities). We excluded prisoners and patients with prior hospital admission for COVID-19. Each clinical center contributed data from up to 125 consecutive patients drawn from that center's contributing hospitals. Because some clinical centers admitted fewer than 125 eligible patients during the study period, clinical centers with excess eligible patients contributed additional participants toward a total study inclusion target of 1500 patients.

Data Collection

Trained personnel obtained data on demographic and clinical characteristics, interventions, and outcomes by manual review of medical records according to a standardized protocol. Abstracted data were entered into a structured data capture interface with integrated real-time data validation.[7] Manual medical record review was supplemented at some sites by electronic data abstraction. Patients were followed until hospital discharge. Additional assessments were performed on hospital days 1, 4, 8, 15, 21, and 28 and (if applicable) on day 1 in the intensive care unit (ICU). Each site was also asked to provide counts and basic demographics of all patients hospitalized with COVID-19 during the study window.

To quantify illness severity, we adapted an 8-point ordinal outcomes scale recommended by the World Health Organization (Supplemental Table 1, available online only at[8] Scale values used the worst available value for the calendar day or, if data were missing, from an adjacent day. Sequential Organ Failure Assessment score calculation did not incorporate urine output but otherwise used standard methods, including assigning component scores of 0 when data were missing.[9] Respiratory support was defined by treatment with supplemental oxygen or positive pressure ventilation. When PaO2 data were unavailable, values were estimated from peripheral oxygen saturation (Spo2) values using a validated nonlinear formula.[10] For patients receiving oxygen by nasal cannula or face mask, the fraction of inspired oxygen (FIO 2) was estimated using the formula FIO 2 = 0.21 + (0.03 × [oxygen flow rate in liters per minute]). Comorbidities, symptoms and their duration, and complications were obtained from clinical documentation.


The primary outcome was hospital mortality. Prespecified secondary outcomes included respiratory failure (defined as treatment with oxygen at ≥11 L/min delivered by face mask, high-flow nasal cannula, noninvasive positive pressure ventilation, or invasive mechanical ventilation) occurring early (≤24 hours) or late (>24 hours) after hospital presentation. Other secondary outcomes included 7-, 14-, and 28-day hospital mortality; COVID-19 ordinal outcome scale values on hospital days 4, 8, 15, and 28; length of hospital stay; respiratory, cardiovascular, and renal support therapies; and survivors' discharge health status.

Statistical Analysis

Continuous data are reported as medians and interquartile ranges (IQRs). Categorical data are reported as numbers and percentages. For descriptive analyses, we did not perform statistical hypothesis testing.

We employed L1 (lasso)–penalized logistic mixed-effects regression[11,12] to identify risk factors for mortality, early-onset respiratory failure, and late-onset respiratory failure from outcome-specific sets of candidate risk factors identified a priori by a team of experienced epidemiologists and clinical researchers on the basis of previously reported association, plausibility, clinical utility, and data availability. To manage missingness among candidate risk factors, we performed penalized regression after multiple imputation of missing data using chained equations.[13] Adjusted effect sizes for selected risk factors were estimated in the multiply imputed data using multivariable mixed-effects logistic regression and combined coefficients using Rubin's rules.[13–16] To account for site-level clustering of patient characteristics, care practices, and outcomes as well as between-site variation in resource strain,[17] we employed a random effect for study site during both penalized regression[18] and multivariable logistic regression model refitting for effect size estimation. Additional details are available in Supplement 1 (available online only).

The vast majority of patients who received invasive mechanical ventilation were also treated with vasopressors.

We assessed our findings' robustness in prespecified sensitivity analyses by reestimating effect sizes after (1) reclassifying patients discharged with hospice services as having the primary outcome (in-hospital mortality); (2) excluding patients who died without respiratory failure from the secondary analysis of late respiratory failure; and (3) excluding support with oxygen by face mask from the definition of respiratory failure. As an additional measure of variable importance, we also report the percentage of models in which each candidate variable was ultimately selected during cross-validation.[19] Analyses were performed with R, version 4.0.3 (R Foundation for Statistical Computing); Stata, version 16.1 (StataCorp); and SAS, version 9.4 (SAS Institute Inc).