Predictors at admission of mechanical ventilation and death in an observational cohort of adults hospitalized with COVID-19

Abstract Background Coronavirus disease (COVID-19) can cause severe illness and death. Predictors of poor outcome collected on hospital admission may inform clinical and public health decisions. Methods We conducted a retrospective observational cohort investigation of 297 adults admitted to eight academic and community hospitals in Georgia, United States, during March 2020. Using standardized medical record abstraction, we collected data on predictors including admission demographics, underlying medical conditions, outpatient antihypertensive medications, recorded symptoms, vital signs, radiographic findings, and laboratory values. We used random forest models to calculate adjusted odds ratios (aORs) and 95% confidence intervals (CI) for predictors of invasive mechanical ventilation (IMV) and death. Results Compared with age <45 years, ages 65–74 years and ≥75 years were predictors of IMV (aOR 3.12, CI 1.47–6.60; aOR 2.79, CI 1.23–6.33) and the strongest predictors for death (aOR 12.92, CI 3.26–51.25; aOR 18.06, CI 4.43–73.63). Comorbidities associated with death (aORs from 2.4 to 3.8, p <0.05) included end-stage renal disease, coronary artery disease, and neurologic disorders, but not pulmonary disease, immunocompromise, or hypertension. Pre-hospital use vs. non-use of angiotensin receptor blockers (aOR 2.02, CI 1.03–3.96) and dihydropyridine calcium channel blockers (aOR 1.91, CI 1.03–3.55) were associated with death. Conclusions After adjustment for patient and clinical characteristics, older age was the strongest predictor of death, exceeding comorbidities, abnormal vital signs, and laboratory test abnormalities. That coronary artery disease, but not chronic lung disease, was associated with death among hospitalized patients warrants further investigation, as do associations between certain antihypertensive medications and death.

M a n u s c r i p t 3

Pandemic coronavirus disease 2019 (COVID-19) is causing severe illness and deaths across
Data from a variety of healthcare settings are needed on patient the United States and the world. characteristics and clinical findings on admission to predict who is most likely to receive invasive mechanical ventilation (IMV) and die.
Several studies examined predictors of adverse outcomes in COVID-19 and proposed predictive criteria based on specialized laboratory testing [19], but some of these studies examined laboratory values obtained days into patients' hospital courses, making them less useful in predicting later outcomes than admission values [19][20][21][22]. Furthermore, predictors of IMV may be different from predictors of death, since many patients with IMV will recover and not all patients who die have received IMV. In this investigation, we gathered descriptive data available to most clinicians on patient hospital admission to examine predictors of IMV and death to inform clinical and public health practice.
A c c e p t e d M a n u s c r i p t 4

Methods
The Centers for Disease Control and Prevention (CDC) and the Georgia Department of Public Health (DPH) partnered with three hospital networks to abstract medical records of patients hospitalized with COVID-19 in eight Georgia hospitals and assess the association between patient characteristics, underlying conditions, pre-hospital medications, and clinical findings on patient presentation with receipt of IMV or death. Seven hospitals were in metropolitan Atlanta, and one was in the southern region of Georgia; all provided tertiary care and included academic medical centers, a public teaching hospital, and community hospitals. CDC and Georgia DPH determined this investigation to be a non-research public health response activity.

Patient population
We collected data on patients hospitalized during March 2020 with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection-the virus that causes COVID-19-confirmed by reverse transcription-polymerase chain reaction; observation stays and deaths in the emergency department were also eligible for inclusion. Patients transferred between participating hospitals or admitted multiple times to the same hospital during March were analyzed as having a single hospitalization. Hospitals provided lists of patients with SARS-CoV-2 infection admitted during March 1-March 30, 2020 (n=698). We abstracted data from medical records of 305 adult patients (≥18 years old) sequentially selected from these lists. For analysis, we included only patients with completed hospitalizations (i.e., discharge or death, n=297) as of May 8, 2020.

Data collection
During March 25-May 8, 2020, investigators abstracted medical records using a secure REDCap form [23] that included elements on patient demographics, underlying medical conditions, pre-hospital medications, whether reason for admission was COVID-19 related, presenting signs and symptoms, laboratory testing, radiographic imaging, and outcomes. Records reviewed included A c c e p t e d M a n u s c r i p t 5 clinician notes and first recorded vital signs, laboratory values, and imaging. The database was continually reviewed to correct missing data and implausible values. History of stroke was included under cardiovascular and not neurologic conditions. Hypoxia on admission was defined as oxygen saturation ≤94% on room air or use of supplemental oxygen. Pre-hospital antihypertensive medications were classified as ACE inhibitors, ARBs, beta blockers, dihydropyridine calcium channel blockers (dCCBs), thiazide diuretics, non-thiazide diuretics (e.g., furosemide, spironolactone), other vasodilators (e.g., hydralazine), and other medications (e.g., clonidine, prazosin).
Immunocompromise was defined as cancer with chemotherapy receipt within the previous year, history of solid organ or stem cell transplant, HIV infection, or current use of immunosuppressive medications. Outcomes were defined as IMV and death and were examined separately.

Analysis
To evaluate independent predictors of IMV and death at hospital admission, we used counterfactual random forest probability machines [24,25] to adjust for all other variables reported, including number of comorbidities and whether reason for admission was COVID-19-related. In brief, separate random forest models were developed using patients with each value of the covariate under consideration. Estimated probabilities of the outcome were generated for all patients, including those who experienced a different value of exposure (i.e., counterfactual). From the sum of the predicted probabilities, two-by-two tables were constructed and scaled to the data's exposure margins. Odds ratios and standard errors were subsequently calculated by standard methods. Continuous variables (age, body mass index [BMI], vital signs, and laboratory values) were entered into random forests as continuous, but we reported associations for individual variables using quantiles or standard categories (for age and BMI) that roughly aligned with quintiles.
Categories with the lowest values were used as reference groups. We used quantiles rather than laboratory reference ranges to provide finer detail than reference ranges allowed, as use of reference ranges only might obscure clinically meaningful differences.
A c c e p t e d M a n u s c r i p t 6 R statistical software (version 3.6.3; The R Foundation) was used to conduct analyses.
Random forests were generated using the randomForestSRC package (version 2.9.3) with default settings (1,000 trees per forest, square root of the number of variables randomly selected as node splitting candidates, 10 random splits considered for each continuous variable). Missing data, ranging from 0% (for 64% [55/86] of variables, e.g., age) to 29% for alkaline phosphatase, were imputed by a single random forest imputation (impute.rfsrc function).
To identify simple algorithms predictive of IMV and death independent of the random forest model, we developed machine-generated decision trees called fast-and-frugal trees (FFTs), which identify the variables and cut-points (for continuous variables) most predictive of outcome. Each decision point, or node, has two branches, one of which continues the tree, and the other is an exit.
For the final node, both branches are exits. FFTs were generated using the ifan algorithm in the R FFTrees package (version 1.5.2) to best balance sensitivity and specificity. Final models rounded continuous variable cut-points to whole numbers to simplify use.

Predictors of IMV and death
In random forest models, increasing age was the strongest predictor of death, with age ≥75 years having an adjusted odds ratio (aOR) of 18 A c c e p t e d M a n u s c r i p t 8 History of hypertension or number of antihypertensive medications before admission was not associated with IMV or death controlling for other variables in the models, including age and comorbidities. However, pre-hospital use of ARBs or dCCBs, specifically, were associated with twice the odds of death (aOR 2.02, 95%CI 1.03-3.96, and aOR 1.91, 95%CI 1.03-3.55, respectively) compared with patients not taking either. Pre-hospital ARB use was also significantly associated with IMV (aOR 1.84, 95%CI 1.02-3.32), but not dCCB use.
Among recorded admission signs and symptoms, altered mental status was significantly predictive of adverse outcomes, having 4.99 times the odds (95%CI 2.07-12.01) of death compared with patients without this condition recorded ( Table 2). Among presenting vital signs, hypoxia and elevated respiratory rate were significantly predictive of IMV. Admission systolic blood pressure <111 mmHg (first quintile) vs. 122-131 mmHg (third quintile) and diastolic blood pressures <65 (first quintile) vs. 78-85 mmHg (fourth quintile) were associated with significantly higher odds of IMV. For respiratory rate, the upper three quartiles of respiratory rate (≥19 breaths per minute) had elevated aORs for death compared with the referent lowest quartile, but respiratory rate was statistically significant only for the second quartile (aOR 3.87, 95%CI 1.62-9.28).
Laboratory tests associated with increased odds of death included thrombocytopenia (lowest quintile, platelets <142 cells/mm 3 ) compared with other quintiles and the highest quintile of AST (≥63 IU/L) compared with the lowest ( Table 3). Compared to those with values in the lowest quintiles, the highest quintile of absolute lymphocyte count (≥1.47 cells/mm 3 ) was protective for death, the highest blood urea nitrogen (BUN) quintile (≥27 mg/dL) had greater odds of IMV and death, and the highest creatinine quintile (≥1.65 mg/dL) was also associated with IMV and was nonsignificantly associated with death (p=0.15). Certain higher quintiles of alanine aminotransferase (ALT), AST, and total bilirubin were also associated with IMV compared with the lowest. Presence of a bilateral or multifocal infiltrate was significantly associated with death (aOR 1.98, 95%CI 1.05-A c c e p t e d M a n u s c r i p t 9 3.76); other abnormality or opacity on chest radiograph was not significantly associated with outcomes.

Discussion
In this observational cohort of nearly 300 predominantly black adults hospitalized with COVID-19 early in the U.S. epidemic, age was by far the strongest predictor of death, with the odds of death increasing markedly among older patients (>12 times greater odds for age 65-74 years and >18 times greater odds for age ≥75 vs. those <45 years). By comparison, death was less strongly associated with underlying conditions (i.e., ESRD, neurologic disorders, and CAD), SNF residence, and clinical findings (aORs ≤4.2). The association between age and death persisted despite adjustment for a wide range of factors, including vital signs and laboratory results, which might be assumed to be more directly predictive of poor outcomes. Why older adults have a markedly higher risk of death merits further study but may relate to immune or vascular system changes that occur with aging.
Chronic lung disease [27], immunocompromise [27], tobacco use [28], and obesity [29] might be expected to be predictive of COVID-19-associated death based on data for influenza, another viral respiratory illness. However, in our analysis, these conditions were not associated with in-hospital death. Some earlier reports found associations between obesity and in-hospital death [13], in-ICU death [30], and severe illness [11] in COVID-19, although these studies also did not identify mortality associations for lung disease, immunocompromise, or smoking. However, various studies have linked all of these conditions to severe COVID-19 [31]. For example, a large U.K. study A c c e p t e d M a n u s c r i p t 10 found obesity and chronic lung disease to be associated with death from COVID-19 among the general population [2]. Although our investigation's relatively small sample size may have limited our power to detect associations, our results suggest that obesity without associated comorbidities was not a strong risk factor for in-hospital death. Further research is still needed to evaluate the associations between underlying conditions and risk of death in hospitalized COVID-19 patients.
That pre-hospital use of ARBs was associated with receiving IMV and in-hospital death, despite extensive adjustment for other factors, including black race, CAD, hypertension, and diabetes, is notable given that a link is biologically plausible and ARB use has been associated with renal dysfunction in COVID-19 [16,32]. However, other studies found no association between ARB use and mortality [33] or a composite adverse outcome [18] in COVID-19 patients. Why dCCB use was also associated with death in our investigation is unclear, and relationships between antihypertensives and COVID-19 outcomes warrant further examination in larger well-controlled studies. Given other studies have not linked pre-hospital antihypertensive use to death in COVID-19 and the limitations of our investigation, outpatients should continue on their prescribed antihypertensive regimens per existing guidance [34].
The simple FFT decision trees predicted outcomes with reasonable accuracy (70-75%), which was lower than a recently proposed 'rule-of-6' algorithm (~90%). However, this algorithm involved specialized laboratory testing (i.e., lactate dehydrogenase, C-reactive protein, ferritin) collected up to 48 hours after admission [35], whereas the FFT involve tests routinely ordered in the emergency department. Only three variables-age, AST, and BUN-were 75% predictive of death in the FFT model; elevated AST and BUN on admission may be markers of multisystem inflammation, which has been associated with severe disease. Several other clinical factors were also predictive of death in the random forest model: altered mental status, thrombocytopenia, and lower lymphocyte counts. Although previous studies have reported associations between elevated admission AST and death [3] and lower lymphocyte counts and severe COVID-19 [22], and others have identified A c c e p t e d M a n u s c r i p t 11 thrombocytopenia as a marker of poor outcomes [36], few studies have examined these factors on admission in a multivariable model. Notably, abnormal respiratory vital signs were less predictive of death, although they were strongly predictive of IMV in both the FFT and random forest models.
Our retrospective observational investigation has several notable limitations. Because data abstraction was limited to medical records, symptom data are less complete than those obtained by questionnaires. We were unable to evaluate certain specialized testing (e.g., C-reactive protein, lactate dehydrogenase, D-dimer) [20,22,37] because they were infrequently ordered on admission.
Second, the outcome IMV is highly influenced by clinical practice, and some clinicians may have pursued early IMV to minimize non-invasive ventilation and avoid emergency endotracheal intubation, given potential risks of viral transmission. As such, using death as an outcome, rather than solely relying on IMV or a composite outcome, allows examination of predictors less dependent on individual medical practices. Third, we used quantiles rather than vital and laboratory reference ranges, and some quantiles included a mix of normal and abnormal values which could have biased those categories toward the null. Fourth, our analysis had limited power to detect weak associations, given the relatively small sample size and adjustment for many factors. However, random forest models allowed robust control for confounders, offering benefits over logistic regression, by allowing examination of more covariates, requiring fewer assumptions, and better accounting for interactions [24,25]. Fifth, we examined nearly 100 admission factors. Although our approach may tend toward a bias to the null with this number of factors [24], we did not incorporate adjustments for multiple testing , and some associations might still have occurred by chance. However, the FFT yielded similar findings to our more well-controlled analyses; FFT also offer benefits over logistic regression because they rarely overfit data and are easy to interpret and use [38,39]. Finally, although records were selected sequentially in the order in which hospitals identified cases, this cohort is ultimately a convenience sample, as it did not encompass all COVID-19 patients admitted to these hospitals during March 2020 [26]. While findings from this cohort, involving predominantly non-Hispanic black patients in a limited geographical area and time, may M a n u s c r i p t 12 not be generalizable to other populations, our investigation provides valuable data on black patients with COVID-19, who have been disproportionately impacted by COVID- 19 [40].
In summary, we provide simple decision trees that found the most important predictors for IMV were hypoxia, elevated respiratory rate, elevated BUN, and low diastolic blood pressure; for death the most important predictors were older age (≥63 years), elevated BUN, and elevated AST.
These predictors were confirmed and augmented by several additional predictors from our multivariable model. Furthermore, the significant association between pre-hospital use of ARBs and IMV and death and dCCBs and death warrants additional investigation.  M a n u s c r i p t 18