Characterizing COVID-19 clinical phenotypes and associated comorbidities and complication profiles

Purpose Heterogeneity has been observed in outcomes of hospitalized patients with coronavirus disease 2019 (COVID-19). Identification of clinical phenotypes may facilitate tailored therapy and improve outcomes. The purpose of this study is to identify specific clinical phenotypes across COVID-19 patients and compare admission characteristics and outcomes. Methods This is a retrospective analysis of COVID-19 patients from March 7, 2020 to August 25, 2020 at 14 U.S. hospitals. Ensemble clustering was performed on 33 variables collected within 72 hours of admission. Principal component analysis was performed to visualize variable contributions to clustering. Multinomial regression models were fit to compare patient comorbidities across phenotypes. Multivariable models were fit to estimate associations between phenotype and in-hospital complications and clinical outcomes. Results The database included 1,022 hospitalized patients with COVID-19. Three clinical phenotypes were identified (I, II, III), with 236 [23.1%] patients in phenotype I, 613 [60%] patients in phenotype II, and 173 [16.9%] patients in phenotype III. Patients with respiratory comorbidities were most commonly phenotype III (p = 0.002), while patients with hematologic, renal, and cardiac (all p<0.001) comorbidities were most commonly phenotype I. Adjusted odds of respiratory, renal, hepatic, metabolic (all p<0.001), and hematological (p = 0.02) complications were highest for phenotype I. Phenotypes I and II were associated with 7.30-fold (HR:7.30, 95% CI:(3.11–17.17), p<0.001) and 2.57-fold (HR:2.57, 95% CI:(1.10–6.00), p = 0.03) increases in hazard of death relative to phenotype III. Conclusion We identified three clinical COVID-19 phenotypes, reflecting patient populations with different comorbidities, complications, and clinical outcomes. Future research is needed to determine the utility of these phenotypes in clinical practice and trial design.


Introduction
The coronavirus disease 2019 (COVID-19), a disease caused by the severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2), has infected over 18 million and led to over 700,000 deaths since first appearing in late 2019 [1]. Researchers are rapidly attempting to understand the natural history of and immune response to COVID-19 [2]. Despite intense research since the arrival of this novel coronavirus [3], only one pharmaco-therapeutic agent, dexamethasone, has been associated with reduced mortality in at-risk individuals [4]. COVID-19 results in a constellation of symptoms, laboratory derangement, immune dysregulation, and clinical complications [5].
Emergency department presentation varies widely, suggesting distinct clinical phenotypes exist and, importantly, it is likely these distinct phenotypes respond differently to treatment. To illustrate, two early phenotypes of respiratory failure likely exist in COVID-19. A classic ARDS phenotype exists with poorly compliant lungs and poor gas exchange; however, a phenotype with normal lung compliance also exists in COVID-19 and is hypothesized to be driven by shunting secondary to pulmonary microthrombi [6,7]. An intricate, multidimensional view is required to adequately understand the disease and account for the variation in clinical outcomes. Furthermore, patients could benefit from phenotype-specific medical care, which may differ from established standards of care.
Despite this need, few studies have characterized COVID-19 clinical phenotypes and evaluated their association with complications and clinical outcomes. The aim of this study was to characterize clinical phenotypes in COVID-19 according to disease-system factors using electronic health record (EHR) data pooled from 14 U.S. Midwest hospitals between March 7, 2020 and August 25, 2020.

Data collection
The data source for this study included EHR reports from 14 U.S. Midwest hospitals and 60 primary care clinics across Minnesota. The healthcare system includes an academic quaternary center along with community hospitals all capable of providing critical care. Patient and hospital-level data were available for 7,538 patients with PCR-confirmed COVID-19. Of these, 1,022 required hospital admission and were included in this analysis. The database included all The study was approved by all hospitals within the MHealth Fairview system which includes ethical approval by the University of Minnesota institutional review board. All patients have the option to opt-out of research upon establishing care within the MHealth Fairview healthcare system. Data is aggregated through the University of Minnesota's centralized informatics center and de-identified prior to analysis. Data were pooled across different electronic health records (EHRs) utilizing a unique patient identifier to account for health care encounters across systems. This study was approved by the University of Minnesota institutional review board (STUDY00001489), which provided a waiver of consent for this study.

Participants
Patient-level data were obtained from the COVID-19 database from March 7, 2020 to August 25, 2020. The inclusion criterion was as follows: PCR-positive COVID-19 test requiring inpatient hospital admission to one of the 14 hospitals providing data. No hospitalized patients were excluded in this analysis to maximize generalizability. Follow-up data were available for a minimum of two weeks following admission for all patients.

Clinical variables for phenotyping
We selected 33 variables for clustering based on their association with COVID-19 mortality, known COVID-19 pathophysiology, and presence in the database (no more than 50% missingness) [8][9][10][11]. The following variables were included: age, body mass index (BMI), heart rate, respiratory rate, oxygen saturation, pulse pressure, systolic blood pressure, total protein, red cell distribution width, mean corpuscular volume, alkaline phosphatase, calcium, anion gap, bicarbonate, hematocrit, aspartate aminotransferase, glucose, absolute monocyte count, absolute neutrophil count, absolute lymphocyte count, white blood cell count, platelet, albumin, bilirubin, international normalized ratio (INR), lactate dehydrogenase, potassium, sodium, Ddimer, hemoglobin, C-reactive protein (CRP), creatinine, and gamma gap. For each variable we selected the first recorded value within the first 72 hours of the emergency department (ED) presentation that ultimately resulted in their hospitalization.

Comorbidities
We selected 68 comorbidities documented for each patient from March 29, 1997 preceding their COVID-19 hospital admission in their electronic health record (S1 Table). All comorbidities were identified based on ICD-9, ICD-10, or problem list documentation within the electronic health record. An indicator variable was created for each comorbidity to denote the presence of the selected ICD-9, ICD-10, or problem list documentation at any time in the medical record. To facilitate analysis, comorbidities were grouped by organ system into the following categories: cardiac, respiratory, hematologic, metabolic, renal, hepatic, autoimmune, cancer, and cerebrovascular disease.

Complications and clinical outcomes
We selected 30 in-hospital complications measured during each patient's hospital stay for COVID-19 categorized into the following systems: cardiovascular, respiratory, hematologic, renal, hepatic, metabolic, and infectious (S2 Table). If applicable, complications could span multiple organ system variables. For example, ventilator associated pneumonia was included in both infectious and respiratory complications. Additional clinical outcomes included hospital length of stay (LOS), need for intensive care unit (ICU) admission, need for mechanical ventilation, and mortality. Mortality was defined as any in-hospital or out-of-hospital death based on death certificate data. All complications and outcomes were followed for a minimum of 2 weeks following hospital admission.

Statistical analysis
The overall rate of missingness of the 33 variables used for phenotyping, which included the first vitals and labs recorded for each inpatient within 72 hours of admission, was 19% (range 0% -50%). We imputed missing values using multivariate imputations by chained equations implemented with the mice package (v.3.10.0) [12,13]. Data were log-transformed before imputing missing values with predictive mean matching. A total of 40 imputed datasets were generated. The diceR package (v.1.0.0) [14] was used to perform k-means-based consensus clustering on each imputed dataset using 80% subsamples and 1,000 iterations. We considered grouping patients into 2-7 phenotypes and determined the optimal number was 3 by evaluating the consensus cumulative distribution function (CDF) plot, the delta area plot, and the consensus matrix heatmap. These figures were generated using the consensus clustering results for each imputed dataset, and all figures were qualitatively similar across datasets. For visualization purposes, these images are provided for a randomly selected imputed dataset in S1-S4 Figs. The final assignment of each patient into one of the three phenotypes was determined by majority voting across the 40 consensus clustering results. Principal component analysis (PCA) was performed on the average covariance matrix to visualize the relationships among the three phenotypes and assess variable contributions [15].
Continuous variables were summarized using the median and interquartile range (IQR) and compared across phenotypes using a Kruskal-Wallis test. Categorical characteristics and outcomes were summarized using counts and proportions and compared across phenotypes using a Pearson's chi-squared test or Fisher's exact test. Multinomial regression models were fit to further compare patient comorbidities across phenotype classification.
We next evaluated the relationship between phenotype and subsequent outcomes using both unadjusted and adjusted models. The adjusted models included sex [16,17], race and ethnicity (white, Black, Asian, Hispanic, other, not reported) [18], and Elixhauser Comorbidity Index [19], since these are known risk factors for the outcomes of interest and were not included in the clustering analysis. The associations between phenotype and complications, ICU admission and need for mechanical ventilation, were estimated using logistic regression models. Mortality was compared across phenotypes using Cox proportional hazard models and patients were censored at the last date of data collection, August 25, 2020. Hospital length of stay was compared across phenotypes using negative binomial regression models. The primary negative binomial model included individuals who died during hospitalization for whom length of stay was defined as the number of days until death. We performed a sensitivity analysis to assess the impact of mortality as a competing risk by refitting the length of stay model after removing the 127 patients who died. Two-sided p-values < 0.05 were considered statistically significant. P-values were not adjusted for multiple comparisons. Visualizations of comorbidities, complications, and outcomes by clinical phenotype were performed using the circlize package for R [20]. Comorbidities and complications were grouped into separate organ systems and the prevalence of each complication/comorbidity type was calculated as a percentage for each phenotype. All analyses were conducted using R version 3.6.3 [21] and Stata version 16.1 (StataCorp).

Results
The database included 1,022 patients requiring hospital admission with COVID-19. Among these patients, the median age was 62.

Variable contributions to clustering
The first two principal components (PCs) from PCA were used to visualize the relationship between phenotypes. PC1 and PC2 captured approximately 11% and 9% of the variance in the clustering variables, respectively. Thirteen components were needed to explain 70% of the variance (S5 Fig). While phenotypes II and III overlay substantially, phenotype I is more clearly defined in the right-hand side of the score plot of the first two principal components (Fig 1). Notably, this figure shows that distinctions between phenotypes are primarily driven by variation in PC1 as opposed to PC2. The variable contributions to PC1 (S6 Fig) demonstrate that the largest contributors to the variation in PC1 are from lactate dehydrogenase (LDH), absolute neutrophil count, and D-dimer. These variables therefore prominently contribute to separating the three phenotypes as shown in the biplot (Fig 2). Univariate tests showed that LDH, D-dimer, and neutrophil count are highest in phenotype I. Other variables influential to phenotype clustering are white cell count (highest in I), C-reactive protein (highest in I), albumin (highest in III), aspartate aminotransferase (highest in I), bilirubin (highest in I), and oxygen saturation (highest in III).

Phenotype characteristics
Differences across phenotypes with respect to patient demographics, admission vitals and labs, complications, comorbidities, and clinical outcomes are presented in Table 1. Patients with phenotype I were older than patients in phenotypes II and III (67.2 [52.9, 79.0] years vs. 60.9 [45.9, 75.4] and 58.6 [34.8, 71.3] years respectively, p < 0.001). Patients with phenotype III were more often female than patients with phenotype I or II (57.6% vs. 41.6% and 53.4%, respectively, p = 0.002). Patients with phenotype I were less likely to be white (38.8% vs. 45.6% vs. 60.7%, respectively, p = 0.002) and more likely to be non-English speaking (47.9% vs. 39.2% vs. 23.7%, respectively, p <0.001). There were no statistically significant differences in BMI or socioeconomic status, as measured using the area deprivation index, between phenotypes ( Table 1). Patients that presented with phenotype III had a more frequent history of smoking, alcohol abuse, and neutropenia. Patients that presented with phenotype II had a less frequent history of hepatic disease than phenotypes I or III (Table 1).

Association between phenotype and clinical outcomes
Clinical phenotypes I and II were associated with increased odds of respiratory .001) increase in risk of mortality, and Phenotype II had a 2.57-fold (HR: 2.57, 95% CI: 1.10-6.00, p = 0.03) increase in the hazard of death compared to Phenotype 3. We performed a sensitivity analysis to assess the impact of mortality as a competing risk by fitting the LOS model before and after removing the 127 patients who died. The estimated effect sizes were similar between these two models (S4 Table).  Table).

Discussion
This is one of the first studies to report on clinical phenotypes associated with COVID-19. We identified three clinical phenotypes for patients with COVID-19 on hospital presentation. Most patients presented with phenotype II, which is associated with a moderate course and an approximately 10% mortality. A subset of patients presented with the more severe phenotype I, which is associated with a staggering 27% mortality. Patients with cardiac, hematologic, and  renal comorbidities were most likely to be characterized by phenotype I. Surprisingly, respiratory comorbidities appeared less related to phenotypes I or II and were most associated with phenotype III, which had the most indolent course. Despite this indolent course, patients with phenotype III had the highest rate of readmission which is likely in part due to the high survival rate. This also suggests patients with pre-existing respiratory comorbidities, while not at highest risk for mortality, may be at highest risk for long term sequalae following COVID-19. Patients that presented with phenotype I were most associated with the development of respiratory, hematologic, renal, metabolic, hepatic, and infectious complications. Surprisingly, cardiovascular complications did not significantly differ between phenotypes.
Elucidating patient risk factors and severe COVID-19 disease markers may allow early treatment implementation that may improve the patient's outcome. Multiple studies have documented COVID-19 risk factors; however, most have done so from a homogenous lens. For example, a prospective cohort study from New York City identified that the most considerable risks for hospital admission were age, male sex, heart failure, chronic kidney disease, and high BMI [22]. A large observational study conducted in the UK reported that increasing age, male gender, comorbidities such as cardiac disease, chronic lung disease, chronic kidney disease, and obesity were associated with higher mortality in COVID-19 positive patients admitted to the hospital. 14 A study from China found that increased odds of in-hospital death due to COVID-19 were associated with older age, higher sequential organ failure assessment (SOFA) score and D-dimers > 1.0 μg/mL on admission [23]. Another retrospective study reported that patients with severe COVID-19 disease and diabetes had increased leucocytes, neutrophils count, and increased C-reactive protein (CRP), D-dimers, fibrinogen levels [24]. A systematic review and meta-analysis found that the biomarkers associated with increased mortality include higher CRP, higher D-dimers, increased creatinine, and lower albumin levels [25]. However it is well known that patients do not have a singular natural history of disease. Multiple studies including this study found that only half of patients suffer a primarily respiratory disease [26,27]. Patients suffer a constellation of cardiovascular, hematologic, renal, or hepatic progression of disease following COVID-19. It is likely patient baseline risk factors related to the virus [28], home medications [16,29], genetic predisposition [30], race/ethnicity [18], and other factors predispose patients to one of the various clinical manifestations and natural history of COVID-19.
Treatment of hospitalized patients should be tailored based on the clinical courses most likely for a patient given their a priori risk. For example, phenotypes with a higher risk of thrombotic events, may benefit from more aggressive anticoagulation. Phenotypes more prone to infectious complications, may benefit from more targeted immunomodulation instead of broad and systemic steroid therapy. A key first step to evaluate these treatment decisions is to characterize and describe clinical phenotypes requiring hospitalization. In this analysis we identified three clinical phenotypes for patients that required hospitalization for COVID-19. Few studies to date have attempted to elucidate clinical phenotypes. One study attempted to characterize clinical phenotypes at ICU admission using a dataset of 85 critically ill patients [31]. Similar to our analysis, they identified three distinct clinical phenotypes. Their The scores (points) and loadings (arrows) of PC1 and PC2 are plotted for each patient and variable in the model. 95% confidence ellipses for the scores are shown. The biplot facilitates interpretation of the scores and loadings, assigning context to the variables which prominently contribute to the phenotypes. Abbreviations: PC1 (principal component 1); PC2 (principal component 2); PCA (principal component analysis); Abs_Nphil_Ct (absolute neutrophil count); LDH (lactate dehydrogenase); CRP (C-reactive protein); WBC (white blood cell count); HCT (hematocrit); HGB (hemoglobin); Tbili (total bilirubin); RDW (red cell distribution width); AST (aspartate aminotransferase); Alk_phos (alkaline phosphatase); RR (respiratory rate); CA (calcium); TP (total protein); INR (internal normalized ratio of prothrombin time); CO2 (carbon dioxide); K (potassium); O2SAT (oxygen saturation); BMI (body mass index); PLT (platelet); PP (pulse pressure); Na (sodium); SBP (systolic blood pressure); Abs_mono_ct (absolute monocyte count); MCV (mean corpuscular volume).
https://doi.org/10.1371/journal.pone.0248956.g002  low mortality cluster which they called cluster 1 was very similar to our phenotype III with a predominance of females, lower mortality rate, lower D-dimer and CRP levels. Similarly, their high mortality cluster was predominantly male, with elevated inflammation markers on ICU presentation. In this study, we not only characterized three clinical phenotypes, but extended findings outside of the ICU by characterizing the association of comorbidities with clinical phenotype and the association of clinical phenotypes with in-hospital complication and clinical outcomes. Phenotype I can be termed the "Adverse phenotype" and was associated with the worst clinical outcomes. Lactate dehydrogenase (LDH), absolute neutrophil count, D-dimer, aspartate aminotransferase (AST), and C-reactive protein (CRP) were most influential in phenotype I determination. The strong association of red cell distribution width (RDW) with phenotype I was interesting. RDW was strongly associated with genetic age which is hypothesized to be a    risk factor in COVID-19 [30]. As people age, variability in red blood cell volumes increases. Similarly, Gamma Gap, a marker of immunoglobulin levels, was elevated in all three phenotypes (median > 3.5) [32]. However, patients with clinical phenotype I were noted to have the largest increase in Gamma Gap. In this scenario elevated Gamma Gap was likely an indicator of systemic inflammation and has been associated in other inflammatory disease processes with prognosis. Other groups have previously reported on the importance of the Absolute Neutrophil to Absolute Lymphocyte count, here we noted that ANC/ALC was lowest for phenotype III and highest for phenotype I, in line with previous reports. Patients with cardiac, hematologic, and renal comorbidities were most prone to develop phenotype I. Phenotype I was associated with numerous complications (hematologic, hepatic, metabolic, renal, respiratory, and infectious) when compared to other phenotypes. It is interesting to note despite a higher rate of baseline cardiac comorbidities phenotype I was not associated with increased cardiac complications. Beyond the pathophysiologic differences, it is important to note the higher proportion of non-White and non-English speaking patients in phenotype I. Moreover, socioeconomic status was similar across all phenotypes, which has been proposed to be a driver of disparate outcomes in healthcare. These findings are consistent with a recent study conducted across this populations of patients which found COVID-19 severity to be associated with minority populations and non-English speaking patients, independent of socioeconomic status. Given race/ethnicity and primary language spoken are social constructs and traits, respectively, which are not biologically grounded; these results require further investigation as to why these populations are at higher risk of developing phenotype I through mediation analysis of external factors (as opposed to these populations being an isolated cause of developing an unfavorable phenotype). Phenotype III was associated with the best clinical outcomes and can be termed the "Favorable Phenotype". Surprisingly, patients with phenotype III had a very high rate of respiratory comorbidities and the best clinical outcomes. What is most surprising is despite the lowest complication rate and mortality, this phenotype was associated with a greater than 10% rate of hospital readmission. Long-term sequelae from the critically ill remains an important target for patient centered improvements in care given the increasing loss of functional status among ICU patients predating the pandemic. It is possible that patients pre-existing respiratory comorbidities predisposed them to longer term sequelae which may have resulted in this readmission rate, although additional studies are needed to better elucidate these findings, specifically controlling for differences in survival. Patients with respiratory comorbidities such as asthma and COPD routinely use medications which may be protective in SARS-CoV-2 pathogenesis which may explain this protective effect. For example, our group has previously identified reduced mortality in COVID-19 for patients with asthma treated with beta2-agonists [16]. Patients with phenotype III were more likely to use inhaled steroids, nasal fluticasone, albuterol, and antihistamines.
Clinical phenotypes are critical during a pandemic when time and resources are scarce. Phenotypes not only enable the identification of risk factors; they also provide essential insight towards the high yield follow up investigations. For example, while noting the respiratory associations with phenotype III (favorable phenotype) is interesting, the more beneficial take away includes further investigations towards how these underlying conditions and/or their medications may mitigate illness severity. Lastly, by phenotyping patients affected by COVID-19; we set the foundation to begin comparing if these phenotypes are unique to SARS-CoV-2 or if similarities exist elsewhere. As the attention paid to personalized medicine accelerates; these studies are just the beginning. Future work will expand upon these phenotypes with the hope that they can assist in 1) identifying those at risk of poor outcomes, 2) precisely treating each phenotype (which may not be uniform across all phenotypes), and 3) preventing further complications in those phenotypes at higher risk. In addition, a deeper investigation into clinical phenotypes and associated genomic, transcriptomic, and proteomic is needed. The ability to classify patients into clinical phenotypes can facilitate the linkage of-omics data to better understand SARS-CoV-2 pathogenesis and natural history. Work is already being done to identify genetic host factors that may play a role in determining not only susceptibility to the virus, but also the clinical trajectory when infection does occur. Understanding COVID-19 severity, its biomarkers, and risk factors is paramount during the COVID-19 pandemic.
Our study has several limitations, including that this is a retrospective study and therefore results may be biased or subject to residual confounding. Second, patients were followed for variable lengths of time. Patients that were admitted in March 2020 thus had approximately 5 months of follow-up whereas patients admitted in late August had limited time. We accounted for this by conducting a Cox proportional hazard analysis when analyzing in-and out-of hospital mortality. Additionally, when the data were pulled, only 54 patients (5%) remained hospitalized. While most patients developed complications within their first 2 weeks of hospital admission, it is possible that they may still develop clinical complications which is not reflected in this analysis. Furthermore, our analysis was completed on hospitalized patients. It is important to recognize that our results are restricted to those who required hospitalization. Our data cannot be extrapolated to those with mild COVID-19 (i.e. not requiring hospitalization).