key: cord-0805468-z8pnyt83
authors: Chen, H.; Varatharajah, Y.; Stewart de Ramirez, S. A.; Arnold, P.; Frankenberger, C.; Hota, B.; Iyer, R.
title: A Retrospective Longitudinal Study of COVID-19 as Seen by a Large Urban Hospital in Chicago
date: 2020-11-30
journal: nan
DOI: 10.1101/2020.11.29.20240606
sha: 5ff49731ebcd310e347ba0d4321b0e11f206b444
doc_id: 805468
cord_uid: z8pnyt83

The rapid spread of the novel coronavirus disease 2019 (COVID-19) has created a high demand for medical resources, including personnel, intensive care unit beds, and ventilators. As thousands of patients are hospitalized, the disease has shown remarkable diversity in its manifestation; many patients with mild to no symptoms recover from the disease requiring minimal care, but some patients with severe disease progression require mechanical ventilation support in intensive care units (ICU) with an increased risk of death. Studying the characteristics of patients in these various strata can help us understand the varied progression of this disease, enable earlier interventions for at-risk patients, and help manage medical resources more efficiently. This paper presents a retrospective analysis of 10,123 COVID-19 patients treated at the Rush University Medical Center in Chicago, including their demographics, symptoms, comorbidities, laboratory values, vital signs, and clinical history. Specifically, we present a staging scheme based on discrete clinical events (i.e., admission to the hospital, admission to the ICU, mechanical ventilation, and death), and investigate the temporal trend of clinical variables and the effect of comorbidities in each of those stages. We then developed a prognostic model to predict ventilation demands at an individual patient level by analyzing baseline clinical variables, which entails (1) a least absolute shrinkage and selection operator (LASSO) regression and a decision tree model to identify predictors for mechanical ventilation; and (2) a logistic regression model based on these risk factors to predict which patients will eventually need ventilatory support. Our results indicate that the prognostic model achieves an AUC of 0.823 (95% CI: 0.765-0.880) in identifying patients who will eventually require mechanical ventilation.

Since the emergence of COVID-19, first identified in Wuhan, China [1] , a global pandemic has ensued [2] , and widespread, sustained transmission has been observed worldwide. As of this writing, the virus has infected ~63 million people and caused over 1.4 million deaths worldwide, including over 13.75 million cases and over 273,000 deaths in the United States.

Studies have been published of the clinical characteristics and treatment outcomes in Chinese cities such as Wuhan [3] , Shanghai [4] , and Chongqing [5] , as well as New York City [6, 7] . These studies suggest that acute respiratory distress syndrome (ARDS) is a major driver of high mortality rates among critically ill patients. Several potentially impactful interventions have been found to reduce the severity of the illness and improve outcomes among critically ill patients with COVID-19: 1) early prone positioning (before intubation), which improves oxygenation and reduces the need for mechanical ventilation [8] [9] [10] ; 2) Remdesivir therapy, when given to those requiring supplemental oxygen, can reduce recovery time [11] ; and 3) dexamethasone, a frequent adjunctive therapy for ARDS and sepsis, has been found to reduce mortality among those with COVID-19 and respiratory compromise [12] . Those findings suggest the importance of early identification of ARDS as part of an overall COVID-19 treatment plan. Therefore, there is a critical need for a tool that can identify the patients who can benefit from earlier interventions.

In this single-center study conducted at the highest-volume clinical center for COVID -19 in the state of Illinois, we characterized the evolution of infections, risk factors for infection, and predictors of severe illness (i.e., ARDS), using a multistage analytic approach. Using LASSO regression and decision tree algorithms, we identified the key risk factors that predicted the need for mechanical ventilation from past data. Finally, we developed an early warning system for identifying those at high risk for requiring mechanical ventilation using the identified risk factors.

An important aspect of the machine learning-based model is its ability to individualize a prediction based on features that are readily available upon hospital admission.

Although recently published comprehensive studies from China address disease progression, our study is unique as it focuses on the United States. We note that work in other urban areas of the United States (i.e., New York and Los Angeles) has found strong differences in severe illness onset and outcomes across racial and ethnic groups [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] . In that context, our research focuses specifically on the population of Chicago, Illinois, which has a diverse and unique urban patient population.

The overall flow of our analysis is illustrated in Figure 1 . Our analysis included 10,123 COVID-19 patients for whom information on demographics and initial symptoms was available.

Among these patients, 1,788 were admitted to the hospital; for them, additional longitudinal information, such as lab test results and vitals, was available. Among the hospitalized patients, 301 had most of their lab measures and vitals collected on the day of admission. We used data collected from those patients to develop and validate the prediction model in a retrospective fashion. Out of the 301 investigated patients, 96 required ventilation. The predictive was developed in two steps:

(1) we used a LASSO regression and a decision tree model to identify top risk factors for mechanical ventilation; and (2) we developed a logistic regression model based on these risk factors to predict which patients would eventually need ventilatory support.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020. 11.29.20240606 doi: medRxiv preprint Data was collected at Rush University Medical Center in Chicago, Illinois, and cover the COVID-19 patients registered between 03/17/2020 and 08/07/2020. Information on patients and treatments was obtained from queries against data warehouses populated from regular exports of clinical data stored in Rush's Epic electronic medical record (EMR) system. The study included 10,123 patients with COVID-19, and their EMRs were retrospectively processed to extract the patients' demographics (Table 1) , laboratory findings, vitals (Table 2) , and comorbidities ( Table   3 ). The age of each patient was typically documented as a numerical value, but for 41 patients whose ages were documented by the text "90+", we assigned an age of 90 for use in our analysis.

The variables investigated in our study include the patients' demographics, laboratory findings, vitals, and comorbidities as collected in the hospital. Demographic variables include patients' age, sex, first race, and smoking status. Laboratory findings include white blood count, absolute neutrophil count, absolute lymphocyte count, absolute monocyte count, neutrophils percentage, lymphocyte percentage, monocyte percentage, albumin, aspartate transaminase, alanine transaminase, d-dimer, red blood count , blood urea nitrogen, creatinine, hemoglobin, ferritin, d-dimer, C-reactive protein, lactate dehydrogenase, blood glucose, platelet count, and creatine phosphokinase. Vitals include oxygen saturation (SpO2), body temperature, respiration rate, and pulse. Comorbidities include hypertension, overweight condition or obesity, type 2 diabetes, chronic kidney disease, and chronic ischemic heart disease.

We defined the stages of disease progression based on clinical events; data on such clinical events can be used by hospitals to allocate medical resources. In our study, we defined (1) hospitalization, (2) admission to ICU, (3) mechanical ventilation, and (4) death as the advancing . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint stages of COVID-19 disease progression. Figure 

LASSO regression [23] was used to select the most discriminative variables for predicting the demand for mechanical ventilation. The "glmnet" packet in R was used to compute the results [24] . A total of 18 variables were entered in an L1-norm LASSO regression. They include age, neutrophil-to-lymphocyte ratio; neutrophil, lymphocyte, and monocyte percentages; white blood, red blood, and platelet counts; and the levels of lactate dehydrogenase, C-reactive protein, blood glucose, blood urea nitrogen, creatinine, albumin, aspartate transaminase, alanine transaminase, hemoglobin, and SpO2. As the regularization term (typically denoted as λ) grows large, only the most important features are left with nonzero coefficients.

In addition to performing a LASSO regression, we also constructed a decision tree [25] for the investigated variables, which. provides another perspective that can be used to identify key risk factors. Using the Gini Index as the impurity metric [26] , the decision tree iteratively splits the current data into two branches. By definition, the variable used for splitting the root node is the most discriminative factor. Furthermore, the key risk factors tend to gather at the high-layer nodes near the root. The key risk factors selected by the LASSO regression and decision tree algorithms . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint were further compared and entered in the logistic regression model for predicting ventilator demands.

We chose to use logistic regression to build the prediction model because an easily interpretable predictive model can help facilitate rapid clinical translation [27] . The key risk factors we identified were used as predictors, and the requirement for ventilation (given as a binary Yes/No variable) was treated as the output label. The task was to predict the demand for mechanical ventilation given data collected at admission. Our positive class (ventilated patients) included 96 patients, and the negative class included 205 patients. The logistic regression classifier was trained and validated using a standard 10-fold cross-validation approach [28] . The accuracy of prediction was evaluated using the area under the receiver-operator characteristic curve (AUC).

The mean AUC and confidence intervals were calculated over all cross-validation splits.

The demographic information, symptoms, and smoking status of infected patients are presented in Table 1 . With a median age of 40 and the 75th percentile at 54, the patients were younger than those reported from China [1] and New York [6] . There were more infected females (53.55%) than males (46.45%). In terms of race, African Americans (32.92%) and Whites is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ;  day 0 to day 7. During that time, the albumin levels of the ventilated patients who ultimately survived behaved much like those patients who subsequently died. However, the albumin levels of the eventual survivors then began to increase, while the albumin levels of those who subsequently died continued to decrease until day 11. The blood urea nitrogen level of the patients who eventually died kept increasing until death and was generally higher than the level found in other groups of patients. The lymphocyte percentages of ICU-but-never-ventilated patients tended to move from the abnormal range (<18%) to the normal range, while the lymphocyte percentages of patients who eventually died tended to worsen. That correlation means that the lymphocyte percentage can be used as an indicator of a patient's condition. A rising lymphocyte percentage indicates a recovery trend in a COVID-19 patient, while a decreasing lymphocyte percentage indicates a worsening condition. As for the red blood count, the initial values for patients who later died were in roughly the same range as those of survivor groups those of survivor groups, but later the value for patients who died showed a sharper decrease compared with the survivors. Thus, the initial value for the red blood count may not be a risk factor for mortality prediction, but its downward trend can be used to forecast deterioration.

We investigated five comorbidities for COVID-19 patients. Ranked by their rates of appearance among the hospitalized patients, they include hypertension (60.29%), overweight or obese condition (57.1%), type 2 diabetes (43.12%), chronic kidney disease (23.04%), and chronic ischemic heart disease (18.06%). As COVID-19 progressed from the hospitalization stage to the ICU stage to eventual death, the percentage of patients who had chronic kidney disease or chronic ischemic heart disease constantly increased, indicating that these two comorbidities are significant risk factors for severe disease progression. In addition, the rates of hypertension and type 2 diabetes . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint were also higher among deceased patients than among the total hospitalized population. The patients who had none of these five comorbidities and presented as relatively healthy accounted for very few (14.65%) of the hospitalized patients and even fewer of the deceased patients (10.71%), meaning that an originally healthy individual is far less likely to suffer serious effects due to COVID-19. These findings indicate that the studied comorbidities are risk factors for mortality. turns to 0, meaning that albumin is the most discriminative feature selected by LASSO, followed by SpO2 and white blood count. The significance of albumin is also indicated by the magnitudes of coefficients; the coefficient of albumin is consistently larger than that of any other feature.

Albumin can be considered a general measure of an individual's overall health. Figure 5 (b) shows the trace of binomial deviance, a type of misclassification error [29] . As the regularization parameter λ becomes large, the binomial deviance first decreases and then increases. The axis above the figure shows the number of nonzero coefficients at a particular λ value. The minimal deviance is achieved with 12 features. They include age, platelet count, white blood count, neutrophil-to-lymphocyte ratio, lymphocyte percentage, and the levels of lactate dehydrogenase, C-reactive protein, blood glucose, blood urea nitrogen, albumin, hemoglobin, and SpO2.

Using the 12 features selected by LASSO regression, we achieved an AUC of 0.819 (95%CI: 0.751-0.888) using a logistic regression approach. To identify the most discriminative features, we used the decision tree algorithm with the Gini Index. Figure 6 shows the result of the decision tree with max layer = 3. The first split is made on albumin = 2.65 g/dl. For the patients . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint with albumin lower than 2.65 g/dl, the second split is made on the white blood count = 6.835 k/ul, indicating that the patients with low albumin and high white blood count are more likely to require mechanical ventilation. Whereas, for the patients with albumin higher than 2.65 g/dl, the second split is made on SpO2 = 92.739%, indicating that the patients with high albumin and high SpO2 values are less likely to need ventilatory support. The top three features selected by the decision tree algorithm exactly match those selected by LASSO regression, highlighting the significance of these features. In clinical practice, this compact set of features may be used to efficiently triage COVID-19 patients. We trained another logistic regression model with these three features alone and achieved an AUC of 0.823 (95%CI: 0.765-0.880), which is marginally higher than the AUC achieved using all 12 features.

We selected a cutpoint of the prediction model by maximizing the sum of sensitivity and specificity, which is also formally equivalent to maximizing Youden's Index [30] . Such a criterion is desirable in clinical setting for it maximizes the overall correct classification rate and thus reduce the overall misdiagnosis rate. We performed a grid search by varying the cutpoint from 0 to 1 with an interval of 0.05. The maximal sum of sensitivity and specificity was achieved at the cutpoint of 0.35, where sensitivity = 0.70 and specificity = 0.82.

This paper presents a novel approach for triaging COVID-19 patients in a hospital setting to predict the need for mechanical ventilation. We created the approach using a multi-stage perspective in a retrospective manner. Our data are unique in that they came from a large metropolitan hospital with an extremely diverse patient population. We characterized the stratification of COVID-19 patients in terms of clinical events and presented laboratory measures in both static and dynamic views, thereby making it easier to understand and track disease . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. This research and the resulting model establish the feasibility of an early triaging tool that can predict the clinical course of COVID-19 at a subject-specific level. This not only has the potential to improve patient outcomes, but also can provide hospital administrations with insights that inform resource management.

Several other previously published modeling approaches have attempted to evaluate predictors of clinical deterioration, mechanical ventilation, and death. Factors associated with disease progression in non-peer-reviewed work include lower platelet and lymphocyte counts; increased markers of DIC, such as fibrinogen, d-dimer, and aPTT; increased LDH, AST, and CK, and abnormal CT scans [31] ; clinical comorbidities, CRP, respiratory rate, and LDH [32] ; and higher SOFA score, age, and d-dimer levels [33] . One non-peer-reviewed study found similar factors and developed a nomogram for prediction [34] . In a peer-reviewed study, clinical comorbidities were the most predictive for severe disease progression [35] . Our work has identified three critical indicators of the deterioration of COVID-19 patients, namely albumin, SpO2, and white blood count, which are readily measurable in a clinical setting.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint

Future work with this model will implement an early warning system for clinical disease progression. The value of using an approach that recognizes respiratory compromise is that it can identify patients for whom early intervention can improve outcomes. With the insights gleaned from emerging clinical data, the use of optimized prone positions, medical therapy with antivirals, and anti-inflammatory medication may abbreviate the inflammatory response, improve oxygenation, reduce the risk of intubation, and reduce mortality in patients with COVID-19. An approach like the one introduced in this paper can also identify patients for whom early discharge is safe. A triage tool for sorting high-vs. low-risk individuals with COVID-19 would be highly useful in resource-constrained situations in which bed capacity must be tightly managed.

Our study is subject to several limitations. It is a single-center study, and practice patterns in that center may not generalize to other centers. In addition, we did not include radiography in the predictive model. Future studies will continue refining the model and validate its performance prospectively.

We thank Kathleen Atchley and Jenny Applequist for assisting with preparing the manuscript. This project has been funded by the Jump ARCHES endowment through the Health Care

Engineering Systems Center at the University of Illinois.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint (2) we developed a logistic regression model based on these risk factors to predict the need for ventilation support at a patient level.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint Among them, 1,788 were treated in the hospital. The patients who suffered more severe disease progression were treated in the ICU and some were ventilated, depending on the severity of their disease. 684 (38.3%) of the patients admitted to the hospital were treated in the ICU, and of those in the ICU, 380 (55.6%) received ventilation support. 155 ventilated patients died from COVID-19.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint Figure 6 Decision tree for distinguishing between patients who did and did not require ventilation.

A maximum tree depth of 3 and the Gini impurity index were utilized to derive the decision tree. The first split is made on albumin = 2.65 g/dl. For the patients with albumin lower than 2.65 g/dl, the second split is made on the white blood count = 6.835 k/ul. Whereas, for the patients with albumin higher than 2.65 g/dl, the second split is made on SpO2 = 92.739%.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 30, 2020. ; https://doi.org/10.1101/2020.11.29.20240606 doi: medRxiv preprint

A Novel Coronavirus from Patients with Pneumonia in China

WHO Director-General's opening remarks at the media briefing on COVID-19 -11

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

Viral and host factors related to the clinical outcome of COVID-19

Clinical and Epidemiological Characteristics of COVID-19 Patients in Chongqing China. Frontiers in Public Health

Clinical Characteristics of Covid-19 in

Presenting Characteristics, Comorbidities, and Outcomes Among 5700 Patients Hospitalized With COVID-19 in the

Prone Position in Management of COVID-19 Patients; a Commentary. Archives of Academic Emergency Medicine

Prone Positioning in Severe Acute Respiratory Distress Syndrome

Prone Positioning in Awake, Nonintubated Patients With COVID-19 Hypoxemic Respiratory Failure

Remdesivir for the Treatment of Covid-19 Preliminary Report

Dexamethasone in Hospitalized Patients with Covid-19

Disparities in Outcomes Among COVID-19 Patients in a Large Health Care System in California

COVID-19 Prevalence and Mortality Rates in Association with Black Race and Segregation in the United States

Social Science Research Network

Black-White Risk Differentials in COVID-19 (SARS-COV2) Transmission, Mortality and Case Fatality in the United States: Translational Epidemiologic Perspective and Challenges

Assessing racial and ethnic disparities using a COVID-19 outcomes continuum for New York State

Racial/Ethnic Disparities in Disease Severity on Admission Chest Radiographs among Patients Admitted with Confirmed COVID-19: A Retrospective Cohort Study

Hospitalization and Mortality among Black Patients and White Patients with Covid-19

Disparities in the Population at Risk of Severe Illness From COVID-19 by Race/Ethnicity and Income

Covid-19 by Race and Ethnicity: A National Cohort Study of 6 Million United States Veterans

COVID-19 and Racial/Ethnic Disparities in Health Risk, Employment, and Household Composition: Study examines potential explanations for racial-ethnic disparities in COVID-19 hospitalizations and mortality

Outcomes of COVID-19: Disparities in obesity and by ethnicity/race

Regression shrinkage and selection via the lasso

Glmnet vignette. 1-30

A survey of decision tree classifier methodology

A note on the calculation and interpretation of the Gini index

Developing prediction models for clinical use using logistic regression: an overview

A study of cross-validation and bootstrap for accuracy estimation and model selection. InIjcai

The elements of statistical learning

Maximization of the sum of sensitivity and specificity as a diagnostic cutpoint criterion

Risk assessment of progression to severe conditions for patients with COVID-19 pneumonia: A single-center retrospective study

Prognostic factors for COVID-19 pneumonia progression to severe symptom based on the earlier clinical features: A retrospective analysis. medRxiv

Machine Learning Model Reveals Older Age and Delayed Hospitalization as Predictors of Mortality in Patients with COVID-19. medRxiv

A Tool to Early Predict Severe 2019-Novel Coronavirus Pneumonia (COVID-19): A Multicenter Study using the Risk Nomogram in Wuhan and Guangdong, China. medRxiv

Neutrophils Percent (%, Low: 46.0

Urea Nitrogen (Mg/Dl, Low: 8

225 (21.99%)

Type 2 diabetes 1