key: cord-290551-a02tueuu
authors: Singh, Shailendra; Bilal, Mohammad; Pakhchanian, Haig; Raiker, Rahul; Kochhar, Gursimran S.; Thompson, Christopher C.
title: Impact of Obesity on Outcomes of Patients with COVID-19 in United States: A Multicenter Electronic Health Records Network Study.
date: 2020-08-21
journal: Gastroenterology
DOI: 10.1053/j.gastro.2020.08.028
sha: 
doc_id: 290551
cord_uid: a02tueuu

nan

During the 2009 H1N1 Influenza A virus pandemic, obesity was significantly associated with increased risk for hospitalization and mortality 1 . In 2020, the COVID-19 pandemic has a higher estimated case fatality rate 2 . It has hit the United States (U.S) at a time when obesity has also reached epidemic status, with the prevalence of obesity increased from 30.5% to 42.4%, and severe obesity increased from 4.7% to 9.2%, over the past decade 3 . Comorbidities associated with obesity are widely recognized risk factors for poor COVID-19 outcomes 4 ; however, larger population-based data evaluating obesity as an independent risk factor continues to be sparse.

We performed a retrospective cohort study using TriNetX (Cambridge, MA, USA), a global federated health research network that provided access to electronic medical records of patients from multiple large member healthcare organizations (HCOs) in United States. Details of the data source are described in the supplementary.

A search query was performed to identify all adult patients (> 18 years) with a diagnosis of COVID-19 between January 20, 2020, and May 31, 2020. The search criteria to identify potential COVID-19 patients were based on specific COVID-19 diagnosis codes or positive laboratory confirmation of COVID-19. Identified COVID-19 patients were stratified based on a body-mass index (BMI) or a diagnosis code for obesity. Patients with a documented BMI > 30 or a diagnosis of obesity within 1 year before the diagnosis of COVID-19 were included in the obesity group. Patients with a documented BMI <30 or with no documented diagnosis of obesity within the last one year were included in the control group. We excluded all patients where BMI varied between > 30 and <30 in the preceding year before the diagnosis of COVID-19 or diagnosis of obesity was present, but BMI was reported as <30 in the preceding year. Details of patient selection are outlined in Supplementary Figure 1.

Obesity group and control groups were compared after 1:1 propensity score matching (PSM). The primary outcome was a composite of intubation or death up to 30 days after diagnosis of COVID-19. Sensitivity analysis and subgroup analysis based on the obesity class were also performed. Details of the statistical analysis, sensitivity analysis, and limitations are also provided in the Supplementary.

A total of 41,513 COVID-19 adult patients from 26 HCOs in the United States were identified. Out of these COVID-19 patients, 8,641 patients with documented BMI >30 (n=5879) or diagnosis of obesity (n=2762) were included in the Obesity group, and 31,273 patients with BMI<30 (n=6437) or without any reported diagnosis of obesity were all included in the control group (Supplementary Figure 1) . Gender, racial and ethnic differences were seen between the groups, and patients in the obesity group had a significantly higher proportion of comorbidities compared to the control group (Table 1 ). In the crude, unadjusted analysis, patients in the obesity group were more likely to have a 30-day composite outcome of death or mechanical ventilation compared to the control group (RR 1.99, 95% CI 1.84-2.15).

After PSM, a relatively balanced cohort of obese and non-obese patients were obtained (N=8112 patients in each group) ( Table 1 ). The risk of composite outcome was higher in the obesity group compared to the control group (RR 1.56, 95% CI 1.41-1.73). Kaplan Meier survival analysis showed that the cumulative probability of being composite event-free up to 30-days remained significantly lower in the obesity group than the control group (87.7% vs. 90.5%, P-log rank <0.0001) (Supplementary Figure 2) . The risk of mortality, intubation, and hospitalization was higher in obesity group compared to the control group in the matched cohort (Table 1 ). In a propensity-matched sub-group analysis based on obesity class, the risk of composite outcome, and other poor outcomes was highest in patients with obesity class 3 (Table 1 ). The results of the sensitivity analysis confirmed the robustness of our main findings (Supplementary).

Our study using a large nationally representative database showed that COVID-19 patients with any degree of obesity had a significantly higher risk of hospitalization and intubation or death compared to patients without obesity. A substantial incremental risk of intubation or death in the obesity cohort persisted even after meticulous PSM to adjust for confounding comorbidities. Patients with severe obesity were at highest risk of these poor outcomes.

The COVID-19 pandemic has exposed the delivery of healthcare in the U.S and has provoked a reckoning regarding our healthcare model moving forward. The U.S. obesity epidemic has continued to grow for decades without any signs of abating. Obesity and its associated comorbidities are now a significant determinant of COVID-19 outcomes 5 in a population where over 90 million adults have obesity and are highly susceptible.

Disproportionate prevalence of obesity and associated comorbidities probably also have played a significant role in the racial and ethnic disparities seen during the COVID-19 pandemic. The obesity cohort derived from our data source showed a higher proportion of African Americans and Hispanics in the obesity group. Obesity increases the risk of poor outcomes in this vulnerable population with limited access to health care. Advanced age and male gender are major risk factors for worse prognosis and higher mortality in COVID-19 patients 6 .

However, a larger proportion of patients with obesity in our cohort were females, and the impact of this can be dramatic enough to shift severe COVID-19 outcomes towards females. Similarly, a large number of younger patients with obesity are also affected by severe COVID-19 with poor outcomes. In the U.S., where obesity is an epidemic, its impact is not only limited to clinical outcomes. Along with the psychosocial impact of social distancing and quarantining that is applicable to the entire society, persons with obesity must contend with "weight stigma." Derogation of persons with obesity is not uncommon and, unfortunately, more socially acceptable than other marginalized groups 7 . These biases and behaviors are not limited to the general public, J o u r n a l P r e -p r o o f and studies have shown that many healthcare workers can also have negative attitudes and stereotypes about persons with obesity 8 .

Our findings highlight the need for a vast improvement in the care of patients with obesity during this pandemic and moving forward. Physicians should manage COVID-19 patients with obesity aggressively as outcomes can be significantly worse than in the general population. In the long-term, preparing for future pandemics or if COVID-19 becomes seasonal, there is also a serious need to develop and implement weightloss strategies. There is a necessity for more healthcare professionals, including gastroenterologists, to play a central role in caring for patients with obesity. J o u r n a l P r e -p r o o f 

TriNetX (Cambridge, MA, USA) is a global federated health research network providing access to electronic health records (EHR) of patients from 34 large member healthcare organizations (HCOs) in United States. COVID-19 data was incorporated in TriNetX using specific diagnosis and terminology following the World Health Organization (WHO) and Centers for Disease Control (CDC) COVID-19 criteria. Real-time access to HIPPA compliant de-identified longitudinal clinical data to member HCOs is provided on a cloud-based platform. A typical HCO is a large academic health center with data coming from the majority of its affiliates. In addition to EHR data available in a structured fashion (e.g., demographics, diagnoses, procedures, medications, lab test results, vital signs), TriNetX can also extract facts of interest from the narrative text of clinical documents using Natural Language Processing. Data is mapped to a standard and controlled set of clinical terminologies and transformed into a proprietary data schema. This transformation process includes an extensive data quality assessment to reject records that do not meet quality standards.

TriNetX data has been granted a waiver from the Western IRB because it is a federated network and only aggregate counts and statistical summaries of the de-identified information without any protected health information received from participating HCO's. Both the patients and HCO's as data sources remain anonymous.

The software checks the basic formatting to ensure, for example, that dates are appropriately represented. It enforces a list of fields that are required (e.g., patient identifier) and rejects those records where the required information is missing. Referential integrity checking is done to ensure that data spanning multiple database tables can be successfully joined together. As the data is refreshed, the software monitors changes in volumes of data over time to ensure data validity. TriNetX requires at least one non-demographic fact for a patient to be counted in our dataset. Patient records with only demographics information are not included in datasets.

Demographics are coded to HL7 version 3 administrative standards. Diagnoses are represented by ICD-10-CM codes. If an HCO provides data in ICD-9-CM, the data source uses a 9-to-10-CM mapping based on GEMs plus custom algorithms & curation to transform data from ICD-9-CM to ICD-10-CM. Diagnoses data are enriched with the Chronic Condition Indicator (CCI). Depending on the coding system used by an HCO, procedure data coded in ICD-10-PCS or CPT. For many procedures, both ICD-10-PCS and CPT codes are added to a query to define a cohort. Medications are represented at the level of ingredients, coded to RxNorm, and organized by NDF-RT Therapeutic Classes. Lab results, vitals, and findings are coded to LOINC. To ease finding and using common labs, LOINC codes are combined up to clinically significant levels for most frequent labs and coded as TNX: LAB.

The search was conducted following the CDC COVID-19 coding guidance. These codes included 2) were excluded to reduce any false positive COVID-19 patients because this ICD-9 code can still be used occasionally as "catch-all' code for more than 50 viral infections.

In addition to the ICD codes, following LOINC codes with positive laboratory results were also used to identify COVID- 

TriNetX has the capability of analyzing data based on a temporal relationship to the index event.

The index event in our study was defined as the diagnosis of COVID-19. Baseline characteristics were estimated from any time before the index event. Presenting laboratory values, and medications were recorded from the time of index event up to two weeks before the index event.

Outcomes were assessed from the index event up to 30 days after the index event. The risk for intubation (mechanical ventilation), hospitalization, and mortality after diagnosis of COVID-19 was recorded. The primary outcome was a composite of intubation or death.

regression was performed in Python 3.6.5 using standard libraries numpy and sklearn. The same analyses were also performed in R 3.4.4 software to ensure the matching of outputs. After the calculation of propensity scores, matching was performed using a greedy nearest-neighbor matching algorithm with a caliper of 0.1 pooled standard deviations. The order of the rows in the covariate matrix can impact the nearest neighbor matching; therefore, the order of the rows in the matrix was randomized to eliminate this bias. For each outcome, the risk ratio with a 95% confidence interval was calculated to compare the association of obesity with the outcome. Kaplan-Meier survival analyses were used to estimate the survival probability of composite outcome at the end of 30 days following the index event. Patients were censored when the time window ended or on the day after the last fact in their record. Hypothesis testing for Kaplan-Meier survival curves was conducted using the log-rank test. A-priori defined two-sided alpha of less than <0.05 was used for statistical significance.

Selection bias in the obesity group and the control group was possible. Therefore, we performed a sensitivity analysis by varying the inclusion criteria. We first included all patients with a diagnosis of obesity in their health records at any time before COVID-19 diagnosis and compared them to a cohort of patients with no record of obesity. Secondly, we compared patients with a diagnosis of obesity in the last three months and one month to a cohort of patients with no reported obesity. Additional sensitivity analyses included the same set of main analyses but also adjusting for medications (angiotensin-converting enzyme inhibitors (ACEI) or angiotensin receptor blockers (ARB)) and presenting laboratory values (ferritin, CRP and LDH). Finally, given the possibility that poor outcomes in patients with obesity might be higher at presentation or related to late presentation and access to health care, we performed an analysis excluding the composite outcomes in the first two days after diagnosis.

An analysis of a larger group of selected patients using diagnostic criteria of obesity as any time before the index event (after PSM n=9769) showed a higher risk for composite outcomes in the obesity group ( J o u r n a l P r e -p r o o f

We acknowledge the limitations due to the retrospective nature of the study. The data derived from EHR based database is susceptible to errors in coding when patient information is translated into codes. However, extensive data quality assessment that includes data cleaning and quality checks minimizes the risk of data collection errors at the investigator's end. Adjustments for missing data is not currently possible on TriNetX platform. Cases of COVID-19 could have been misdiagnosed as other cases of Pneumonia or viral infections due to diagnosis or coding errors, especially early in the pandemic. We likely missed patients who were asymptomatic or had mild disease and did not seek medical attention; therefore, our cohort may represent the more severe spectrum of the COVID-19 disease. Data on exposure history, incubation time, and dynamic changes in patients' clinical condition could not be estimated from the EHR database. Socioeconomic and structural determinants, psychological elements, geographical factors, and health care delivery during COVID-19 could have impacted the care of patients with obesity but were beyond the scope of our study. Despite these limitations, our study uses a large national database to evaluate the impact of obesity in COVID-19 patients. Given our study population is representative of multiple centers across the U.S., the results are more generalizable than singlecenter or regional experiences. In addition, even though our study was not randomized, we performed a robust statistical analysis using propensity score matching. 

Morbid obesity as a risk factor for hospitalization and death due to 2009 pandemic influenza A(H1N1) disease. PLoS One

Assessment of Deaths From COVID-19 and From Seasonal Influenza

Prevalence of Obesity and Severe Obesity Among Adults: United States

Clinical course and risk factors for mortality of adult inpatients with COVID-19

China: a retrospective cohort study

Obesity is associated with worse outcomes in COVID-19: Analysis of Early Data From New York City

Clinical features of patients infected with 2019 novel coronavirus in Wuhan

CPT: Intubation, endotracheal, emergency procedure) OR "1015098" (CPT: Ventilator management) OR "5A1935Z" (ICD10: Respiratory Ventilation, Less than 24 Consecutive hours) OR "5A1945Z" (ICD10: Respiratory Ventilation, 24-96 Consecutive hours) OR "5A1955Z" (ICD10: Respiratory Ventilation, Greater than 96 Consecutive hours) OR "0BH17EZ" (ICD10: Insertion of Endotracheal Airway into Trachea, Via Natural or Artificial Opening) OR 0BH18EZ (ICD10: Insertion of Endotracheal Airway into Trachea, Via Natural or Artificial Opening Endoscopic) OR 0BH13EZ (ICD10: Insertion of Endotracheal Airway into Trachea

Statistical Analysis All statistical analyses were performed in real-time using TriNetX. The means, standard deviations, and proportions were used to describe and compare patient characteristics. Categorical variables were compared using the Pearson chi-square test and continuous variables using an independent-samples t-test. We performed a 1:1 propensity score matching (PSM) to reduce the effects of confounding. Covariates included in the propensity score model included age, race, ethnicity, dyslipidemia, diabetes mellitus, chronic lower respiratory diseases (Chronic obstructive pulmonary disease (COPD) and Asthma), ischemic heart diseases, heart failure, pulmonary heart diseases, cerebrovascular diseases, chronic kidney disease, fatty liver, cirrhosis of liver, malignant neoplasm, and nicotine use (Table 1)

J o u r n a l P r e -p r o o f Codes used for patient characteristics included in the propensity score matching.

Coding system and Codes