key: cord-0987605-6kuefmol
authors: Experton, B.; Tetteh, H.; Lurie, N.; Walker, P.; Carroll, C.; Elena, A.; Hein, C.; Schwendiman, B.; Burrow, C.
title: A Multi-Factor Risk Model for Severe Covid-19 for Vaccine Prioritization and Monitoring Based on a 15 Million Medicare Cohort
date: 2020-11-03
journal: nan
DOI: 10.1101/2020.10.28.20219816
sha: 114922fb39554eff6c001c133df386668764164f
doc_id: 987605
cord_uid: 6kuefmol

Background: Public Health interventions to slow the spread of the Covid-19 pandemic focus on protecting individuals at risk for severe disease. Risk categorization is essential to effective pandemic response. However, existing risk models for severe Covid-19 lack needed integration of both socio-demographic and clinical risk factors, and geographic characteristics. Methods: We present an integrated multi-factor risk model for severe Covid-19 using de-identified Medicare claims from which we extracted demographic and clinical data for a cohort of 15 million Medicare beneficiaries with 770,000 Covid-19 cases, and socio-economic data at the county and zip code level from the CDC Social Vulnerability Index. The model and associated digital maps were developed as part of Project Salus of the Department of Defense Joint Artificial Intelligence Center, for use by the National Guard and other military personnel in their support mission to hospitals and local jurisdictions impacted by the pandemic. Results: The model affirms ethnicity (Black: OR 1.64; 95% CI 1.61-1.68, American Indian: OR 2.21; 95% CI 2.01-2.42), age over 85 (OR 1.75, 95% CI 1.69-1.81), the socio-economic factor of residing in a zip code in the lowest quartile of income (OR 1.23; 95% CI 1.21-1.26), ESRD (OR 2.35; 95% CI 2.25-2.45) and chronic lung disease (OR 1.95; 95% CI 1.90-2.00) as leading risk factors for Covid-19 hospitalizations, but reveals low risk for COPD (OR 1.15; 95% CI 1.13 -1.17) and minimal or no risk for diabetes (OR 1.03; CI 1.01-1.05), CHF (OR 1.10, 95% CI 1.08-1.12) or hypertension (OR 0.96; 95% CI 0.94-0.98), and demonstrates an association between prior herpes zoster immunization (OR 0.74; 95% CI 0.71-0.77), and to a lesser degree prior influenza and pneumococcal vaccines with less severe Covid-19. Conclusions: This multi-factor risk model and derived digital maps can be applied for use by national and local health authorities to augment existing tools for pandemic response, including monitoring of post Covid-19 sequelae, prioritization of Covid-19 vaccine, and vaccine monitoring for both safety and efficacy.

Over eight million people in the United States have been infected with SARS-CoV-2, with more than 225,000 deaths of whom 80% are over the age of 65. 1, 2 Public health interventions to slow the spread of the pandemic have focused on protecting individuals most at risk for exposure to SARS-CoV-2, or at risk for severe Covid-19. The Centers for Disease Control (CDC) has identified people of color, individuals with certain comorbidities and those over the age of 65 at highest risk. 3 This risk-based approach to pandemic response is now directing Covid-19 vaccine prioritization to allocate the first available vaccine doses to individuals most at risk. All Covid-19 vaccine prioritization frameworks reviewed by the Advisory Committee for Immunization Practices (ACIP) are risk-driven with a time-phased approach. 4 The National Academies of Science, Engineering and Medicine recommend a four-phase equitable plan determined by risk categories. 5 The "jumpstart phase 1a" allocates the first vaccine doses to front line healthcare professionals and emergency responders, followed by phase 1b for high risk individuals with comorbidities or older individuals living in congregate or overcrowded settings. In addition, the National Academies recommend special efforts to deliver vaccines to residents of high-vulnerability areas by using the CDC's Social Vulnerability Index (SVI) 6 or the Covid-19 Community Vulnerability Index (CCVI). 7 Multiple-factor risk categorization is needed for vaccine allocation schemes to simultaneously include individual demographics, comorbidities, and socioeconomic characteristics. Integration of these various data should also consider overlap of risk categories, as the CDC highlighted during the ACIP October meeting. 8 This is important when considering that more than half of the 53 million Americans over age 65 suffer from two or more chronic conditions, 9 qualifying them for phase 1b vaccine allocation.

At a population level, both clinical and sociodemographic risk need to be considered in tandem. The current CDC listing of risk factors for severe Covid-19 10 is derived from single hospital-based studies with limited sample sizes, 11, 12 or hospital reporting, both of which lack nationwide representation. 13 The Center for Medicare and Medicaid Services (CMS) issues monthly Medicare Covid-19 Data Snapshots that include national demographic characteristics and prevalence of common chronic conditions of hospitalized fee-for-service (FFS) Medicare beneficiaries, but lack more detailed diagnosis, medication, and procedure data as well as socio-economic data needed for identifying risk factors for severe Covid-19 at the individual and population levels. 14 There are no analyses that fully support operationalization of the National Academies recommendation to use both clinical and SVI or CCVI risk factor data. The CCVI provides an integrated risk model of both CDC identified risk factors for severe Covid-19 and SVI data, at the state and county, but not at the local zip code level. This and other models of this type, such as the Pandemic Vulnerability Index, 15 have further limitations. They offer population (not individual) based risk data, include limited clinical data with certain chronic conditions with their geographic distributions, and use dissociated geographic data sources (national data for comorbidities, and census tract level data for SVI socio-economic risk factors).

This study integrates multiple factors in its risk models for severe Covid-19 in the Medicare population, based on de-identified Medicare claim data from which detailed clinical and past medical history data were extracted, the CDC SVI and other public datasets. The risk models were developed in part based on prior analysis of Medicare patient outcomes in the frail elderly population, 16 and the Medicare claim data processing conducted by Humetrix with our CMS approved technology as part of the CMS Blue Button program since 2011. 17 The resulting risk models are both individual and population based and has been used to create an interactive dashboard currently hosted on a secure government intranet, enabling visualization of calculated population risk at national, county and zip code levels.

The risk models and associated mapping were developed as part of a Department of Defense (DoD) Joint Artificial Intelligence Center (JAIC) Covid-19 related data analytics project named Salus (the Roman goddess for safety and health), to provide predictive visualization tools to the National Guard and other military personnel in charge of resourcing Covid-19-related healthcare and other vital sector supply and personnel assets, especially serving local hospitals and health departments affected by the pandemic. 18 Following on the National Academies recommendation for use of existing systems across all levels of government, we discuss the possible use of the Salus Medicare risk models and its derived dashboard as a tool for supporting equitable allocation, distribution of Covid-19 vaccine and vaccine monitoring, as well as for Covid-19 surveillance in the higher risk Medicare population.

The Salus Medicare risk models are based on an observational study of all FFS Medicare beneficiaries who since January 1, 2020, either had a Covid-19 test or diagnosis, or for any medical reason had an emergency department, urgent care, or telehealth visit, or were hospitalized. The CMS Office of Enterprise Data and Analytics provides for this cohort weekly Medicare claim outputs from the CMS Chronic Condition Warehouse with Project Salus specified outpatient and inpatient institutional, Part B professional, skilled nursing facility, and hospice claims extending back to October 1, 2019, and Part D claims starting January 1, 2020. De-identified claims are received in the ECS Federal Secure Unclassified Network (SUNet) for processing and analysis by Salus partner Humetrix to build risk models for severe Covid-19. In addition to the clinical and demographic data generated from Medicare claims, the models used selected socio-economic variables from the CDC SVI.

The Humetrix Enterprise Platform with its claim and other health data analytics software provides weekly data processing of over 100 million CMS de-identified Medicare claim records in its secure JAIC SUNet enclave, including data upload into a database to generate Covid-19 outcome datasets for: Covid-19 confirmed cases and related hospitalizations, ICU and ventilator use, outpatient care only, and death. Medicare claim-generated clinical, demographic and outcome variables are then merged with CDC SVI variables for the residential zip codes of individuals in the cohort.

The independent variables included in our severe Covid-19 risk models are: beneficiary age, sex, ethnicity, insurance coverage and residential zip code, prior health care utilization (prior hospitalization(s), skilled nursing home admissions, etc.) as a measure for disease severity and frailty, the individual's multiple chronic conditions with in addition to the CMS chronic condition flags, Humetrix compiled diagnostic categories using specific ICD-10 code algorithms, medications grouped by pharmaceutical class, vaccinations before Covid-19 diagnosis, and other variables starting October 1, 2019 (see Supplemental Methods in the Appendix). Socio-economic variables (e.g., income quartile, education, residential density, and other factors) were defined at the individual residential zip code level, after conversion from the census track based data found in the CDC Social Vulnerability Index.

To determine significant predictors of Covid-19 related hospitalization, ICU admission and death outcomes, we used logistic regressions (R statistical software, version 3.6 with rms, glmnet and pROC packages). For these regressions, we defined the following groups of Covid-19 cases: those who received All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint outpatient care only (defined as cases that did not require hospitalization or didn't die at least thirty days after diagnosis), hospitalizations attributed to Covid-19, and deaths (defined as cases who died of SARS-CoV-2 infection within 60 days of diagnosis).

The logistic regression models were built on 60% training and 40% validation sets obtained by random selection of Covid-19 confirmed cases. We used a 50:50 ratio of Covid-19 cases who required either hospitalization, or who only received outpatient care for training the hospitalization model, and a 50:50 ratio for Covid-19 survivors versus deceased cases, for training the death model. Determination of correlation coefficients between independent variables, as well as lasso regression, were used to eliminate independent variables which demonstrated significant correlations and collinearities in both models. A stepwise backward variable selection procedure based on the Akaike Information Criterion (AIC) was also used to remove non-significant variables, verifying that independent variables discarded by the lasso regression corresponded with variables identified by the stepwise backward selection procedure. Computation of the 95% confidence intervals for the coefficient estimates were generated by bootstrapping (2000 repetitions) on the training set.

Model performance was measured by the Area Under the Receiver Operating Characteristics (AUROC) curve value. The hospitalization risk model AUROC was computed on a validation set composed of cases that were not used in the training set, with adjustments to give a 60:40 ratio of outpatient to hospitalized cases (which is the observed ratio for hospitalization in the Covid-19 study population). For the death model AUROC we used an 85:15 ratio (which is the observed case fatality ratio in the Covid-19 study population).

Logistic regression coefficients were used to compute individual predicted probabilities of hospitalization in the event of SARS-CoV-2 infection for the entire Salus Medicare cohort of 15 million beneficiaries. The regional percentage of the cohort population over a predicted probability of 0.55, is displayed on a digital risk map at the county and zip code levels. Additional mapping data was also produced to display Covid-19 confirmed cases and their outcomes at the national and local level, with recent, cumulative and times series displays.

Socio-demographic and clinical characteristics of the 15 million study population, as of September 16, 2020, and its subsets of Covid-19 cases (n=770,000) who were either hospitalized with severe disease (n=264,000) or who only required outpatient Covid-19 care (n=299,000), or who died (n=105,000) are summarized in Table 1 . Table 1 displays comparisons between non-SARS-CoV-2 infected FFS Medicare beneficiaries and Covid-19 cases, and among them, hospitalized vs. non-hospitalized and deceased cases. Among Covid-19 cases, the severe disease groups of hospitalized and deceased patients show higher frequencies of diabetes, COPD, ESRD, hypertension, ischemic heart disease, cerebrovascular disease, chronic kidney disease, chronic lung disease, chronic liver disease, and congestive heart failure although the effect sizes for these differences are small (see Supplemental Appendix).

We used logistic regression analysis to determine the relative importance of significant predictors of severe Covid-19. The predictor variables for hospitalization are shown on Figure 1 and for death in Figure 2 .

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. The regression analyses indicate that prior herpes zoster immunization (with the recombinant GlaxoSmithKline Shingrix ® vaccine) is also associated with lower odds of Covid-19 hospitalization or death (hospitalization OR 0.74; 95% CI 0.71-0.77; death OR 0.71; 95% CI 0.66-0.77). Use of this vaccine was not correlated with any other demographic, socioeconomic or other variable in the model. We observed a smaller association of prior influenza vaccination on hospitalization (OR 0.89; 95% CI 0.88-0.91), but none against death. Pneumococcal vaccination is associated with reduced odds of death (OR 0.87; 95% CI 0.84-0.91), as recently proposed, 19 but not hospitalization.

To test the hypothesis that these vaccinations simply represent preventive behaviors that are also associated with less hospitalization and death, we created a preventive behavior variable based on the use of preventive screenings (colonoscopy and mammography) as well as vaccinations. Indeed, these preventive behaviors were associated with lower odds of developing severe Covid-19 (hospitalization OR 0.73; 95% CI 0.71-0.74, death OR 0.65; 95% CI 0.63-0.67). However, the use of Shingrix was poorly correlated with these other behaviors (preventive screenings and other vaccinations), and vaccination with Shingrix retained a significant reduction in odds ratio, even when preventive behaviors were accounted for in the models. This raises the intriguing hypothesis that this recombinant adjuvanted vaccine may have an unrecognized biological effect which reduces the probability of severe Covid-19.

The hospitalization model logistic regression coefficients were used to calculate the predicted probability for hospitalization in the event of SARS-CoV-2 infection, at the individual level, for the entire 15 million cohort. We mapped areas in which the percentage of this cohort with a predicted probability of hospitalization was greater than 0.55 for every residential zip code in the U.S. Figure 3 shows a portion of the Los Angeles metropolitan area with zip codes displaying a wide range of population percentages over this threshold which were positively correlated with the cumulative percentage of Covid-19 patients requiring hospitalization in these zip codes (Pearson correlation coefficient 0.65 p < 0.001). We further conducted the same correlation analysis in a set of zip codes drawn from other metro areas shown in Figure S1 in the Appendix. All correlation coefficients in these regions (Washington D.C., 0.61; Houston, 0.66; Miami, 0.72; Phoenix, 0.72; and New York City, 0.52) had p values < 0.001. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint

The severe Covid-19 risk models and the mapping of their outcomes are based, to our knowledge, on the largest Covid-19 dataset assembled to date for this purpose. All members of the study cohort are active users of healthcare services, presenting on average a frailer clinical profile than the general Medicare population, 20 as shown by their clinical characteristics in Table 1 . While our observed 770,000 Medicare FFS cases represent a little more than 10% of the total number of Covid-19 cases in the U.S., they contribute close to 50% of all CDC estimated Covid-19-related hospitalizations and a majority of Medicare hospitalizations. 21 The scale of this dataset has allowed us to quantitively and qualitatively validate our hospitalization risk model with actual hospitalizations in multiple geographic areas.

The countrywide distribution, and local representation of the Salus Medicare cohort also enabled Project Salus to offer the first county and zip code level mapping for this population at higher risk for severe Covid-19 and related hospitalizations. Beyond its use for epidemic mapping, the Salus Medicare Covid-19 platform provides a disease surveillance tool for Covid-19, which can be extended to influenza to evaluate the potential compounding effect of these two respiratory diseases on hospitalizations as influenza undergoes its seasonal progress in the coming months.

Our logistic regression analysis, as applied to this very large Medicare population, allowed us to test the current CDC listing of risk factors for severe Covid-19. Our analysis affirms the critical risk factors of ethnicity and older age but reveals the lack of or more modest effect of hypertension, diabetes, COPD, asthma, and cardiovascular diseases.

The new findings of the significant and independent association of less severe Covid-19 with prior use of a recombinant herpes zoster vaccine calls for further investigation. If the effect of herpes zoster vaccination is confirmed, use of this vaccination may be even more beneficial than previously recognized. The independent association of prior use of preventive screenings, a possible marker for a higher rate of Covid-19 preventive behavior (social distancing, hand washing, mask wearing), with less severe Covid-19, is a reminder of the importance of such behaviors.

In preparation for a Covid-19 immunization campaign, the models we have developed integrate both socio-economic factors and individual clinical data, and thus respond to the recommendations of the National Academies for prioritization and allocation of Covid-19 vaccines when they will become available. They could be used by the Medicare program, in collaboration with local health officials to affirmatively invite or encourage highest risk beneficiaries to seek early vaccination and can be used as a tool for vaccine allocation. In Figure 4 , we show on a histogram the distribution of the predicted probabilities of hospitalization for SARS-CoV-2 infected patients in the 15 million Salus Medicare population which can identify priority groupings for Covid-19 vaccination, according to the quantities of vaccine doses available. Further, once receipt of vaccination is linked to Medicare claims, the system could be used to support post-licensure pharmacovigilance and effectiveness studies. These are of paramount importance, especially in the early phases of a vaccination campaign.

To conclude, we believe that this quantitative analysis of risk for severe Covid-19 provides important new insights which should be considered before finalization of ACIP Covid-19 vaccine recommendations. State and local governments could consider asking the DoD, through the Mission Assignment process, to provide the Salus Medicare risk mapping to their jurisdictions. The data rich Salus Medicare dataset, which underscores the value of Medicare claim data for epidemiologic surveillance, with its size and nationwide representation, can also augment both the ILINet and COVID-NET disease surveillance systems and complement the existing vaccine monitoring systems for tracking both the safety and efficacy of Covid-19 vaccination in the high risk Medicare population. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. The views expressed in this article are those of the authors and do not necessarily reflect the official policy or position of the Department of Defense, nor the U.S. Government. Authors are military service members (or an employee of the US Government). This work was prepared as part of their official duties. Title 17 USC §105 provides that copyright protection under this title is not available for any work of the US Government. Title 17 USC §101 defines a US Government work as a work prepared by a military service member or employee of the US Government as part of that person's official duties.

The independent variable odds ratios were determined by binary logistic regression analysis of confirmed Covid-19 cases that required hospitalization for the disease and for those that were managed with outpatient care only. In addition to the thirty nine variables shown, HIV (OR 0.90; 95% CI 0.83 -0.96), a past history of anemia (OR 0.87; 95% CI 0.86 -0.89); a prescription for Azithromycin overlapping the date of Covid-19 diagnosis (OR 1.18; 95% CI 1.15 -1.22), Medicare-Medicaid dual coverage and the Medicare defined ethnicity of "Other" were included in the model based on the criteria described in the Methods section. Variables excluded from the model on the basis of these criteria (or for a regression coefficient p value of > 0.05) included prescriptions for Chloroquine or Hydroxychloroquine overlapping the time of Covid-19 diagnosis; the Medicare defined "Unknown" ethnicity, a history of ischemic heart disease, cerebrovascular disease, breast cancer, endometrial cancer; or a history of prescribed immunosuppressive drugs, Histamine type-2 receptor antagonists, inhaled corticosteroids or opioids. The AUROC for the model was 0.65. The independent variable odds ratios were determined by binary logistic regression analysis of confirmed Covid-19 cases that survived and those that died within 60 days of Covid-19 diagnosis. In addition to the thirty nine variables shown, a prescription history which included immunosuppressive drugs (OR 0.93; 95% CI 0.89 -0.98), and the Medicare defined "Other" ethnicity were included in the model based on the criteria described in the Methods section. Variables excluded from the model on the basis of these criteria (or for a regression coefficient p value of > 0.05 ) included the Medicare defined ethnicity of "Unknown"; a history of prescribed histamine type-2 receptor antagonists, anticoagulants, inhaled corticosteroids or opioids; a past history of endometrial cancer, colorectal cancer, breast cancer, hypertension, ischemic heart disease, acute myocardial infarction in 2019 or HIV. The AUROC for the model was 0.71. All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint

Medicare claim data processing; The Humetrix SaaS platform installed in the SUNet secure classified government network provides automated pre-processing of weekly updates downloaded to the Humetrix SUNet enclave of Medicare Part A inpatient and outpatient, Hospice and SNF, Part B Carrier claims and Part D (PDE) claims data to generate output files containing derived variables used to train logistic regression models as described below. The platform's medical terminology service includes a complete collection of AMA CPT-4, FDA NDC, NNPES NPI, ICD-10-CM, CMS Level II HCPCS, and NLM RxNorm codes and provides automated identification grouping of ICD-10-CM codes to identify chronic condition categories and NDC drug code to RxNorm ingredient code mappings to identify pharmaceutical classes of active pharmaceutical ingredients from Medicare claim data.

Demographic and Coverage Variables: weekly updates of CMS "Master Beneficiary Summary 2020 File" (MBSF_2020) data were processed, extracting variables: ORIG_REASON_FOR_ENTITLEMENT (Disability), ZIP_CD (residential zip code), YOB, SEX_CODE, RACE_CODE, and DUAL_STUS_CD 01-12 (Dual Medicare-Medicaid insurance).

Social Vulnerability Index (SVI) variables associated with zip codes: these were derived from CDC data, which are categorized by census tract. The variables analyzed using binary logistic regression to determine significant predictor variables for hospitalization or death due to Covid-19 included the following SVI variables: EPL_PCI, EPL_POV, EPL_NOHSDP, RPL_THEME1, EPL_CROWD, EPL_GROUPQ, EPL_MUNIT, EPL_THEME4, RPL_THEMES.

Chronic Condition Summary Variables: Chronic conditions in the MBSF_2019 chronic condition segment file was extracted, including variables: or 10 days before the Covid-19 diagnosis date, or by finding Part A Inpatient claims where the data of admission no more than 14 days after or 10 days before the Covid-19 diagnosis date. 2. Covid-19 cases managed by outpatient care only were defined as cases who were not hospitalized in the 30 days after the Covid-19 diagnosis date and who did not die within 60 days of the diagnosis date. 3. Deaths due to Covid-19 were defined as either deaths which occurred during a Covid-19 hospitalization using the inpatient PTNT_DSCHRG_STUS_CD code, or a death which occurred within 60 days of the Covid-19 diagnosis as reported in the DEATH_DT field of the MBSF_2020 file.

Effect size calculation ( All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint Figure S1 : Metropolitan region risk maps for severe Covid-19 (A) Los Angeles County, (B) New York City, (C) Houston, (D) Miami Dade, (E) Phoenix Arizona, (F) Washington DC. In each metropolitan region, zip codes with the higher predicted probabilities of hospitalization with Covid-19 based on the logistic regression hospitalization model are shown in darker shades of lavender. Circles denote observed cumulative hospitalizations due to Covid-19 extracted from claims data with the size of the circles scaled to the number of hospitalizations. The percentage of cases requiring hospitalization is displayed by the color of the circles on a beige to dark orange-red color scale with the darker circles indicating higher zip codes with higher percentages of cases requiring hospitalization for Covid-19.

All rights reserved. No reuse allowed without permission.

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted November 3, 2020. ; https://doi.org/10.1101/2020.10.28.20219816 doi: medRxiv preprint

People at Increased Risk

ACIP COVID-19 Vaccines Work Group

National Academies Release Framework for Equitable Allocation of a COVID-19 Vaccine for Adoption by HHS, State, Tribal, Local, and Territorial Authorities

Surgo Foundation. The COVID-19 Community Vulnerability Index (CCVI

Centers for Disease Control and Prevention. ACIP COVID-19 Vaccines Work Group

Preliminary Estimates of the Prevalence of Selected Underlying Health Conditions Among Patients with Coronavirus Disease 2019 -United States

COVID-19 in older people: a rapid clinical review

Hospitalization Rates and Characteristics of Patients Hospitalized with Laboratory-Confirmed Coronavirus Disease 2019 -COVID-NET, 14 States

Centers for Medicare & Medicaid Services. Preliminary Medicare COVID-19 Data Snapshot

The COVID-19 Pandemic Vulnerability Index (PVI) Dashboard: Monitoring county-level vulnerability using visualization, statistical modeling, and machine learning

The impact of payor/provider type on health care use and expenditures among the frail elderly

AI in Defense DoD's Artificial Intelligence Blog. The JAIC Forges Ahead

Possible Cross-Reactivity between SARS-CoV-2 Proteins, CRM197 and Proteins in Pneumococcal Vaccines May Protect Against Symptomatic SARS-CoV-2 Disease and Death

Prevalence of Chronic Conditions among Fee-for-Service Beneficiaries

COVID-19 Laboratory Confirmed Hospitalizations

AMI) 2. chronic kidney disease (CHRONIC_KIDNEY_EVER)

diabetes mellitus (DIABETES_EVER)

ischemic heart disease (ISCHEMICHEART_EVER)

stroke/transient ischemic attack (STROKE_TIA_EVER)

CANCER_LUNG) 11. recent endometrial cancer

hypertension (HYPERT_EVER)

End Stage Renal Disease (ESRD) was identified by finding a "Y

Humetrix SaaS platform identification of additional chronic conditions: The following chronic condition variables were identified by analyzing ICD10 codes in these CMS claim files: (Beneficiary Part A institutional Inpatient and Outpatient claims, Part B Carrier claims, SNF and Hospice claims) from October 1

chronic liver disease (includes alcoholic cirrhosis, primary biliary cirrhosis, chronic viral hepatitis due to hepatitis B and C, alcoholic fatty liver

HIV/AIDS 5. organ transplant (includes following transplants: lung, bone, heart, liver, pancreas, intestine, kidney, bone marrow

BMI over 40

Note coding of non-chemotherapy drugs as "True" signifies at least one applicable prescription fill was identified in 2020. These variables were all identified by mapping NDC drug product codes to RxNorm ingredient codes for these pharmaceutical classes of drugs: 1. chemotherapy: a value of "True" signifies that a beneficiary at any time in 2020 either had an ICD10 code for chemotherapy in part A Institutional or Part B Carrier claims, had a CPT-4 code indicating administration of parenteral chemotherapy in a Part B claim, or had a pharmacy (PDE) claim with an NDC code which mapped to an RxNorm ingredient code for an active pharmaceutical

A value of "True" was assigned if a sufficient quantity of these drugs was filled to extend up to

Immunization Variables: Influenza and pneumococcal vaccinations were identified using their respective CPT-4 codes found in Part B claims from October 1

Covid-19 hospitalizations were identified either by Part B Carrier claims with place of service code and CPT codes indicating inpatient care with a date of service no more than 14 days after