Changes in COVID-19-related outcomes, potential risk factors and disparities over time

To investigate temporal trends in coronavirus disease 2019 (COVID-19)-related outcomes and to evaluate whether the impacts of potential risk factors and disparities changed over time, we conducted a retrospective cohort study with 249 075 patients tested or treated for COVID-19 at Michigan Medicine (MM), from 10 March 2020 to 3 May 2021. Among these patients, 26 289 were diagnosed with COVID-19. According to the calendar time in which they first tested positive, the COVID-19-positive cohort were stratified into three-time segments (T1: March–June, 2020; T2: July–December, 2020; T3: January–May, 2021). Potential risk factors that we examined included demographics, residential-level socioeconomic characteristics and preexisting comorbidities. The main outcomes included COVID-19-related hospitalisation and intensive care unit (ICU) admission. The hospitalisation rate for COVID-positive patients decreased from 36.2% in T1 to 14.2% in T3, and the ICU admission rate decreased from 16.9% to 2.9% from T1 to T3. These findings confirm that COVID-19-related hospitalisation and ICU admission rates were decreasing throughout the pandemic from March 2020 to May 2021. Black patients had significantly higher (compared to White patients) hospitalisation rates (19.6% vs. 11.0%) and ICU admission rates (6.3% vs. 2.8%) in the full COVID-19-positive cohort. A time-stratified analysis showed that racial disparities in hospitalisation rates persisted over time and the estimates of the odds ratios (ORs) stayed above unity in both unadjusted [full cohort: OR = 1.98, 95% confidence interval (CI) (1.79, 2.19); T1: OR = 1.70, 95% CI (1.36, 2.12); T2: OR = 1.40, 95% CI (1.17, 1.68); T3: OR = 1.55, 95% CI (1.29, 1.86)] and adjusted analysis, accounting for differences in demographics, socioeconomic status, and preexisting comorbid conditions (full cohort: OR = 1.45, 95% CI (1.25, 1.68); T1: OR = 1.26, 95% CI (0.90, 1.76); T2: OR = 1.29, 95% CI (1.01, 1.64); T3: OR = 1.29, 95% CI (1.00, 1.67)).

Emerging evidence shows an appreciable decline in mortality rates among patients with severe COVID-19 outcomes after the initial wave in March and April 2020 [16][17][18]. A study conducted in a single health system in New York City found that the hospital mortality rate dropped from 25.6% at the start of the pandemic in March to 7.6% by mid-August 2020, adjusting for demographic and clinical factors [17]. The same trend of improved survival over time was seen among COVID-19 patients requiring hospitalisation [18] or critical care management [16] in England. Racial and ethnic disparities still remain, but recent data showed reduced differences in age-adjusted case fatality rate across ethnic groups [19]. Introduction of vaccines in the beginning of the year 2021 has also impacted COVID outcomes [20,21].
To understand the dynamics of potential risk factors on COVID-19 outcomes over time, time-stratified analyses are required. Understanding how risk factors changed over the course of the pandemic can enhance the coordination of healthcare resources and help protect the most vulnerable. It can also help us understand whether targeted outreach/prevention efforts in the early months of the pandemic to specific vulnerable subgroups had an effect on improving their outcomes or if disparities persisted over time.
As a follow-up to the smaller cross-sectional study of Gu et al. [9], this time-stratified retrospective cohort study aims to examine the sociodemographic and clinical characteristics that are associated with various COVID-19 outcomes by race/ethnicity in a much larger COVID-19 cohort over a 15-month time period, as well as to characterise the risk factor trajectories over time, using electronic health records (EHRs) from Michigan Medicine (MM), a large academic health care system in the State of Michigan.

Study cohort and COVID-19 testing
The study sample consisted of 249 075 patients tested or treated for COVID-19 at MM, the University of Michigan Health System, from 10 March 2020 to 3 May 2021, presenting 26 289 testing positive.
The descriptive statistics of the tested cohort are summarised in Table 1. Patients in the tested cohort received one of the five types of diagnostic tests: an in-house polymerase chain reaction (PCR) test (127 615 patients (52.1%)), a point of care PCR test (5781 patients (2.4%)), a commercial PCR test (Viracor; 418 patients (0.2%)), COVID-19 nasopharynx or oropharynx PCR tests deployed by the Michigan Department of Health and Human Services (52 patients (0.02%)), and a small fraction of ribonucleic acid (RNA) tests (three patients (0.001%)); 111 098 tested patients (45.4%) were transferred, tested elsewhere, or had no information on the type of testing they received.

COVID-19 outcomes
In this study, we primarily focused on three COVID-19-related outcomes among individuals who tested positive: hospitalisation, intensive unit care (ICU) admission, and all-cause mortality. All patients who died post COVID-19 were included in the mortality counts. We performed sensitivity analysis by only considering outcomes that happened within 6 months of a COVID-19 diagnosis. We also investigated the factors associated with being tested (a question related to both access to testing and infection rates) and testing positive for COVID-19 (using a randomly selected untested comparison group described in the Supplementary Material). The corresponding results for these two secondary outcomes are largely reported in the Supplementary Material. The definitions of this ensemble of COVID-19 outcomes are listed in Supplementary Table S1, and the sample sizes for each outcome category are presented in Supplementary Figure S1.
Definition of demographics, socioeconomic status, comorbidities, and other covariates We followed the same process described in Gu et al. [9] to extract the relevant variables from the EHRs. The sociodemographic variables considered included age, self-reported sex, race/ethnicity, smoking status, alcohol consumption, body mass index (BMI), Neighbourhood Socioeconomic Disadvantage Index (NDI) [22], and population density (in persons per square mile) [22]. NDI is a scale ranging from 0 to 1, with larger values indicating more disadvantaged communities. To characterise the underlying medical conditions of the patients, we considered seven comorbid conditions: respiratory conditions, circulatory conditions, any cancer, type 2 diabetes, kidney diseases, liver diseases and autoimmune diseases, all of which were coded binary. The general health status of patients was represented by a comorbidity score ranging from 0 to 7 that sums up the aforementioned seven conditions. In addition, all patients receiving medical care at least once at any MM primary care location since 1 January 2018 were considered as having MM as their primary care provider.

Statistical analysis
For each COVID-19-related outcome Y COVID , we fit a Firth's biascorrected logistic regression to avoid the potential separation issues in the traditional maximum likelihood estimates. The analysis model is given by where X and Covariate denote the predictors of interest (including Race) and the vector of adjustment covariates, respectively. To assess the sensitivity of our analysis to the choice of confounders/ adjustment variables, we explored four nested sets of variables for Covariate, respectively labelled adjustment 0-3 (defined in Supplementary Table S1 and results reported in Supplementary  Table S2). Our final model (adjustment 3) adjusted for age, sex, race/ethnicity, NDI and the composite comorbidity score. We excluded the cumulative comorbidity score when the factor of interest X was one of the individual comorbid conditions. The models for being tested and testing positive further adjusted for population density as that is likely to be related to disease transmission. We further conducted a time-stratified analysis to assess the changes in racial/ethnic disparities and the effects of other risk factors for severe COVID-19-related outcomes over time since the start of the pandemic. According to the time period in which they first tested positive, the COVID-19-positive cohort were stratified into three groups: Time Period 1 (T1), 10 March 2020-30 June 2020; Time Period 2 (T2), 1 July 2020-31 December 2020; Time Period 3 (T3), 1 January 2021-3 May 2021. Similarly, the COVID-19-negative cohort were stratified based on the time period in which they first received the tests. The logistic regression model (1) was fitted using data for each time period.
The Race variable was self-reported and consisted of four categories: non-Hispanic White (White), non-Hispanic Black (Black), other known race/ethnicity, and unknown race/ethnicity. While the disproportionate burden of disease for Hispanics/ Latinos is an important public health issue [11,23], our cohort did not have a sufficiently large sample size to afford the analysis. Therefore, we focus on comparing White and Black patients to evaluate racial/ethnic differences.
To evaluate the possibility of effect modification by race/ethnicity on the effect of potential risk factors, we fit a set of models with interactions by race: For each model, we report the Firth's bias-corrected estimates of odds ratios along with their associated 95% Wald confidence intervals (CI) and P values. For the interaction models, we also obtained the P values of the differences in subgroup effects for White and Black patients by testing the null hypothesis H 0 : β int = 0. All analyses were performed using complete cases in R statistical software version 4.0.3 (R Project for Statistical Computing). Among the 26 289 patients who tested positive, 3071 (11.7%) were hospitalised, 844 (3.2%) were admitted to the ICU, and 485 (1.8%) died. Older patients, male sex, ever smokers, and increased comorbidity burden tended to be associated with worsened disease severity among COVID-19-positive patients ( Table 1). From T1 to T3, the mean and median ages decreased over time for patients who were hospitalised or admitted to the ICU. For example, the mean ages for hospitalised patients were 58.2, 54.2, and 49.0 years in T1, T2, and T3, respectively (Supplementary Table S3). This could potentially be attributed to the increased availability of vaccines for the elderly, the younger population in working-age groups returning to more in-person activities, and a harvesting effect.

Descriptive statistics and unadjusted analysis
Stratifying the descriptive statistics by White and Black patients, we noted differences in covariate distributions between the two groups (Supplementary Table S4). For example, in the tested cohort, Black patients tended to be younger (mean (S.D.) age, 41.4 (21.7) vs. 45.7 (23.2)), live in higher populated areas (mean (S.D.) population density in persons per square mile, 3439.4 (2270.7) vs. 1985.5 (2201.6)), and be of lower socioeconomic status (mean (S.D.) NDI, 0.21 (0.13) vs. 0.08 (0.06)) than White patients. Similar patterns existed in the subgroups (e.g. hospitalised patients) of the tested cohort. We assumed that the data were missing completely at random and used observed complete cases for the analysis. The fractions of missingness for each covariate are presented in Supplementary Table S5. Figure 1 presents the changes in the overall and race-stratified COVID-19 outcomes over time. The overall hospitalisation rate among patients testing positive decreased from 36.2% in T1 to 14.2% in T3, and the overall ICU admission rate decreased from 16.9% to 2.9%. In the sensitivity analysis that only considers outcomes that happened within 6 months of a COVID-19 diagnosis, similar patterns were noted for the hospitalisation and ICU Table 1.   Figure S2 and Table S6). Furthermore, age, male sex, ever smoker, and lower NDI were inversely associated with the odds of testing positive. All the comorbidity conditions considered were associated with an increased risk of testing positive. , possibly due to the reduced sample size after stratification. Additionally, age, male sex, smoking, living in densely populated areas or disadvantaged neighbourhoods, and heavier overall comorbidity burden were positively associated with hospitalisation and ICU admission in the full COVID-positive cohort ( Table 2). Most of the associations in each time period stayed in the same direction as in the full cohort analysis ( Interaction analysis by race/ethnicity within the COVID-19-positive cohort Figure 2 displays the ORs of potential risk factors stratified by race/ethnicity for the hospitalisation outcome ( Fig. 2A) and ICU admission (Fig. 2B). The main effects of the seven comorbid conditions considered were significant in both Black and White patients for hospitalisation, while none of the comorbidities exhibited significant interaction effects. For example, the OR of kidney diseases for Black and White patients were 3.88 (95% CI (3.07, 4.91), P < 0.001) and 3.86 (95% CI (3.38, 4.40), P < 0.001), respectively, and the interaction P value was 0.96. Higher NDI (i.e. lower socioeconomic status) had significant positive associations with hospitalisation in White patients (OR = 9.29, 95% CI (4.15, 20.8), P < 0.001), but not in Black patients (OR = 1.64, 95% CI (0.69, 3.87), P = 0.263), with the interaction being significant (P int = 0.004). For the ICU admission (Fig. 2B)

Discussion
In this study, we evaluated the potential risk factors and racial/ ethnic disparities for COVID-19 prognosis using a larger cohort from the same healthcare system that covers a longer time frame than Gu et al. [9]. We observed that Black patients and patients with higher comorbidity burden were more susceptible to COVID-19 and more likely to be hospitalised, a result consistent with previous findings in the literature. On the other hand, the association between age and positive test results reversed compared to what was reported in Gu et al. [9], such that younger people, especially those aged 18-35 years, were more likely to be tested positive. The finding aligns with recent data from the Centers for Disease Control and Prevention, which shows an increase in the COVID-19 incidence in persons aged <30 years and a decrease in the median age for confirmed cases in May−August 2020 [24].
A fraction of patients who were hospitalised (n = 249 of 3071 (8.1%)) in our sample were transferred from external hospitals, mainly from the Detroit Metro area. These transferred patients tended to have more severe symptoms and different distributions in sociodemographic characteristics than non-transferred patients. It is important to note the temporal changes in the context of care at MM (Supplementary Table S7). There was an enrichment in transferred patients in T1 (n = 123 (19.3%)) compared to T2 (n = 76 (5.6%)) and T3 (n = 50 (4.7%)) for those hospitalised. However, a sensitivity analysis that excluded patients who did not receive their primary care at MM was consistent with the overall analysis, namely improved outcomes and reduced but persisting health disparities over time (Supplementary  Table S8).
Our definition of COVID-19 related hospitalisation, ICU admission, and mortality may be liberal. We performed a sensitivity analysis restricting the time period to six months since the first positive diagnosis. For example, of the 485 deaths, 458 (94.4%) happened within six months. The results of the sensitivity analysis (Supplementary Figure S3) showed consistent patterns with the analysis based on unfiltered outcomes ( Fig. 1) with respect to the changes in the rates and ORs of the three outcomes. Specifically, the overall hospitalisation rate, ICU admission rate, and mortality rate decreased from 34.2%, 16.5%, and 7.1% to 9.6%, 1.9%, and 0.8%, respectively.
Our findings suggest that the COVID-19-related hospitalisation, ICU admission, and mortality rates declined from T1 to T2, which may be partly explained by the rapid increase in testing and adoption of preventive measures. The rates of these severe 6 Youfei Yu et al. outcomes remained relatively steady from T2 to T3. However, racial/ethnic disparities persisted over time. This study also confirms a number of potential risk factors for COVID-19-related outcomes: older age, male sex, Black race/ethnicity, preexisting conditions, and lower socioeconomic status. Among the examined comorbidities, underlying kidney diseases in general had the largest odds ratios for the COVID-19-related outcomes. The interaction analysis identified several risk factors that affected the hospitalisation and ICU admission differently in Black and White patients. Specifically, underlying respiratory diseases and liver diseases increased the risk of hospitalisation in the subgroup of White patients, while the effects on Black patients were not significant. On the other hand, obesity and autoimmune diseases were positively associated with ICU admission in Black patients but not in White patients.
There are several limitations of our study. First, the sample cannot be considered a population-based random sample, as MM prioritised testing to those presenting COVID-19 symptoms or at greatest risk of exposure, particularly during the early stages of the pandemic when the availability of tests was limited. Second, since the hospitalisation records were available only for those treated at MM, not all hospitalised patients were captured by our data. Patients hospitalised elsewhere may be classified as nonhospitalised, which may lead to biased estimates. Finally, 486 out of the 52 325 people who were first tested in T3 received at least one vaccination dose by May 3, but we did not consider vaccination as a covariate, which may explain part of the reduction in hospitalisation and mortality. Despite the limitations, our study assessed the temporal trend in COVID-19-related outcomes and evaluated whether the impacts of potential risk factors and disparities changed over time. Elderly adults, Black patients, and patients with underlying health conditions were disproportionately affected by COVID-19 over the course of the pandemic, which calls for vaccine prioritisation and targeted outreach in these population groups.  Institutional Review Board Statement. The University of Michigan Medical School institutional review board reviewed the study and determined that it is exempt. Ethical review and approval were waived for this study per institutional policy.
Informed Consent Statement. The institutional review board waived the need to obtain consent for the collection, analysis, and publication of the anonymised COVID-19 data, per institutional policy.