Abstract
Background
Studies of new-onset diabetes as a post-acute sequela of SARS-CoV-2 infection are difficult to generalize to all socio-demographic subgroups.
Objective
To study the risk of new-onset diabetes after SARS-CoV-2 infection in a socio-demographically diverse sample.
Design
Retrospective cohort study of electronic health record (EHR) data available from the OneFlorida + clinical research network within the National Patient-Centered Clinical Research Network (PCORnet).
Subjects
Persons aged 18 or older were included as part of an Exposed cohort (positive SARS-CoV-2 test or COVID-19 diagnosis between 1 March 2020 and 29 January 2022; n = 43,906), a contemporary unexposed cohort (negative SARS-CoV-2 test; n = 162,683), or an age-sex matched historical control cohort (index visits between 2 Mar 2018 and 30 Jan 2020; n = 40,957).
Main Measures
The primary outcome was new-onset type 2 diabetes ≥ 30 days after index visit. Hazard ratios and cases per 1000 person-years of new-onset diabetes were studied using target trial approaches for observational data. Associations were reported by sex, race/ethnicity, age, and hospitalization status subgroups.
Key Results
The sample was 62% female, 21.4% non-Hispanic Black, and 21.4% Hispanic; mean age was 51.8 (SD, 18.9) years. Relative to historical controls (cases, 28.2 [26.0–30.5]), the unexposed (HR, 1.28 [95% CI, 1.18–1.39]; excess cases, [5.1–10.3]), and exposed cohorts (HR, 1.64 [95% CI, 1.50–1.80]; excess cases, 17.3 [13.7–20.8]) had higher risk of new-onset T2DM. Relative to the unexposed cohort, the exposed cohort had a higher risk (HR, 1.28 [1.19–1.37]); excess cases, 9.5 [6.4–12.7]). Findings were similar across subgroups.
Conclusion
The pandemic period was associated with increased T2DM cases across all socio-demographic subgroups; the greatest risk was observed among individuals exposed to SARS-CoV-2.
Supplementary Information
The online version contains supplementary material available at 10.1007/s11606-024-09035-8.
KEY WORDS: long COVID, diabetes risk, post-acute sequelae of COVID-19, COVID-19
INTRODUCTION
The USA reported over 100 million confirmed cases of SARS-CoV-2 since the onset of the pandemic.1 Recent reports of new-onset diabetes in adults and children after SARS-CoV-2 infection have prompted concerns for a future increase in type 2 diabetes mellitus (T2DM) cases. 2–6 Such increases could be especially detrimental to socioeconomically disadvantaged groups such as Black and Hispanic communities that experience disproportionately higher rates of SARS-CoV-2 infection, COVID-19-related morbidity and mortality, and T2DM in the USA. 7–9
Limited diversity in previous reports of new-onset T2DM after SARS-CoV-2 makes the results difficult to generalize to all socio-demographic subgroups, including racial/ethnic minorities, younger adults, and women. Data from the Department of Veterans Affairs (VA) found the increased risk for T2DM persisted in stratified analyses by age, race, sex, BMI, and baseline risk for T2DM.3 But their study population was limited to older veterans, mostly White and mostly male; categorization of race was limited to White, Black, or Other; and Hispanic ethnicity was not reported in these analyses. Higher excess risk of diabetes was observed among White and non-White patients (Asian, Black, and/or Other combined) in a study utilizing primary care records from the National Health Service in the UK.10 Patient-level race/ethnicity data were also not available in analyses of T2DM after COVID-19 using administrative claims and electronic health records from USA and Germany.11–13
There is a critical need to investigate associations between SARS-CoV-2 infection and diabetes risk in diverse samples to better inform post-infection diabetes screening and management guidelines. Using analytical approaches for target trial emulation, we examined if the risk of new-onset diabetes differed by socio-demographic (sex, race/ethnicity, age) category and hospitalization status in a socio-demographically diverse, real-world cohort of patients from Florida prior to and during the COVID pandemic.
METHODS
Study Population
We conducted a secondary analysis of patient-level electronic health record (EHR) data from OneFlorida + , a clinical research network (CRN) within the National Patient-Centered Clinical Research Network (PCORnet), built on the PCORnet Common Data Model.14 The OneFlorida + Data Analytics team independently queried the data warehouse to extract socio-demographic and clinical characteristics of eligible patients.15
We built three retrospective cohorts of adult individuals: an exposed cohort, an unexposed cohort, and a historical cohort (Fig. 1). The index date was defined based on results of the COVID-19 test (positive or negative) or COVID-19 diagnosis code or the earliest date of visit in the pre-pandemic period for the historical cohort. The exposed cohort consisted of individuals exposed to SARS-CoV-2 during the pandemic period (1 March 2020 to 29 January 2022). The unexposed cohort consisted of individuals with no recorded exposure to SARS-CoV-2 but utilized the healthcare system during the pandemic period. The historical cohort consisted of individuals utilizing the healthcare system from 2 March 2018 to 30 January 2020. Patients who were not included in the exposed or unexposed cohorts were sampled into the historical cohort after matching on sex and 5-year age intervals. The historical cohort was intended to study confounding by period effects on the association of SARS-CoV-2 with new-onset T2DM. The three cohorts were followed up starting from the post-acute period, i.e., 30 days from index date.
Figure 1.
Cohort selection. The index date was defined based on results of the positive SARS-CoV-2 test or COVID-19 diagnosis code for the exposed cohort, the first negative SARS-CoV-2 test for the unexposed and the earliest date of visit in the pre-pandemic period for the historical cohort.
To be eligible for inclusion in the study sample, individuals would have sought care before and after the index date at least three times in the OneFlorida + clinical sites.14 Individuals were included if they were adults (18 or older) with at least two encounters (including at least one clinical encounter) recorded in OneFlorida + in the lookback period (24 months preceding the index date, and at least one clinical encounter in the post-acute follow-up period after index date (≥ 30 days from index date until last visit or 28 February 2022 [29 February 2020 for Historical cohort]). Individuals were excluded from the study sample if they had pre-existing diabetes mellitus, as defined by previously validated computable phenotypes for T1DM and T2DM prior to the index date.16
This real-world data analysis followed guidelines recommended by the REporting of studies Conducted using Observational Routinely collected health Data (RECORD) Statement.17
Exposure Cohorts
The exposed cohort met at least one of two criteria between 1 March 2020 and 29 January 2022: any positive test (nucleic acid amplification tests [NAAT] or antigen) or COVID-19 related ICD-10-CM codes.18 We defined the index date of infection exposure (T0) as the date of the first positive test or date of record for the ICD-10-CM code. The timeline for cohort selection and follow-up is presented in Supplementary Fig. 1.
The unexposed cohort met three criteria between 1 March 2020 and 29 January 2022: no positive test during follow-up period, at least one negative test (NAAT or antigen) and no COVID-19 related ICD-10 CM codes. The index date was the date of the first negative test to maximize follow-up duration.
The index dates of historical cohort ranged from 2 March 2018 and 30 January 2020; we selected this interval, so the last date of possible follow-up (29 February 2020) did not overlap with the start of the calendar period of identification (1 March 2020) for the exposed and unexposed cohorts.
The analytic sample was restricted from the study sample to exclude individuals who did not have a measurement of weight in the preceding year (Supplementary Fig. 1, n = 138,317). We also excluded 2197 individuals with T2DM diagnosed in the acute phase (0 to 30 days) to minimize misclassification of critical-illness hyperglycemia with newly diagnosed T2DM. The final analytic sample consisted of 43,906 exposed, 162,683 unexposed controls, and 40,957 historical controls. All analyses accounted for the sampling of historical controls.
Outcome—New-Onset Type 2 Diabetes Mellitus
We defined the computable phenotype of new-onset T2DM in our analytic sample in the 30 days after index date (post-acute phase) as those who meet at least two of three distinct criteria within 90 days of each other: ICD-9/10CM diagnosis code for T2DM, antidiabetic medication, and/or HbA1c ≥ 6.5%.16 Patients without the computable phenotype for new-onset T2DM were censored at the date of the last recorded lab or diagnosis code or prescription.
Effect Modifiers—Socio-demographic Characteristics and Hospitalization Status
We included sex (male, female), race/ethnicity (non-Hispanic [NH] White, NH Black, Hispanic), and age on index date (18–39, 40–64, 65, and older) as key socio-demographic characteristics in the analysis. All patients with an inpatient hospital encounter in the acute phase (30 days from index date) were classified as hospitalized, while all others were considered not hospitalized.
Confounders
We used investigator-specified covariates and empirically identified covariates to adjust for baseline differences between cohorts using inverse probability weighting. We derived investigator-specified covariates known to be associated with higher cardiometabolic risk from data collected before the index date. Supplementary Table 1 provides medication, laboratory, and diagnosis codes used as investigator-specified covariates. We additionally identified the health system and primary insurance type on the index date.
We computed empirically identified covariates across available data dimensions: diagnoses (Clinical Classifications Software Refined categories), procedures (Current Procedural Terminology or ICD-10-CM inpatient or outpatient codes), prescriptions (Anatomical Therapeutic Chemical Level 3 classes), and laboratory tests (Logical Observation Identifiers Names and Codes). We identified the top 200 covariates in each data domain based on their prevalence in the sample. For each covariate, we then computed three intensity dummy variables (at least once, sporadic [> median participant], frequent [> 75th percentile]).19
Statistical Analysis
Inverse Probability Weighting for Cohort Membership and Follow-up
Baseline characteristics may differ between exposed, unexposed, and historical control cohorts resulting in non-random cohort membership in observational studies, including this one. We used a machine learning-based propensity score, developed using investigator defined and algorithmically selected covariates, to balance covariates between the exposed, unexposed and historical cohorts.20 We assessed covariate balance between the unexposed and historical control cohorts, relative to the exposed, after inverse probability weighting (IPW) using population standardized bias.21 Bias less than 0.1 after weighting was considered as indicative of covariate balance. We additionally constructed IPW for availability of follow-up data to minimize selection bias, since not all participants had data during the follow-up period.22 IPW were also constructed separately for each effect modifier of interest with numerator reflecting probability of cohort membership under levels of each characteristic (sex, race/ethnicity, age, hospitalization) when assessing differences between levels.23
Risk and Excess Burden of Type 2 Diabetes
First, we used a marginal structural Cox Proportional Hazards regression, after IPW, with robust sandwich variance to estimate the relative risk (hazard ratio) and risk difference (excess cases) per 1000 person-years of T2DM in the unexposed and exposed cohort, relative to the historical cohort. We additionally estimated the hazard ratio and excess cases for those exposed, relative to the unexposed cohort. We adjusted for imbalanced covariates.24,25 Second, we added statistical interactions to characterize the relative risk and excess burden across categories of sex (male vs female), race/ethnicity (NH White vs NH Black vs Hispanic), age (18–39 vs 40–64 vs 65 and older) and hospitalization status in the post-acute period (time from 30 days since index date).
Sensitivity Analysis
First, to minimize the influence of patients who rarely used the health system, we restricted the analytic sample to those patients with at least 100 days of follow-up and at least one encounter when laboratory parameters were measured in the follow-up period. Second, since our computable phenotype for new-onset T2DM was not previously validated, we used an alternative definition of new-onset T2DM that more closely matches the SUPREME-DM computable phenotype of incident T2DM. Third, to account for differences in timing of follow-up visits, we adjusted for county-level COVID-19 transmission. Fourth, we used multiple imputation to impute missing values of laboratory parameters in lookback period and adjusted for imbalanced covariates in the regression model. Fifth, to minimize the possibility of spurious associations, we used two negative controls: traumatic fractures and ear infections. Finally, we used overlap weights to address extreme propensity scores.26
All analysis was carried out using R 4.2.3 using survival package (version 3.5–3) and Python 3.11.4. The Emory University Institutional Review Board determined this study to be exempt (Protocol 4932).
RESULTS
The analytic sample for studying the association of SARS-CoV-2 infection and new-onset T2DM consisted of 247,546 adults of whom 154,086 (62%) were female with a mean age of 51.8 years (SD, 18.9). Of the analytic sample, 43,906 belonged to the exposed cohort, 162,683 belonged to the unexposed cohort, and 40,957 belonged to the historical control cohort. Nearly half the analytic sample were racial-ethnic minorities, with 21.4% identifying as non-Hispanic Black and 21.4% identifying as Hispanic (Table 1).
Table 1.
Descriptive Characteristics of Analytic Sample, n = 247,546
| Overall | Exposed | Unexposed | Historical | Population standardized biasa | |
|---|---|---|---|---|---|
| N | 247,546 | 43,906 | 162,683 | 40,957 | |
| Follow-up available | 218,848 (88%) | 38,673 (88%) | 142,878 (88%) | 37,297 (91%) | |
| Follow-up duration (in days) | 243 (98, 423) | 176 (75, 358) | 215 (88, 373) | 462 (266, 594) | |
| New T2DM cases | 6000 (2.7%) | 1231 (3.2%) | 3608 (2.5%) | 1161 (3.1%) | |
| CP1 (diagnosis and prescription) | 1652 (0.8%) | 368 (1.0%) | 804 (0.6%) | 480 (1.3%) | |
| CP2 (diagnosis and HbA1c) | 288 (0.1%) | 62 (0.2%) | 139 (< 0.1%) | 87 (0.2%) | |
| CP3 (HbA1c and prescription) | 4060 (1.9%) | 801 (2.1%) | 2665 (1.9%) | 594 (1.6%) | |
| Censored | 212,848 (97%) | 37,442 (97%) | 139,270 (97%) | 36,136 (97%) | |
| Female (%) | 154,086 (62%) | 28,731 (65%) | 99,236 (61%) | 26,119 (64%) | 0.019 |
| Age (years) | 51.8 (18.9) | 48.7 (18.4) | 53.5 (19.0) | 48.6 (18.4) | 0.117 |
| Race/ethnicity | |||||
| NH White | 127,686 (51.6%) | 19,386 (44.2%) | 87,452 (53.8%) | 20,848 (50.9%) | |
| NH Black | 52,918 (21.4%) | 11,657 (26.5%) | 32,163 (19.8%) | 9098 (22.2%) | 0.105 |
| Hispanic | 52,946 (21.4%) | 10,497 (23.9%) | 33,755 (20.7%) | 8694 (21.2%) | 0.031 |
| NH Other | 13,945 (5.6%) | 2336 (5.3%) | 9293 (5.7%) | 2316 (5.7%) | 0.008 |
| Smoking | |||||
| Yes | 116,243 (47%) | 18,462 (42%) | 77,617 (48%) | 20,164 (49%) | 0.101 |
| Primary insurance | |||||
| Medicare | 38,064 (18%) | 5639 (13%) | 24,666 (19%) | 7759 (19%) | 0.032 |
| Medicaid | 17,930 (8.3%) | 3305 (7.7%) | 11,062 (8.4%) | 3563 (8.7%) | 0.016 |
| Other government | 5255 (2.4%) | 1236 (2.9%) | 3296 (2.5%) | 723 (1.8%) | 0.015 |
| Private | 99,030 (46%) | 23,048 (54%) | 54,541 (41%) | 21,441 (52%) | 0.057 |
| No insurance | 4168 (1.9%) | 567 (1.3%) | 1872 (1.4%) | 1729 (4.2%) | 0.032 |
| No information | 51,138 (24%) | 9240 (21%) | 36,156 (27%) | 5742 (14%) | 0.038 |
| Missing | 31,961 | 871 | 31,090 | 0 | |
| Hospitalization within 30 days of index date (acute phase) | 64,880 (26%) | 12,069 (27%) | 49,806 (31%) | 3,005 (7.3%) | 0.325 |
| Medication (any within 1 year prior to index date) | |||||
| Antidepressants | 31,805 (13%) | 5887 (13%) | 20,739 (13%) | 5179 (13%) | 0.021 |
| Antipsychotics | 10,869 (4.4%) | 2192 (5.0%) | 7111 (4.4%) | 1566 (3.8%) | 0.02 |
| Antihypertensives | 67,675 (27%) | 12,244 (28%) | 46,011 (28%) | 9420 (23%) | 0.069 |
| Statins | 31,508 (13%) | 5556 (13%) | 21,949 (13%) | 4003 (9.8%) | 0.073 |
| Immunosuppressants | 22,143 (8.9%) | 4661 (11%) | 14,524 (8.9%) | 2958 (7.2%) | 0.059 |
| Anthropometry within 1 year prior to index date | |||||
| Height | 66.2 (4.0) | 66.1 (4.0) | 66.2 (4.1) | 66.1 (4.0) | 0.02 |
| Missing | 1438 | 360 | 969 | 109 | |
| BMI | 29.4 (7.4) | 30.5 (7.7) | 29.3 (7.3) | 29.0 (7.2) | 0.129 |
|
Normal (18.5–24.9 kg/m2) |
67,676 (27%) | 10,414 (24%) | 45,109 (28%) | 12,153 (30%) | |
|
Overweight (25.0–29.9 kg/m2) |
74,552 (30%) | 12,403 (28%) | 49,678 (31%) | 12,471 (30%) | |
| Obese (≥ 30 kg/m2) | 99,208 (40%) | 20,214 (46%) | 63,590 (39%) | 15,404 (38%) | |
| Systolic BP | 125.7 (18.6) | 125.0 (17.8) | 125.9 (18.9) | 125.7 (17.9) | 0.017 |
| Missing | 30,288 | 7813 | 19,361 | 3114 | |
| Labs within 1 year prior to index date | |||||
| HbA1c (%) | 6.2 (1.6) | 6.3 (1.7) | 6.2 (1.6) | 6.1 (1.5) | 0.095 |
| Missing | 200,133 | 33,669 | 130,418 | 36,046 | |
| Glucose (mg/dL) | 112.8 (45.7) | 113.6 (47.9) | 113.1 (45.6) | 109.7 (42.1) | 0.07 |
| Missing | 108,292 | 17,026 | 67,578 | 23,688 | |
| ALT | 17.0 (12.0, 26.0) | 17.0 (12.0, 26.0) | 17.0 (12.0, 26.0) | 17.0 (12.0, 26.0) | 0.023 |
| Missing | 129,254 | 20,314 | 81,846 | 27,094 | |
| AST | 20.0 (16.0, 26.0) | 19.0 (15.0, 25.0) | 20.0 (16.0, 26.0) | 19.0 (16.0, 25.0) | 0.05 |
| Missing | 127,827 | 19,972 | 81,193 | 26,662 | |
| Serum creatinine | 0.8 (0.7, 1.1) | 0.8 (0.7, 1.0) | 0.9 (0.7, 1.1) | 0.8 (0.7, 1.0) | 0.076 |
| Missing | 110,270 | 17,251 | 68,920 | 24,099 | |
| HDL | 50.9 (16.9) | 50.2 (15.7) | 51.0 (16.8) | 51.8 (19.2) | 0.089 |
| Missing | 192,594 | 32,393 | 126,018 | 34,183 | |
| LDL | 83.1 (50.0) | 80.3 (53.5) | 81.7 (49.8) | 96.0 (42.1) | 0.314 |
| Missing | 190,329 | 31,630 | 124,776 | 33,923 | |
| Diagnosis codes for comorbidities within 2 years prior to index date | |||||
| Obesity | 9979 (4.0%) | 2222 (5.1%) | 5898 (3.6%) | 1859 (4.5%) | 0.022 |
| Cardiovascular | 12,238 (4.9%) | 2127 (4.8%) | 7700 (4.7%) | 2411 (5.9%) | 0.023 |
| Cerebrovascular | 536 (0.2%) | 103 (0.2%) | 298 (0.2%) | 135 (0.3%) | 0.016 |
| Hypertension | 23,402 (9.5%) | 4162 (9.5%) | 14,231 (8.7%) | 5009 (12%) | 0.036 |
| Pulmonary disease | 7370 (3.0%) | 1491 (3.4%) | 4670 (2.9%) | 1209 (3.0%) | 0.023 |
| Hyperlipidemia | 18,191 (7.3%) | 3578 (8.1%) | 10,845 (6.7%) | 3768 (9.2%) | 0.03 |
| Follow-up encounters | |||||
| Ambulatory visits | 4 (1, 12) | 3 (1, 10) | 4 (1, 12) | 6 (2, 14) | |
| Telehealth visits | 0 (0, 0) | 0 (0, 1) | 0 (0, 1) | 0 (0, 0) | |
| Other outpatient visits | 0 (0, 4) | 0 (0, 3) | 0 (0, 4) | 0 (0, 2) | |
Values were mean (standard deviation) or median (25th percentile, 75th percentile) or frequency (percentage%)
aPopulation standardized bias was computed after inverse probability weighting. Descriptive statistics after weighting is provided in Supplementary Table 2. Population standardized bias less than 0.1 after weighting is considered as indicative of covariate balance. Values greater than or equal to 0.1 are bolded
CP1: diabetes diagnosis and medication use within 90 days; CP2: diabetes diagnosis and HbA1c ≥ 6.5% within 90 days; CP3: HbA1c ≥ 6.5% and medication use within 90 days
The exposed cohort were more likely to be female (65%) and younger (48.7 years [SD, 18.4]) and had shorter follow-up (176 days [IQR, 75–358]), relative to the unexposed cohort (female, 61%; age, 53.5 years [SD, 19.0]; follow-up, 215 days [IQR, 88–373]) (Supplementary Fig. 2). Relative to the unexposed (46.2%) and historical (49.1%) cohorts, the exposed (55.8%) cohort was also more likely to consist of racial-ethnic minorities. The exposed and historical cohorts were similar in age and sex distribution as expected. the exposed (27%) and unexposed (31%) cohorts were more likely to be hospitalized, relative to the historical (7%) cohort. The exposed cohort was heavier on average on the index date (BMI 30.5 kg/m2 [SD, 7.7]), relative to the unexposed (29.3 kg/m2 [SD, 7.3]) and historical (29.0 kg/m2 [7.2]) cohorts. After IPW, the cohorts were imbalanced on age, proportion of Non-Hispanic Black ethnicity, health system on index date, hospitalization within 30 days of index date, BMI, and LDL cholesterol.
Follow-up information on laboratory, diagnosis codes, and prescriptions were available for 218,848 (88%) of the analytic sample, with a median follow-up of 243 days (IQR, 98–423). During the follow-up period, there were 6000 cases of new-onset T2DM observed with 212,848 not meeting the T2DM criteria during follow-up. We note that those excluded because of missing BMI (n = 124,208) or detection of new-onset T2DM before start of follow-up (n = 16,306) were less likely to belong to the unexposed cohort (54% versus 66% among the analytic sample), were younger (49.4 years versus 51.8 years) and less likely to use statins and antihypertensives (Supplementary Table 2). For instance, those without follow-up were younger (available, 52.1 years [SD, 18.8]; unavailable, 49.9 years [SD, 19.5]), more likely to be Non-Hispanic White (available 51.2%, unavailable 54.8%), and more likely to be hospitalized (available 25%, unavailable 32%). We used IPW for loss to follow-up to balance the differences in socio-demographic and clinical characteristics by follow-up status in subsequent analysis.
Risks and Burden of New-Onset Type 2 Diabetes Mellitus
Relative to the Historical cohort (cases per 1000 person-years, 28.2 [95% CI, 26.0–30.5]), the unexposed cohort (HR, 1.28 [1.18–1.39]; excess cases, 7.8 [5.1–10.4]) and the exposed cohort (HR, 1.64 [95% CI, 1.50–1.80]; excess cases, 17.5 [13.9–21.0]) had higher risks of T2DM (Table 2). Relative to the unexposed cohort, the exposed cohort had a higher relative risk (HR, 1.28 [1.20, 1.37]) and excess cases (9.7 [6.6–12.8]).
Table 2.
Risk and burdens of new-onset type 2 diabetes by exposure group
| Historical | Unexposed | Exposed | |
|---|---|---|---|
| Hazard ratio (95% CI) | Ref | 1.28 (1.18, 1.39) | 1.64 (1.50, 1.80) |
| Hazard ratio (relative to unexposed) | 0.78 (0.72, 0.84) | Ref | 1.28 (1.20, 1.37) |
| Burden per 1000 person-years (95% CI) | 28.2 (26, 30.5) | 36 (34.6, 37.4) | 45.7 (42.9, 48.5) |
| Burden relative to historical cohort per 1000 person-years (95% CI) | Ref | 7.8 (5.1, 10.4) | 17.5 (13.9, 21.0) |
|
Burden relative to unexposed cohort per 1000 person-years (95% CI) |
− 7.8 (− 10.4, − 5.1) | Ref | 9.7 (6.6, 12.8) |
All estimates are from the marginal structural model with statistical interaction of exposure group and socio-demographic characteristic (i.e., sex, age, race/ethnicity) and hospitalization status, adjusting for imbalanced covariates. Burden per 1000 people at 12 months (1000 person-years) were estimated after standardization to covariate distribution of analytic sample
The higher relative hazards and cases (per 1000 person-years) of the exposed cohort, followed by the unexposed cohort, relative to the historical cohort, were observed across all socio-demographic subgroups and by hospitalization status (Fig. 2). For instance, relative to the historical cohort, the unexposed and exposed cohorts had higher relative hazards of new-onset T2DM among both females (unexposed, 1.24 [95% CI, 1.12–1.38]; exposed, 1.59 [95% CI, 1.41–1.80]) and males (unexposed, 1.17 [9% CI, 1.05, 1.31]; exposed, 1.59 [95% CI, 1.40–1.82]) (Supplementary Table 3). Relative to the unexposed cohort, the relative and absolute burden was higher among the exposed cohort (females, 8.6 [95% CI, 4.9–12.2]; males, 14.1 [95% CI, 8.6–19.5]). Relative to the historical cohort, the unexposed has similar but the exposed cohorts had higher relative hazard of new-onset T2DM among those aged 18 to 39 years (unexposed, 0.94 [95% CI, 0.75–1.19]; exposed, 1.33 [95% CI, 1.02–1.74]). Associations were otherwise consistent among those aged 40 to 64 years, and 65 and older. Relative to the unexposed cohort, the relative hazards (NH White, 1.40 [95% CI, 1.25–1.56]; NH Black, 1.31 [95% CI, 1.17–1.46]; Hispanic, 1.16 [95% CI, 1.00–1.35]) and cases (NH White, 10.4 [95% CI, 6.3–14.6]; NH Black, 14.9 [95% CI, 7.7–22.2]; Hispanic, 6.5 [95% CI, − 0.6, 13.7]) were higher among the exposed cohort. Relative to the unexposed cohort, the relative hazards (hospitalized, 1.52 [95% CI, 1.34–1.73]; not hospitalized, 1.27 [95% CI, 1.17–1.38]) and cases (hospitalized, 19.1 [95% CI, 11.8–26.3]; not hospitalized, 8.8 [95% CI, 5.4–12.2]) were higher among the exposed cohort.
Figure 2.
Relative and absolute risk of T2DM in the post-acute period after SARS-CoV-2 infection. Our historical cohort corresponds with the green lines, symbols, and bars. Our unexposed cohort corresponds with the blue lines, symbols and bars. Our exposed cohort corresponds with the orange lines, symbols, and bars. All estimates (95% robust confidence intervals) are from the marginal structural model with statistical interaction of exposure group and socio-demographic characteristic (i.e., sex, age, race/ethnicity) and hospitalization status, adjusting for imbalanced covariates. Number of new cases per 1000 person-years are after marginal standardization to covariate distribution of analytic sample and estimated as absolute risk (B) and risk difference (C). Risk of T2DM in the post-acute period after SARS-CoV-2 infection for NH Other were not estimated due to small sample sizes.
The observed associations remained consistent under the different sensitivity analyses. After restricting to those with 100 days of follow-up and at least one lab encounter, the hazards relative to the historical cohort were higher for the unexposed (1.20 [95% CI, 1.07–1.35]) and exposed (1.57 [95% CI, 1.38–1.80]) cohorts as well as socio-demographic subgroups and hospitalization status. However, the absolute risk per 1000 person-years for exposed (37.0 cases [95% CI, 33.7–40.4]), unexposed (28.6 cases [95% CI, 27.2–30.1]), and historical (24.0 cases [95% CI, 21.3–26.6]) cohorts decreased in magnitude relative to the overall analytic sample (Supplementary Table 4). The SUPREME-DM computable phenotype identified 11,771 cases (exposed, 2257; unexposed, 7395; historical, 2119) of incident diabetes. Associations were similar upon using the SUPREME-DM computable phenotype (Supplementary Table 5), after adjusting for COVID-19 cases and test positivity on the index date (Supplementary Table 6), and after multiple imputation of missing laboratory parameters (Supplementary Table 7). Relative to the exposed cohort, the unexposed cohort had similar hazards of traumatic fractures (Supplementary Table 8) and higher hazards of otitis media (Supplementary Table 9). Relative to the exposed cohort, the historical cohort had lower hazards of traumatic fractures and similar hazards of otitis media. Using overlap weights did not change the observed associations (Supplementary Table 10).
DISCUSSION
Using a diverse, real-world data source, we observed the exposed and unexposed cohorts from the pandemic period both had higher new-onset T2DM risk than age and sex matched historical controls. Furthermore, in the post-acute phase of SARS-CoV-2 infection, patients who tested positive (exposed) had higher risk of new-onset T2DM than the contemporary controls who tested negative (unexposed). The excess risk of T2DM was similar across socio-demographic groups of age, sex, and race/ethnicity as well as hospitalization status. Our analysis also showed that those who tested positive for SARS-CoV-2 were more likely to be younger and female, to belong to a minority racial/ethnic group, and to have a higher BMI than those who tested negative.
SARS-CoV-2 infection was previously reported to be associated with new-onset T2DM, although magnitude of excess burden varied by definition and calendar year of the control group used.2 In a real-world study of over 8 million adults from the US Department of Veteran Affairs (VA), those with SARS-CoV-2 infection had 1.40 times higher risk and 13.46 excess cases per 1000 people at 12 months, relative to contemporary controls. In contrast with our study that suggests similar magnitude of association across socio-demographic groups, the VA study suggested the burden was higher among those hospitalized and admitted to intensive care, among non-Hispanic Black patients and those older than 65 years. Relative to historical or contemporary controls with and without acute respiratory infection, higher excess burden of new-onset T2DM were also observed among those with SARS-CoV-2 infection in other studies from USA, England, and Germany.10,12,13 However, no excess burden of new-onset T2DM was observed in another study among those with SARS-CoV-2 infection relative to contemporary and historical controls admitted for pneumonia in England.27 Therefore, whether the observed association of SARS-CoV-2 infection and T2DM is causal or can be alternatively explained by other mechanisms is unknown,2 particularly because individuals in both control populations included in this study were uninfected.
A variety of biological and social mechanisms have been proposed to explain the observed association between SARS-CoV-2 infection and new-onset T2DM. First, SARS-CoV-2 may infect pancreatic beta cells through the ACE2 receptors, mediated by Neuropilin-1 (NRP-1) and Transmembrane Protease, Serine 2 (TMPRSS2),28–31 and reduce insulin production, increasing the risk for diabetes.32 Second, risk factors for T2DM may predispose an individual to symptomatic SARS-CoV-2 infection and/or prompt COVID-19 testing.33 These individuals may eventually develop T2DM regardless of their COVID-19 history, but SARS-CoV-2 infection may accelerate this process or reduce time to T2DM detection. Third, other PASC-associated symptoms (such as shortness of breath and fatigue) may promote physical inactivity which increases T2DM risk.34 Fourth, reports of new-onset T2DM among patients hospitalized with SARS-CoV-2 may represent misclassification of time-limited sources of hyperglycemia associated with steroid therapy and/or stress from critical illness.11
The higher incidence of T2DM we observed in the unexposed cohort (36 per 1000 person years), relative to the historical cohort (28 per 1000 person years) suggests unique features of the pandemic period may also affect rates of T2DM. For example, shifts in patterns of healthcare utilization during the pandemic, particularly when shelter-in-place orders were in effect, may have altered the population of patients included in studies that utilize electronic medical record data which could impact reported rates of incident T2DM.35 Nevertheless, both rates are higher than the incidence of self-reported diabetes in the general population (6 per 1000 adults) from the National Health Interview Survey in 2021,36 although we do not know the burden of undiagnosed diabetes during the pandemic. The incidence rate of self-reported diabetes remained similar from 2000 to 2021 in USA.36
This study of new-onset T2DM after SARS-CoV-2 infection has several strengths. The validated computable phenotype for prevalent T2DM was used to exclude adults with pre-existing T2DM.16 The analytic sample is also among the largest of its kind nationally, and is more socio-demographically diverse and younger than other published reports. We used methods for target trial emulation, attempted to minimize instrumental variable bias in confounding adjustment, and explored negative controls to rule out spurious associations.
This real-world study also has several limitations. First, the process of excluding adults with pre-existing T2DM and then identifying incident T2DM using the same computable phenotype has not been previously validated.16 However, our sensitivity analysis using a modified version of SUPREME-DM phenotype suggested similar results.37 Second, although we used a high dimensional propensity score for confounding adjustment, there may still be residual confounding, including by reasons for visit as suggested by the results from the analysis with negative controls. Third, we were unable to account for patients testing positive and seeking care for diabetes at a healthcare system that does not contribute data to the OneFlorida + CRN. However, we were able to minimize misclassification due to home testing, since it became widespread only towards the end of follow-up for this study, i.e., February 2022. Finally, we note that the observed association may be potentially explained by unmeasured baseline differences in risk of T2DM and rates of screening for T2DM between cohorts, as suggested by differences in HbA1C values and missingness in the year prior to index date. The high rates of missingness for blood-based biomarkers is expected in real-world settings. However, we used different imputation approaches and high dimensional propensity scores to account for this unmeasured confounding and minimize concerns from missing data that may bias these findings.38
In conclusion, those who sought care during the pandemic period and tested positive for SARS-CoV-2 infection had the highest risks of new-onset T2DM relative to those who tested negative and historical controls. These associations were also observed among younger adults and racial/ethnic minorities who have been disproportionately impacted by COVID-19. Although the observed association between SARS-CoV-2 and T2DM, if causal, may be mediated through viral pathways, current observational studies cannot rule out alternative explanations such as higher T2DM risk or undiagnosed T2DM before infection and changes in healthcare utilization after infection leading to a higher likelihood of diagnosis. Given these uncertainties, it is unclear if the observed increased risk of T2DM among patients testing positive for SARS-CoV-2 portends a major shift in diabetes incidence in the USA. Post-acute care strategies for patients with SARS-CoV-2 infection should continue to emphasize guideline-concordant screening for diabetes across all socio-demographic groups. Future research should investigate mechanisms underlying the observed increased risk of incident T2DM to determine if this finding ultimately signals a significant shift in diabetes incidence.2
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements:
The authors thank the OneFlorida+ Data Trust team (Kathryn Shaw, Meggen Kaufman, Jiang Bian, Elizabeth Shenkman) for the support on query development and data extraction. The authors thank Shihab Chowdhury for the administrative support.
Abbreviations
- BMI
Body mass index
- CP
Computable phenotype
- EHR
Electronic health record
- HDL
High-density lipoprotein
- LDL
Low-density lipoprotein
- PASC
Post-acute sequelae of COVID-19
- T2DM
Type 2 diabetes
Author Contribution:
RJC, WTD, and JSV conceptualized the study with inputs from MKA and YG. JSV conducted the analysis. JSV wrote the first draft with inputs from RJC. All authors reviewed and edited subsequent drafts. RJC is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Funding
This research was supported by the National Institute for Diabetes and Digestive and Kidney Diseases (NIDDK) of the National Institutes of Health, award number 3R01DK120814-05S1. MKA was partially supported by the Georgia Center for Diabetes Translation Research which is funded by the National Institutes of Health (P30DK111024).
Data and Resource Availability
Information about OneFlorida+ CRN is provided at https://onefloridaconsortium.org/, and OneFlorida+ data are made available to researchers with an approved study protocol at https://onefloridaconsortium.org/front-door/prep-to-research-data-query/. For questions regarding OneFlorida+, email OneFloridaOperations@health.ufl.edu. The code for the analysis is available on https://github.com/jvargh7/pasc_diabetes.
Declarations:
Consent for Publication:
Not applicable.
Competing Interests:
The authors declare no competing interests.
Footnotes
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.World Health Organization. World Health Organization COVID-19 Emergency Dashboard [Internet]. 2023. Available from: https://covid19.who.int/region/amro/country/us. Accessed 22 Feb 2024
- 2.Cefalu WT. COVID-19 and Rising Incidence of Diabetes: Despite Evolving Data, an Enigma Still to Be Solved. Diabetes Care. 2023 May 1;46(5):913–5. [DOI] [PubMed] [Google Scholar]
- 3.Xie Y, Al-Aly Z. Risks and burdens of incident diabetes in long COVID: a cohort study. The Lancet Diabetes & Endocrinology. 2022;10(5):311–21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Zhang T, Mei Q, Zhang Z, Walline JH, Liu Y, Zhu H, et al. Risk for newly diagnosed diabetes after COVID-19: a systematic review and meta-analysis. BMC Med. 2022 Nov 15;20(1):444. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kim SH, Arora I, Hsia DS, Knowler WC, LeBlanc E, Mylonakis E, et al. New-Onset Diabetes After COVID-19. The Journal of Clinical Endocrinology & Metabolism. 2023 Oct 18;108(11):e1164–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Izzo R, Pacella D, Trimarco V, Manzi MV, Lombardi A, Piccinocchi R, et al. Incidence of type 2 diabetes before and during the COVID-19 pandemic in Naples, Italy: a longitudinal cohort study. eClinicalMedicine. 2023 Dec;66:102345. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Mude W, Oguoma VM, Nyanhanda T, Mwanri L, Njue C. Racial disparities in COVID-19 pandemic cases, hospitalisations, and deaths: A systematic review and meta-analysis. J Glob Health. 2021 Jun 26;11:05015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Dalsania AK, Fastiggi MJ, Kahlam A, Shah R, Patel K, Shiau S, et al. The Relationship Between Social Determinants of Health and Racial Disparities in COVID-19 Mortality. J Racial and Ethnic Health Disparities. 2022 Feb;9(1):288–95. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Beckles GL, Chou CF. Disparities in the Prevalence of Diagnosed Diabetes — United States, 1999–2002 and 2011–2014. MMWR Morb Mortal Wkly Rep. 2016 Nov 18;65(45):1265–9. [DOI] [PubMed] [Google Scholar]
- 10.Ayoubkhani D, Khunti K, Nafilyan V, Maddox T, Humberstone B, Diamond I, et al. Post-covid syndrome in individuals admitted to hospital with covid-19: retrospective cohort study. BMJ. 2021 Mar 31;n693. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Hernandez-Romieu AC, Carton TW, Saydah S, Azziz-Baumgartner E, Boehmer TK, Garret NY, et al. Prevalence of Select New Symptoms and Conditions Among Persons Aged Younger Than 20 Years and 20 Years or Older at 31 to 150 Days After Testing Positive or Negative for SARS-CoV-2. JAMA Netw Open. 2022 Feb 4;5(2):e2147053. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Daugherty SE, Guo Y, Heath K, Dasmariñas MC, Jubilo KG, Samranvedhya J, et al. Risk of clinical sequelae after the acute phase of SARS-CoV-2 infection: retrospective cohort study. BMJ. 2021 May 19;n1098. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Rathmann W, Kuss O, Kostev K. Incidence of newly diagnosed diabetes after Covid-19. Diabetologia [Internet]. 2022 Mar 16 [cited 2022 Apr 22]; Available from: https://link.springer.com/10.1007/s00125-022-05670-0 [DOI] [PMC free article] [PubMed]
- 14.Hogan WR, Shenkman EA, Robinson T, Carasquillo O, Robinson PS, Essner RZ, et al. The OneFlorida Data Trust: a centralized, translational research data infrastructure of statewide scope. Journal of the American Medical Informatics Association. 2022 Mar 15;29(4):686–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.PCORnet. Common Data Model (CDM) Specification, Version 6.0 [Internet]. 2022. Available from: https://pcornet.org/wp-content/uploads/2022/01/PCORnet-Common-Data-Model-v60-2020_10_221.pdf
- 16.Wiese AD, Roumie CL, Buse JB, Guzman H, Bradford R, Zalimeni E, et al. Performance of a computable phenotype for identification of patients with diabetes within PCORnet: The Patient‐Centered Clinical Research Network. Pharmacoepidemiol Drug Saf. 2019 May;28(5):632–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Benchimol EI, Smeeth L, Guttmann A, Harron K, Moher D, Petersen I, et al. The REporting of studies Conducted using Observational Routinely-collected health Data (RECORD) Statement. PLoS Med. 2015 Oct 6;12(10):e1001885. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zhang H, Zang C, Xu Z, Zhang Y, Xu J, Bian J, et al. Data-driven identification of post-acute SARS-CoV-2 infection subphenotypes. Nat Med. 2023 Jan;29(1):226–35. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Rassen JA, Blin P, Kloss S, Neugebauer RS, Platt RW, Pottegård A, et al. High‐dimensional propensity scores for empirical covariate selection in secondary database studies: Planning, implementation, and reporting. Pharmacoepidemiology and Drug. 2023 Feb;32(2):93–106. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Varghese JS, Guo Y, Ali MK, Donahoo WT, Chakkalakal RJ. Body mass index changes and their association with SARS-CoV-2 infection: a real-world analysis. International Journal of Obesity. 2024 September. https://www.nature.com/articles/s41366-024-01628-x [DOI] [PMC free article] [PubMed]
- 21.McCaffrey DF, Griffin BA, Almirall D, Slaughter ME, Ramchand R, Burgette LF. A tutorial on propensity score estimation for multiple treatments using generalized boosted models. Statist Med. 2013 Aug 30;32(19):3388–414. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Weuve J, Tchetgen Tchetgen EJ, Glymour MM, Beck TL, Aggarwal NT, Wilson RS, et al. Accounting for Bias Due to Selective Attrition: The Example of Smoking and Cognitive Decline. Epidemiology. 2012 Jan;23(1):119–28. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Robins JM, Hernán MÁ, Brumback B. Marginal Structural Models and Causal Inference in Epidemiology: Epidemiology. 2000 Sep;11(5):550–60. [DOI] [PubMed] [Google Scholar]
- 24.Lee H. Causal Inference for Real-World Evidence: Propensity Score Methods and Case Study [Internet]. The ASA Biopharmaceutical Section Regulatory-Industry Statistics Workshop: US Food & Drug Administration; 2020 Sep. Available from: https://ww2.amstat.org/meetings/biop/2020/onlineprogram/handouts/SC4-Handouts.pdf. Accessed 22 Feb 2024
- 25.Rubin DB, Thomas N. Combining Propensity Score Matching with Additional Adjustments for Prognostic Covariates. Journal of the American Statistical Association. 2000 Jun;95(450):573–85. [Google Scholar]
- 26.Li F, Thomas LE. Addressing Extreme Propensity Scores via the Overlap Weights. American Journal of Epidemiology [Internet]. 2018 Sep 5 [cited 2024 Mar 22]; Available from: https://academic.oup.com/aje/advance-article/doi/10.1093/aje/kwy201/5090958
- 27.Holman N, Barron E, Young B, Gregg EW, Khunti K, Valabhji J, et al. Comparative Incidence of Diabetes Following Hospital Admission for COVID-19 and Pneumonia: A Cohort Study. Diabetes Care. 2023 May 1;46(5):938–43. [DOI] [PubMed] [Google Scholar]
- 28.Wu CT, Lidsky PV, Xiao Y, Lee IT, Cheng R, Nakayama T, et al. SARS-CoV-2 infects human pancreatic β cells and elicits β cell impairment. Cell Metabolism. 2021 Aug;33(8):1565-1576.e5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Lim S, Bae JH, Kwon HS, Nauck MA. COVID-19 and diabetes mellitus: from pathophysiology to clinical management. Nat Rev Endocrinol. 2021 Jan;17(1):11–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Tang X, Uhl S, Zhang T, Xue D, Li B, Vandana JJ, et al. SARS-CoV-2 infection induces beta cell transdifferentiation. Cell Metabolism. 2021 Aug;33(8):1577-1591.e7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Jackson CB, Farzan M, Chen B, Choe H. Mechanisms of SARS-CoV-2 entry into cells. Nat Rev Mol Cell Biol. 2022 Jan;23(1):3–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Apicella M, Campopiano MC, Mantuano M, Mazoni L, Coppelli A, Del Prato S. COVID-19 in people with diabetes: understanding the reasons for worse outcomes. The Lancet Diabetes & Endocrinology. 2020 Sep;8(9):782–92. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Landstra CP, de Koning EJP. COVID-19 and Diabetes: Understanding the Interrelationship and Risks for a Severe Course. Front Endocrinol. 2021 Jun 17;12:649525. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Gil S, Gualano B, Araújo AL, Júnior GNO, Damiano RF, Pinna F, et al. Association between Post-acute Sequelae of SARS-CoV-2 and Physical Inactivity: A Cohort Study of 614 COVID-19 survivors [Internet]. In Review; 2022 May [cited 2022 May 29]. Available from: https://www.researchsquare.com/article/rs-1638885/v1
- 35.Chen JL, Krupp GR, Lo JY. The COVID-19 Pandemic and Changes in Health Care Utilization Among Patients With Type 2 Diabetes. Diabetes Care. 2022 Apr 1;45(4):e74–6. [DOI] [PubMed] [Google Scholar]
- 36.Centers for Disease Control & Prevention. National Diabetes Statistics Report [Internet]. 2023 [cited 2024 Feb 13]. Available from: https://www.cdc.gov/diabetes/data/statistics-report/index.html
- 37.Nichols GA, Schroeder EB, Karter AJ, Gregg EW, Desai J, Lawrence JM, et al. Trends in Diabetes Incidence Among 7 Million Insured Adults, 2006–2011. American Journal of Epidemiology. 2015 Jan 1;181(1):32–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Madley-Dowd P, Hughes R, Tilling K, Heron J. The proportion of missing data should not be used to guide decisions on multiple imputation. Journal of Clinical Epidemiology. 2019 Jun;110:63–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Information about OneFlorida+ CRN is provided at https://onefloridaconsortium.org/, and OneFlorida+ data are made available to researchers with an approved study protocol at https://onefloridaconsortium.org/front-door/prep-to-research-data-query/. For questions regarding OneFlorida+, email OneFloridaOperations@health.ufl.edu. The code for the analysis is available on https://github.com/jvargh7/pasc_diabetes.