The study inclusion criteria were any individual who: (1) donated allogeneic whole blood or plasma (2009–2018) at a Vitalant blood collection location that had implemented the current electronic database system; and (2) completed ≥1 COVID-19 survey (2021–2023). At each blood donation visit, every individual met all Food and Drug Administration (FDA) eligibility requirements for routine blood donations, such as age (≥16 years) and weight (≥110 lb). We excluded participants who had missing key variables, specifically: (1) survey responses regarding COVID-19 hospitalization and PCC; (2) cardiometabolic indicators (systolic and diastolic blood pressure, total circulating cholesterol, BMI) at ≥3 donation timepoints; and (3) other key covariates. Based on these eligibility criteria, the study participant flow diagram is in Supplementary Fig. 1.

During every routine blood donation, demographic and cardiometabolic indicators are collected by donor health questionnaire, donation site staff, and blood assays. Self-reported information included demographics (e.g., age, gender, race-ethnicity, height [in], and weight [lb], geographic location). During each donation visit, donation site staff measured the donor’s diastolic and systolic blood pressure (millimeters of mercury [mm Hg]) by automated digital sphygmomanometer. If the first blood pressure measurement was within prespecified ranges (90–180 mm Hg for systolic, 50–100 for diastolic), this measurement was recorded. If the first blood pressure measurement was outside these ranges, a second measurement was recorded and included in this analysis. Following blood collection organization donation eligibility criteria, donors with blood pressure outside of the prespecified ranges were deferred from donating. As exceptions, a medical director can approve allogeneic donors based on in person assessment of the donor and their completed health questionnaire; donors that are otherwise healthy and at low risk for adverse consequences caused by blood donation can be cleared for donations. Non-fasting blood samples were assayed for total cholesterol concentrations (mg/dL) by automated clinical chemistry analyzers (e.g., Beckman Coulter AU analyzer [Brea, CA]) by laboratory staff at Creative Testing Solutions.

As part of a larger CDC-funded COVID-19 serosurveillance study [34, 35], blood donors were invited to complete electronic COVID-19 surveys in multiple waves between 2021–2023. Surveys included questions regarding COVID-19 symptomology and hospitalizations as well as PCC.

PCC was defined as any health consequences (e.g., signs, symptoms, conditions) present ≥4 weeks after initial acute SARS-CoV-2 infection, per the US Department of Health and Human Services (DHHS) definition [36]. BMI was calculated as weight (kg) divided by height squared (m²) and categorized with standard World Health Organization (WHO) categories [37]. For each participant, we defined a baseline visit as the first donation date that was included in this analysis.

Intraindividual variability was considered as the coefficient of variation (CV) of a cardiometabolic indicator across all available timepoints for the participant. For example, if a donor had 14 donations during the study time period, their intraindividual variability of cholesterol was calculated as the CV across all 14 measurements of total cholesterol (mg/dL). Donors with the top 20 highest BMI CV are visually illustrated in Fig. 1. For each cardiometabolic indicator (BMI, systolic and diastolic blood pressure, total circulating cholesterol), we categorized participants into quartiles of CV (Fig. 1).

As descriptive statistics, we reported measures of central tendency as the mean (SD) or median (IQR) of continuous variables and N (%) for categorical variables. We compared differences between subgroups with Wilcoxon rank-sum and Mantel-Haenszel chi-squared test statistics. Statistical significance was based on two-sided tests and alpha values of 0.05. SAS (version 9.4; SAS Institute; Cary, North Carolina), GraphPad Prism (version 9.3.1.; GraphPad Software, LLC; San Diego, California), and BioRender.com (Toronto, Ontario) were used for statistical analysis and visualizations. We used a complete-case analysis approach to address missingness of key variables (Supplementary Fig. 1).

We assessed the associations between intraindividual variability of cardiometabolic health indicators and the risk of COVID-19 hospitalization or PCC by fitting binomial regressions. COVID-19 related hospitalization and PCC were the primary dependent variables of interest. Intraindividual variability of cardiometabolic indicators (CV quartiles of BMI, systolic and diastolic blood pressure, total cholesterol) were the independent variables of interest; in all models, the lowest quartile was the reference group. For our confounder selection approach, we initially selected a set of potential covariates a priori per previous literature [38]. We first reported all univariable regressions for bivariate associations between each independent variable and a primary dependent variable; subsequently, we evaluated fully adjusted multivariable regression results. To confirm the dose-response, we also evaluated the linear test for trend between cardiometabolic indicator variability quartiles and each COVID-19 outcome in the final multivariable models. As a subgroup analysis, we evaluated the prevalence of specific PCC symptoms among donors with PCC and available symptomology data. Among donors with a particular PCC symptom (e.g., brain fog), we compared the proportions in differing BMI CV quartiles. As a sensitivity analysis, we utilized a log-binomial causal mediation model framework [39] to evaluate COVID-19 hospitalization as a mediator of the association between high BMI variability (highest CV quartile vs other quartiles) and PCC. Specifically, we considered a single mediator model with the highest quartile of BMI CV as a binary variable, with the SAS CausalMed procedure and alpha values as 0.05. We calculated estimates of bootstrap-corrected standard errors and confidence intervals with 1000 replicates and an alpha value of 0.05 as the confidence level for constructing bootstrap intervals.

At baseline, the median age of participants was 48 years (IQR 38, 56; Table 1); ages ranged between 16–80 years. 59.2% of participants identified as female. The proportions of self-reported race-ethnicity were: White, non-Hispanic (87.6%); Black, non-Hispanic (0.8%), and other races and ethnicities (11.6%). In terms of highest educational attainment, 0.8% had not completed high school, 37.5% had a high school diploma, 37.1% had a bachelor’s degree, and 24.6% had a graduate degree. The baseline median values of cardiometabolic indicators were: 26.6 kg/m² (IQR 23.7, 30.2) BMI, 183.0 mg/dL (IQR 161.0, 209.0) total circulating cholesterol, 121.0 mm Hg (IQR 112.0, 132.0) systolic blood pressure, 76.0 mm Hg (IQR 70.0, 83.0) diastolic blood pressure.

Based on WHO categories of BMI, there were: 0.2% of participants with underweight, 36.5% with normal weight, 36.9% with overweight, and 26.5% with obesity (Table 1). The percentages of participants in BMI categories (under- or normal weight, overweight, obesity) differed by their BMI CV quartile (p < 0.01; Fig. 2). Among participants with obesity, a greater proportion were in the highest variability subgroup (4th CV quartile; 35.8%) and a lower proportion were in the lowest variability subgroup (1st CV quartile 16.5%).

At the first donation visit, systolic blood pressure (126.5 mm Hg [IQR 117.0, 140.0]) was higher among participants with hospitalization, compared to among those without (121.0 [112.0, 132.0]; p < 0.05). Median BMI was higher among donors with COVID-19 related hospitalization (30.4 kg/m² [IQR 27.5, 34.1]), relative to those without PCC (26.5 [23.6, 30.2]; p < 0.01; Table 1). Median total cholesterol and diastolic blood pressure did not significantly differ by hospitalization (both p > 0.05).

At the first donation visit, median diastolic blood pressure (78.0 mm Hg [IQR 71.0, 83.0]) was higher among participants with PCC, compared to among those without PCC (76.0 [70.0, 82.0]; p = 0.02; Table 1). Median BMI was higher among donors with PCC (27.5 kg/m² [IQR 24.2, 31.7]), relative to those without PCC (26.4 [23.6, 30.0]; p < 0.01). Systolic blood pressure and total cholesterol were similar by PCC status (both p > 0.05).

Comparing intraindividual variability of cardiometabolic health, the median CV of total cholesterol was higher among those hospitalized from COVID-19 (9.3 [IQR 7.3, 12.7]), compared to those without hospitalization (8.4 [IQR 6.3, 10.9]; p = 0.04; Supplementary Table 1). Median CV of BMI, systolic and diastolic blood pressure did not differ by hospitalization (all p > 0.05). The median CV of BMI was higher among those with PCC (3.1 [IQR 2.1, 4.8], relative to those without PCC (2.8 [IQR 1.8, 4.3]; p < 0.01). Median CV of blood pressure (systolic, diastolic) and cholesterol did not differ by PCC (all p > 0.05).

BMI variability was positively associated with COVID-19 hospitalization risk (4th quartile aRR 4.15 [95% CI 1.31, 13.11], p = 0.02; 3rd quartile aRR 3.41 [95% CI 1.09, 10.69], p = 0.04) relative to the lowest [1st] quartile and accounting for age, gender, geographic region, and the total number of donations (Fig. 3B and Supplementary Table 2). In this multivariable regression, we also confirmed the dose-response (linear test of trend) of BMI quartiles in association with hospitalization risk (p_trend = 0.02). Total cholesterol, systolic and diastolic blood pressure variability were not significantly associated with COVID-19 hospitalization risk (all p > 0.05); none of the trend tests were significant (all p_trend > 0.05).

Participants with higher variability of BMI had greater risk of PCC (4th CV quartile aRR 1.29 [95% CI 1.02, 1.64]; p = 0.04), relative to those with low variability (1st quartile), adjusting for age, gender, race-ethnicity, educational attainment, geographic region, COVID-19 vaccination, and total number of donations between 2009–2008 (Fig. 3A and Supplementary Table 3). In this multivariable regression, we also confirmed the dose-response (linear test of trend) of BMI quartiles in association with PCC risk (p_trend = 0.03). Intraindividual variability (CV quartiles) of blood pressure (systolic, diastolic) and total cholesterol were not significantly associated with PCC risk (all p > 0.05); tests for trend were also not significant (all p_trend > 0.05).

Among 519 participants who had PCC, 77.8% reported specific PCC symptomology. Prevalent PCC symptoms included fatigue or weakness (72.3%), difficulty thinking or concentrating (60.4%), change in taste (49.0%), and change in smell (48.4%), cough (46.0; Fig. 4A). Proportions of donors in BMI CV quartile subgroups differed by three of the top 10 PCC symptoms (fatigue or weakness, difficulty thinking or concentrating, difficulty sleeping; all p < 0.05; Fig. 4B).

Based on causal mediation modeling framework and associated terminology, the total effect of the association between the highest quartiles of BMI CV, relative to all lower quartiles, and PCC risk was significant (aRR 1.24 [95% CI 1.03, 1.47; p < 0.05; Fig. 4C). Assessing COVID-19 hospitalization as a proxy for disease severity and mediator, the decomposed natural direct effect was aRR 1.22 (95% CI 1.02, 1.45; p < 0.05) and the natural indirect effect was aRR 1.01 (0.98, 1.04; p > 0.05; Fig. 4C).

There were several key limitations of this study. First, there is no standard measurement of intraindividual variability. However, with respect to our primary outcomes, it is likely that unmeasured variability was non-differential, which would potentially bias results towards the null hypothesis. Second, COVID-19 hospitalizations and PCC were self-reported, and could have potential recall and missing data bias. There is no universal definition of PCC, including symptoms and duration of persistence (e.g., >4, >8, or >12 weeks); since this remains an active area of investigation, self-reporting of PCC is subjective [7, 57]. Heterogeneous definitions of PCC are currently used in the literature [58] and recommended by the WHO [59], US DHHS (CDC) [36], and UK National Institute for Health and Care Excellence (NICE) [60]. These multiple PCC definitions further exacerbate the adjacent challenges, which span from elucidating biological mechanisms of immunopathogenesis [61], prevalence, incidence, modifiable risk factors, etiology, and prevention and treatment guidelines. SARS-CoV-2 diagnostics (availability, accessibility, quality) was extensively variable in the US during the early phase of the COVID-19 pandemic, therefore the self-reporting of whether hospitalization was due to COVID-19 could be less accurate in this period. Third, we did not have viral strain sequencing data to account for the SARS-CoV-2 variants and sub-variants, in light of differing immune escape capacity [62], that caused infections resulting in severe COVID-19 and PCC. We also did not have complete data regarding other key factors, such as previous infection history, vaccination (e.g., COVID-19 vaccines [dosages, timing]), prior clinical diagnoses (e.g., Type 2 diabetes mellitus, ischemia heart disease), other routine clinical assessments (e.g., HDL- and LDL-cholesterol, waist circumference), medication history (e.g., statins, nirmatrelvir or COVID-19 antivirals, blood pressure and diabetes medications), socioeconomic status, and lifestyle behaviors (diet, physical activity) for all participants. Fourth, our causal mediation model was limited by the relatively small numbers of individuals with COVID-19 hospitalization and PCC. We were not able to comprehensively account for other potential confounders, given small sample cell sizes. We considered this a sensitivity analysis of our primary findings. Our findings emphasize the need to identify specific biological mechanisms underlying this association as well as larger observational studies to evaluate causal inference models. Fifth, blood donors meet the FDA and blood collection organization donor eligibility criteria at each successful donation timepoint; therefore, blood donors are not representative of the general population. Relatedly, there is potential competing risk bias, especially due to COVID-19 related mortality.

One major strength of this study was the larger number of timepoints (median of 9 donations [IQR 5, 17] per participant) with complete cardiometabolic indicators over a 10-year exposure window (2009–2018) prior to first global detection of SARS-CoV-2 infection cases during 2019. This allowed for characterization of intraindividual variability of cardiometabolic health indicator data during period that was temporally antecedent of SARS-CoV-2 infections and COVID-19 consequences. Second, cardiometabolic indicators were collected during and assayed from routine blood donations, which minimizes the potential for recall bias of exposures. Third, the multi-site study design allowed for the inclusion of free-living individuals in their late adolescence and adulthood in 24 US states.