The Strong Heart Study (SHS) and Strong Heart Family Study (SHFS) are population-based longitudinal studies of CVD in several American Indian communities in Arizona, North and South Dakota, and Oklahoma. Details of the study designs have been described previously.(21) Briefly, in 1989, SHS recruited a cohort of 4,549 individuals who participated in clinical exams in 1989-91, 1993-95, and 1998-99. At the 1998-99 exam, 1st, 2nd, and 3rd degree relatives of SHS participants were invited to participate in a family study, which became the Strong Heart Family Study (SHFS). The SHFS included two examinations: 2001-03 and 2006-09. The institutional review boards from each Indian Health Service region and all communities approved the studies, and written informed consent was obtained from all participants at each exam.
Analyses were restricted to participants with prevalent diabetes at the time of their sphingolipid measurement. Prevalent diabetes was defined as a fasting plasma glucose ≥ 126 mg/dl or use of insulin or oral anti-diabetic medications at the relevant study exam; in SHS, where 2-hour oral glucose tolerance tests were performed, we also identified prevalent diabetes as an oral glucose tolerance test ≥ 200 mg/dL. (22).
Sphingolipids were measured prospectively at two timepoints in SHFS, allowing for the inclusion of 508 participants with prevalent diabetes at the 2001-03 exam and an additional 202 participants who were subsequently diagnosed with diabetes before or at the 2006-09 exam. Of these 710 participants, we excluded 96 who had a history of CVD at the time of their sphingolipid measurement and 17 who had no available follow-up CVD data, leaving 597 SHFS participants eligible for this analysis.
In SHS, we used a nested case-control design and samples from the 1993-95 exam. Selected cases and controls had prevalent diabetes and no history of CVD at the 1993-1995 exam. For cases, we randomly selected 100 participants out of 444 who subsequently developed CVD during follow-up in order to preserve scarce samples. We used risk set matching; for each case, we selected two age (within 5 years) and sex matched controls who had not developed incident CVD. Of the 300 SHS participants we sampled, 281 had sufficient blood samples and were able to have their sphingolipids measured; 14 participants from the Arizona site were excluded due to a lack of appropriate controls (13 cases and 1 control), leaving 267 participants (78 cases and 189 controls) in the analysis.
Each SHFS and SHS examination included a standardized personal interview, physical examination, and laboratory work-up. The data collection procedures were identical at all SHFS and SHS exams except for physical activity and diabetes ascertainment; these procedures have been described in detail previously (21, 23).
Sphingolipids were measured from blood samples collected at the 2001-03 and 2006-09 SHFS study examinations and the 1993-95 SHS study examination. The same collection protocols were used at each study exam; blood samples were collected after a 12-hour fast and were stored at -70°C until analyzed. A detailed description of the laboratory procedures can be found elsewhere; briefly, plasma sphingolipids were quantified from baseline EDTA-blood samples using liquid chromatography-tandem mass spectrometry.(24)
This analysis includes eight sphingolipids that carry a saturated fatty acid acylated to the sphingoïd backbone: ceramide and sphingomyelin with palmitic acid (16:0 [16 carbons, 0 double bonds]; Cer-16 and SM-16), with arachidic acid (20:0; Cer-20 and SM-20), with behenic acid (22:0; Cer-22 and SM-22), and with lignoceric acid (24:0; Cer-24 and SM-24).
Incident Cardiovascular Disease
Incident CVD was defined as the first occurrence of myocardial infarction (MI; definite, probable, or fatal), ischemic stroke (definite or fatal), atherosclerotic cardiovascular disease (ASCVD; definite), or heart failure (HF; including fatal). Incident events in both SHFS and SHS were identified at each study examination and through subsequent surveillance of medical records and death certificates; after the 2006-09 examination and through December 2017, cases were ascertained through phone interviews with the participants and confirmed by medical record documentation. Stroke and HF were identified via chart review, MI was identified through chart reviews and/or evidence of MI by study visit ECG, and ASCVD was defined as incident MI, coronary heart disease (CHD), ischemic stroke, or abnormal study ECG with positive Rose Angina Questionnaire.
In the SHFS cohort, Cox regression models were used to examine the associations of each sphingolipid with the risk of incident CVD. SHFS participants began accruing time-at-risk at the earliest study exam at which they had prevalent diabetes and measured sphingolipids (n=434 from the 2001-03 exam, and n=163 from the 2006-09 exam), and were followed to the earliest of: date of incident CVD, death, or loss to follow up. Because the SHFS is comprised of extended families, robust standard error estimates that account for clustering of risk factors among family members were used. In the SHS cohort, logistic regression with robust standard errors was used to evaluate the associations of plasma sphingolipids and incident CVD.
We used 3 sets of models to examine associations of sphingolipids with incident CVD. Model 1 (minimally-adjusted model) included terms for age, sex, and study site (SHFS only). Model 2 (multivariable model) additionally adjusted for education (years), smoking (never, former, current), alcohol consumption (never, former, current), physical activity (in SHFS: average number of pedometer measured steps per day; in SHS: metabolic equivalent task hours per week), Body Mass index (BMI; log-transformed), waist circumference, low density lipoprotein (LDL), systolic blood pressure, treated hypertension, duration of diabetes (log-transformed), type of diabetes medication used (insulin + oral hypoglycemic, insulin only, oral hypoglycemic only, or neither), and treated hyperlipidemia. Model 3 (sphingolipid adjusted model) included Model 2 with additional adjustment for one of the other species: Cer-16 and SM-16 models include adjustment for Cer-22 and SM-22, respectively; Cer-20, -22, -24, and SM-20, -22, and -24 models include adjustment for Cer-16 and SM-16, respectively. In both cohorts, adjustment covariate values were drawn from the same study exam at which the sphingolipids were measured. Sphingolipid exposures were log base-2 transformed, so the relative risk represents the associations with incident CVD per doubling of each of the sphingolipid species concentrations (µM), which is roughly equivalent to the difference between the 10th and 90th percentiles (Table 2). With the exception of study site in SHS, the same set of adjustment variables were used in each of the models for both SHFS and SHS cohorts.
As a secondary analysis, we evaluated associations of sphingolipids with each of incident MI, stroke, heart failure, and ASCVD within the SHFS cohort. These models followed the same analytic approach as the primary analysis.
We conducted a number of sensitivity analyses. To assess whether associations were further confounded by high density lipoprotein (HDL), fibrinogen, or triglyceride levels, or by estimated glomerular filtration rate (eGFR), we repeated the multivariable analyses with additional terms for HDL, triglycerides, and fibrinogen in one model and chronic kidney disease (CKD; eGFR<60) in another. We also examined whether associations of sphingolipids with incident CVD differed by age, sex, BMI, CKD, or duration of diabetes by adding product interaction terms to the multivariable models.
Inverse-variance-weighted fixedeffects meta-analyses were conducted to combine results from the two cohorts using the log hazard ratios (SHFS), log odds ratios (SHS), and their corresponding standard errors in STATA 16.0 (Stata Corporation, College Station, TX); for the purposes of this manuscript, we refer to resulting meta-analyzed parameter as “relative risk.” Inverse-variance-weighted fixed-effects meta-analyses approximate results that would be obtained if the data from all studies could be analyzed together with adjustment for study (25). Heterogeneity was assessed via I2 variance (26).
Multiple imputation with chained equations was used to impute missing values for smoking (SHFS: n=2; SHS: n=9), alcohol consumption (SHFS: n=30; SHS: n=5), education (SHFS: n=9), BMI (SHFS: n=7), waist circumference (SHFS: n=4), LDL (SHFS: n=10; SHS: n=18), duration of diabetes (SHFS: n=32; SHS: n=1), physical activity (SHFS: n=206; SHS: n=12), systolic blood pressure (n=1), fibrinogen (SHFS: n=3; SHS: n=3), triglycerides (SHFS: n=3; SHS: n=1), HDL (SHFS: n=7; SHS: n=2), and CKD (SHFS: n=2; SHS: n=4) using information on age, sex, site, treated hypertension, treated hyperlipidemia, and type of diabetes treatment. (18, 27, 28). Schoenfeld residuals were evaluated to test the proportional hazards assumption in SHFS Cox models; Martingale residuals were used to assess the linearity of associations; and delta-betas were reviewed to assess the impact of potential influential outliers.
A Bonferroni correction was used to adjust for multiple comparisons; a significance threshold of 0.0063 (0.05/8 sphingolipid species) was used.