Association of BMI, Fitness, and Mortality in Patients With Diabetes: Evaluating the Obesity Paradox in the Henry Ford Exercise Testing Project (FIT Project) Cohort

There is conflicting evidence on whether BMI is associated with adverse cardiovascular outcomes and mortality in patients with diabetes. A number of studies have demonstrated a lower mortality risk among individuals with diabetes who are overweight or obese compared with normal-weight individuals, a finding that has been termed the “obesity paradox” (1,2). Conversely, other studies among individuals with diabetes have not demonstrated a lower mortality risk or have demonstrated a higher risk of mortality for overweight or obese individuals (3,4).

Understanding whether the obesity paradox exists among patients with type 2 diabetes is of particular importance, because a higher BMI is one of the strongest risk factors for the development of type 2 diabetes, which is independently associated with an higher risk for cardiovascular disease (CVD) and all-cause mortality (5–7). Overweight or obese individuals with a high fitness level have been termed “fat but fit,” and it has been suggested that individuals with a higher BMI who are metabolically healthy may partly account for the observed obesity paradox (8,9). While there is significant heterogeneity in CVD risk for patients with diabetes, they are often considered as a CVD risk equivalent group, and diabetes is one of the four groups identified by the 2018 American Heart Association/American College of Cardiology Cholesterol Guidelines to benefit from statin therapy (10,11). However, fitness modifies the relationship between BMI and mortality among patients with CVD, and we therefore hypothesized that 1) among individuals with diabetes, those with a higher fitness level would not have a paradoxical relationship between BMI and mortality; and 2) individuals with a higher fitness level would have a lower risk of mortality regardless of BMI category (12,13). Accordingly, we investigated the association between obesity and mortality among individuals with diabetes in the Henry Ford Exercise Testing Project (FIT Project) and whether fitness modified this relationship.

This analysis includes 8,528 individuals with diabetes from the Henry Ford Exercise Testing Project (FIT Project) who performed a clinically indicated, physician-referred Bruce protocol exercise treadmill stress test between 1991 and 2009 at the Henry Ford Health System medical centers in metropolitan Detroit, MI, as has previously been described in detail elsewhere (14). Patients were at least 18 years old and had a diagnosis of diabetes, which was defined by patient self-report, use of a blood glucose–lowering medication, or based on electronic medical record (EMR) diagnosis. We required that an EMR diagnosis of diabetes or any other medical condition (e.g., hypertension or hyperlipidemia) be coded on at least three separate encounters in the EMR in order to be included as a diagnosis in our database. Among patients diagnosed with diabetes, 85% had an HbA_1c ≥6.5% and/or were taking a glucose-lowering medication, 11% had an HbA_1c ≥5.7–6.4% (39–46 mmol/mol), and 4% had an HbA_1c <5.7% (39 mmol/mol). We excluded persons with a BMI <18.5 kg/m² (n = 252) and those with prevalent cancer (n = 604).

Total mortality was the primary outcome and was ascertained through a search of the Social Security Death Index with follow-up through the year 2013. A previously described algorithm using a combination of first name, last name, date of birth, and Social Security number was used to perform matching (14). Follow-up was calculated from date of the exercise test to the date of death or through April 2013.

Bruce protocol treadmill stress testing was performed using standard methodology, and the test was stopped if the patient experienced chest pain, dyspnea, or other exercise-limiting symptoms (e.g., chest pain, dyspnea, or dizziness) as determined by the supervising clinician or if the patient requested that the test be stopped. The test could also have been stopped if the patient had an abnormal blood pressure response, significant ST segment changes, or a clinically significant arrhythmia as defined by the American Heart Association/American College of Cardiology guidelines (15,16). Each patient’s maximal exercise capacity (e.g., fitness) was estimated by calculating their METs, which were calculated by the Quinton treadmill controller (Q-Stress; Quinton Instrument Company, Bothell, WA) using their peak exercise workload (treadmill speed and grade) achieved during the stress test based on equations published by the American College of Sports Medicine (17). We categorized fitness as low (<6 METs), moderate (6–9.9 METs), and high (≥10 METs), as consistent with our prior work (18). The stress test indication was categorized into common indications based on the physician referral information, which primarily included chest pain, dyspnea, and preoperative evaluation.

A trained nurse and/or clinical exercise physiologist recorded the patients’ demographics and CVD risk factors along with current medication use and past medical history immediately preceding the treadmill stress test. Patients reported their race, height, and current smoking status. Weight was measured at the time of the treadmill stress test, and the EMR recorded weight was used for any missing values. BMI was categorized as normal (18.5–24.9 kg/m²), overweight (25–29.9 kg/m²), or obese (≥30 kg/m²). A diagnosis of hypertension and hyperlipidemia was based on patient self-report, the use of a disease-specific medication, or a database-verified diagnosis. Patients were classified as having a family history of coronary artery disease if they reported a first-degree relative with a history of a clinical coronary artery disease event. Laboratory values for tests performed within 90 days of the stress test were obtained through a retrospective search of the EMR and associated laboratory databases. Hemoglobin A_1c values were only available for 5,786 individuals (68%), and fasting glucose values were not available. For patients who participated in the Henry Ford Health System integrated health plan, a retrospective search of the EMR, administrative databases, and/or pharmacy claims files was performed to obtain additional data on medication use and past medical history.

We calculated age-adjusted mortality rates per 1,000 person-years’ follow-up stratified by BMI and fitness group. We also performed progressively adjusted Cox proportional hazards modeling to examine the association of BMI and total mortality within each fitness group. Using Cox proportional hazards modeling, we also examined the association of fitness and total mortality within each BMI group. Model 1 included age, sex, and ethnicity. Model 2 additionally adjusted for hypertension, current smoking, hypertension medication use, lipid-lowering medication use, oral glucose-lowering medication use, and a history of CVD. Model 3 additionally included insulin use. We also used Cox proportional hazards modeling within each fitness group to examine whether there were differences in the relationship between BMI and mortality for prespecified subgroups of interest.

We performed adjusted cubic spline modeling with BMI as a continuous variable and a reference value of BMI of 22.5 kg/m² (consistent with prior publications) within each fitness group in order to examine the continuous association of BMI with total mortality (19). We also calculated an adjusted cubic spline figure that displays the relative association for each fitness group using a reference value of BMI of 22.5 kg/m² for the lowest fitness group.

We performed multivariable-adjusted logistic regression modeling (model 3) to evaluate the association between 1) per 1 kg/m² increase in BMI and 2) 1 MET increase in fitness. In addition, we performed interaction testing using BMI and fitness as continuous variables. Statistical analyses were conducted using Stata/SE version 15.1 (Stata Corporation, College Station, TX).

Overall, the mean age was 57.9 years (SD 11.3), 49% of individuals were women, 40% were African American, and there were 1,319 deaths over a mean follow-up of 10.0 years (SD 4.1) (Table 1). In general, individuals who were obese were less likely to exercise to ≥10 METs or have a diagnosis of coronary artery disease, but they were more likely to have traditional CVD risk factors, be prescribed a glucose-lowering medication, and have a higher hemoglobin A_1c. Patients in the highest fitness group were younger, were less likely to be a woman, and had a lower prevalence of traditional CVD risk factors (Supplementary Table 1). Patients in the highest fitness group had the lowest age-adjusted mortality rate, and there was little absolute difference in the mortality rate between normal-weight (8/1,000 person-years) and obese patients (5/1,000 person-years) who achieved ≥10 METs (Fig. 1). The age-adjusted mortality rate for patients who achieved <6 METs was more than double that of patients who achieved six to nine METs regardless of BMI. Within each categorical fitness group, there was an ∼30% lower risk for total mortality for patients who were obese compared with normal weight, except for patients in the highest fitness group in whom the association was directionally similar, but not significant (hazard ratio 0.72 [95% CI 0.49–1.10]) (Table 2). For example, within the least fit group (<6 METs), the hazard for total mortality was 0.74 (95% CI 0.60–0.93) for obese compared with normal-weight patients.

Within each BMI group, there was an ∼70% lower risk for total mortality for patients in the highest fitness group (≥10 METs) compared with the least fit patients (<6 METs), which was consistent even after adjusting for traditional CVD covariables (Supplementary Table 2). For example, among obese patients, the hazard for total mortality was 0.26 (95% CI 0.20–0.35) for the most fit compared with least fit group.

In adjusted cubic spline modeling, we found a consistent and significantly lower risk for all-cause mortality with higher BMI for patients in the lowest fitness group (Fig. 2). Among patients in the moderate fitness group, the association between BMI and mortality was significant only above a BMI of ∼30 kg/m². Among patients in the highest fitness group, the association between BMI and all-cause mortality was lower only among patients with a BMI of ∼33–43 kg/m². We performed additional cubic spline modeling using the lowest fitness group and a BMI of 22.5 kg/m² as the reference point, which showed a uniformly lower hazard for all-cause mortality among patients with higher fitness levels, and that slope flattened with higher fitness (Supplementary Fig. 1). In addition, the association between BMI and a lower mortality risk flattened with higher fitness. We observed a similar relationship between BMI and mortality for each subgroup across the fitness categories, supporting the consistency of our overall findings (Supplementary Fig. 2).

Multivariable logistic regression modeling showed that for every 1 kg/m² increase in BMI, there was a 0.96 (95% CI 0.95–0.97; Z score −6.94) lower odds of mortality, while for every 1 MET increase in fitness, there was a 0.76 (95% CI 0.73–0.78; Z score −17.95) lower odds of mortality. Interaction testing using continuous values for BMI and fitness demonstrated a P value of 0.02.

Our results demonstrate that in the overall cohort, there was an inverse relationship between BMI and total mortality for patients with diabetes. However, when we examined this relationship by fitness level, the relationship was only consistently significant among patients in the lowest fitness group. We also found that within each BMI category, there was a significantly lower mortality risk with a higher fitness level. Therefore, while we observed a consistently significant lower risk for total mortality with higher BMI among patients who completed <6 METs, a higher fitness level was associated with an even lower relative risk for total mortality regardless of BMI.

There are a number of reasons that may explain why an inconsistent observation of the obesity paradox has been reported across different studies. For instance, in a study of >3 million people who were linked to the U.K. national mortality database, Bhaskaran et al. (20) demonstrated a higher mortality risk with higher BMI using a reference BMI of 25 kg/m². However, the median age of individuals in this study was 37 years old, and individuals who were overweight or obese had a median age that was 10 years older than normal-weight individuals. In another study that pooled data from 10 prospective U.S. cohorts, Khan et al. (21) found that obese individuals had a higher risk of CVD morbidity and mortality. However, for men, the cumulative incidence of non-CVD death was slightly lower for obese (20.1) and morbidly obese men (19.1) compared with men with a normal BMI (22.2), and in the overall pooled cohort, overweight individuals had a similar risk of mortality compared with normal-weight individuals.

Our overall results demonstrating an obesity paradox among patients with diabetes are consistent with other cohorts of patients with diabetes. However, these cohorts demonstrated significant interactions that we did not observe in the FIT Project cohort. In a study of 23,842 individuals with type 2 diabetes from the U.K. Biobank cohort, Jenkins et al. (19) demonstrated an obesity paradox. However, they found a significant smoking interaction that nullified the obesity paradox among never smokers, which they attributed to reverse causality and confounding. In our subgroup analysis, we demonstrate that our results are similar for current smokers and nonsmokers. Potential reasons for these different findings include that Jenkins et al. (19) did not exclude individuals with a BMI <18.5 kg/m², among whom there was the highest percentage of current smokers and patients with a baseline diagnosis of cancer. In an analysis of a Swedish national cohort that included ∼90% of all patients with diabetes, Edqvist et al. (22) demonstrated that the short-term (<5 year) mortality risk was lower for obese patients with a BMI of 30–35 kg/m², while long-term follow-up (≥5 years) showed a higher mortality risk with higher BMI. We did not observe a difference in the relationship between BMI and mortality for older versus younger individuals with diabetes, which may be attributable to the Swedish cohort only including patients with new-onset diabetes, among whom the mean duration of diabetes was 1.5 years and the mean age of onset was 58 years.

However, our results are consistent with previous data examining the impact of fitness on the obesity paradox among patients in the FIT Project cohort. For example, in an analysis of patients with heart failure by McAuley et al. (13), our group demonstrated a lower mortality risk with higher BMI among individuals who achieved <4 METs during a treadmill stress test, but not among those with a higher fitness level. In a separate analysis of individuals from the FIT Project without CVD or diabetes, we also demonstrated that obese individuals had a lower mortality rate compared with nonobese individuals only among the lower fitness group (23).

One limitation of this study is that we do not have information on each individual’s maximal lifetime BMI or change in BMI, which some have proposed may in part explain the obesity paradox (24). We also do not have information on duration of diabetes, and we have only incomplete information or glucose control, although the median hemoglobin A_1c was only 0.3% greater for obese versus normal-weight individuals. Obese patients referred for treadmill stress testing are likely to be healthier compared with obese patients deemed unsuitable for treadmill stress testing, which may have resulted in a higher mortality risk. While we adjusted for age in our Cox proportional hazards models, patients with a BMI ≥30 kg/m² were ∼4 years younger than patients with a BMI between 18.5 and 25 kg/m², although this age difference is less than observed in other studies (20). We also do not have measurements of regional adiposity, such as waist-to-hip ratio or imaging-based anthropometry. While BMI does not incorporate measures of fat distribution such as central adiposity, it is the most commonly used measure of obesity/adiposity in clinical practice, closely aligns with DEXA measurement of adiposity, and is an excellent predictor of excess adiposity, especially when BMI ≥30 kg/m² (25,26). In addition, we do not have information on the percentage of patients with type 1 and type 2 diabetes. However, only 21.9% of patients included in this study reported using insulin, and among U.S. adults with diabetes, the prevalence of type 1 diabetes is 5.8% (27). Finally, the observational design of this study cannot determine a cause-and-effect relationship.

Strengths of this study include 1) a nearly equal proportion of men and women; 2) a large proportion of African American patients; 3) a long follow-up period; and 4) a large number of primary outcome events (i.e., deaths). In addition, fitness was assessed using a Bruce protocol treadmill stress test, which is the most common protocol used in clinical practice.

Our results demonstrate that a higher BMI at baseline testing was associated with a lower mortality rate among patients with diabetes predominantly when fitness was low (<6 METs) or moderate (6–9.9 METs). We found no evidence of a consistently significant relationship between BMI and mortality among patients with a high fitness level (≥10 METs). Importantly, a higher fitness level was associated with a consistent and significantly lower mortality rate regardless of BMI. Accordingly, fitness significantly affects the association of BMI and mortality risk among patients with diabetes. While weight loss and improved fitness should both be recommended for patients with diabetes, these observational results highlight the need for further research to test whether prevention strategies focusing on improving fitness may potentially provide a greater reduction in mortality than weight-loss interventions.

Funding. S.P.W. was supported by the PJ Schafer Memorial Foundation.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. S.P.W. wrote the manuscript and researched data. P.A.M. wrote the manuscript and researched data. Z.D. performed statistical analyses and reviewed and edited the manuscript. O.A.O. reviewed and edited the manuscript. C.A.B. reviewed and edited the manuscript. J.K.E. reviewed and edited the manuscript. S.J.K. reviewed and edited the manuscript. M.A.-M. reviewed and edited the manuscript. M.J.B. wrote the manuscript and researched data. S.P.W. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. This study was presented in poster form at the American College of Cardiology’s 68th Annual Scientific Session & Expo, New Orleans, LA, 16–18 March 2019.