Cardiorespiratory fitness, muscular strength and risk of type 2 diabetes: a systematic review and meta-analysis

PubMed and EMBASE were searched for cohort studies on the associations among cardiorespiratory fitness, muscular strength and risk of incident type 2 diabetes (electronic supplementary material [ESM] Table 1 and ESM Table 2). No restrictions on date of publication were set. The final search was conducted on 12 December 2018. We additionally went through reference lists from studies included in the review and searched Web of Science for studies citing these publications.

Cohort studies were eligible if they (1) followed individuals free of type 2 diabetes at baseline, but we included cohorts of individuals with diabetes-associated conditions (e.g. dyslipidaemia or obesity); (2) assessed cardiorespiratory fitness using a maximal or sub-maximal test of any form at baseline or assessed muscular strength using a test requiring a maximal effort at baseline; (3) considered incidence of type 2 diabetes as an isolated outcome; and (4) were published in Scandinavian or English language. Conference abstracts were included if relevant. Restriction to muscular fitness operationalised by maximal strength (thereby excluding muscular power and endurance) [10] was chosen to maximise the potential for exposure harmonisation. Studies were excluded if they (1) considered a cohort of individuals with a chronic disease (e.g. cancer); or (2) assessed muscular endurance or power (full details in ESM Table 3). To be included in meta-analysis, studies additionally had to provide (1) HR, OR or RR for type 2 diabetes for one or more cardiorespiratory fitness or muscular strength estimates (linear or categorical); (2) estimates of variance or data to calculate it; and (3) had cardiorespiratory fitness estimates convertible to metabolic equivalents (METs) [10] or muscular strength estimates convertible to per SD. Two researchers (J. Tarp and A. P. Støle) independently screened titles and abstracts using Endnote X7.7.1 (Clarivate Analytics, PA, USA) according to pre-specified criteria. When eligibility was ambiguous, the full text was retrieved. Disagreement was resolved by discussion including a third researcher (A. Grøntved).

Data from eligible studies were independently extracted by two researchers (J. Tarp and A. P. Støle) using a piloted template. Disagreements were resolved by discussion. The following were extracted if available: first author, country, cohort name/title, cohort recruitment period (years), sex, ethnicity, baseline age, length of follow-up, method of outcome ascertainment, cumulative diabetes incidence, method of exposure ascertainment, levels of exposure, case count and total participant count in fitness categories, HRs/ORs/RRs and associated variance for linear or categorical estimates, and control variables applied in retrieved estimates. From each study or cohort we extracted estimates with and without control for an index of adiposity if available. We extracted adiposity- and non-adiposity-controlled estimates from the same report if possible, but included other reports from the same cohort if this increased sample sizes or the number of cases or facilitated a more direct BMI contrast. When a study did not provide either (1) estimates with and without adiposity control; or (2) estimates from at least two categories compared with a common reference, we contacted corresponding authors and requested additional information using a standardised template (template available on request to the corresponding author). The Newcastle–Ottawa Scale (NOS) [19] with modifications informed by the study question was used to rate overall study quality (details of the NOS rating and criteria in ESM Table 4 and ESM Table 5). Age, sex, ethnicity, cardiorespiratory/muscular fitness, smoking [20], family history of diabetes, dietary intake [21], alcohol consumption [22], TV viewing [23] and socioeconomic status were considered putative confounding variables which could potentially result in biased measures of association. When multiple publications from the same cohort were identified, we used the manuscript presenting the largest case and participant count with harmonisable exposure data (table of overlapping cohorts presented in ESM Table 6). If additional data (e.g. linear/categorical or with/without control for adiposity) from the cohort were available in other publications, we retrieved estimates from both of these papers.

A detailed overview of assumptions, calculations and unpublished data provided by contacted authors used in exposure harmonisation is provided in ESM Table 7 and ESM Table 8. If possible, cardiorespiratory fitness estimates were converted to METs for the non-linear analysis and to per 1 MET increase for the linear meta-analysis. Muscular strength was converted to per SD increase. Harmonisation of linear estimates was performed using transformation of the log-ratio estimate (using the natural logarithm) [24] under the assumption of fitness measures following a normal distribution and a log-linear association with type 2 diabetes incidence. Estimates for cardiorespiratory fitness based on treadmill duration were converted using exercise protocol-appropriate equations [25]. When the exposure level was unclear, but distributional assumptions allowed estimation, we assumed an SD of cardiorespiratory fitness of 2.0 METs [26, 27] in subsequent calculations. Cardiorespiratory fitness presented as watts per kg was converted to ml O₂ kg⁻¹ min⁻¹ using a linear equation based on measurement of maximal oxygen uptake by indirect calorimetry (variance explained was 71% [ESM Table 7]). Data provided as ml O₂ kg⁻¹ min⁻¹ were converted to METs by dividing by 3.5 [10]. When original estimates did not have the lowest fitness category as reference, we converted the lowest fitness level to the reference using the Hamling method [28].

Harmonised estimates for cardiorespiratory and muscle strength were pooled using a random-effects model [29] under the assumption of a linear dose–response relationship and that each study provides an estimate in a distribution of ‘true’ estimates. We provide fixed-effects estimates for comparison. If studies provided HRs/RRs/ORs pertinent to a continuous form we used this estimate. Otherwise, and if at least two categories with a common reference were available, we used generalised least squares trend (GLST) estimation to estimate the study-specific dose–response association [30] taking into account the common reference group [31] and the non-zero exposure in the reference group by centring the exposure corresponding to the natural log RR [32]. In addition, we modelled the dose–response association between cardiorespiratory fitness and type 2 diabetes using restricted cubic splines with knots placed at the 25th, 50th and 75th percentiles [33] of the cardiorespiratory fitness distribution in the data. Departure from linearity was assessed by a Wald test examining the null hypothesis that the coefficient of the second spline was equal to zero [34]. At least two categories with a common reference were needed to be included in non-linear analysis. Insufficient data precluded non-linear analysis of muscular strength. Analyses were performed separately on estimates reported with and without inclusion of adiposity indices as control variables. Adiposity-adjusted estimates were included as primary analysis as we consider this the more conservative analysis. We present data for men and women separately if available but used a fixed-effects meta-analysis to pool estimates within a study based on other stratifications if relevant. ORs were assumed to approximate RRs [35]. Between-study heterogeneity was formally assessed using I² as a measure of the proportion of variance not explained by random error and by visual interpretation of the forest plots. Sources of heterogeneity were explored by stratification on cohort and participant characteristics. Robustness of estimates was assessed by repeating the analysis excluding a single study at a time. Risk of small-study bias was investigated by funnel plot and Egger’s test for asymmetry. Estimates are presented with a 95% CI. Assuming that estimates represent causal effects, we calculated the risk difference for a 1 MET and a 1 SD increase in cardiorespiratory fitness and muscular strength, respectively. The risk differences (with 95% CIs) were calculated using the formula: risk difference = background incidence rate × (RR − 1) [36]. We calculated risk differences based on background annual rates in the age strata 18–44, 45–65 and ≥ 65 years based on 2015 US incidence data [37]. We calculated potential impact fractions (PIFs) [38, 39] as a measure of the percentage of new annual diabetes cases in the population that could hypothetically be prevented by interventions affecting the population distribution of cardiorespiratory fitness. Our hypothetical interventions were modelled on sex-specific population estimates of cardiorespiratory fitness for 40–59-year-olds from the US Fitness Registry and the Importance of Exercise National Database (FRIEND) [40]. PIFs were calculated under four counterfactual scenarios: (1) a structural intervention resulting in a 1 MET increase in the bottom 50% of the cardiorespiratory fitness distribution; (2) the same intervention but resulting in a population-wide 1 MET fitness increase; (3) achievement of the same cardiorespiratory fitness distribution as observed in age-matched individuals from the Norwegian population-based HUNT study [41]; and (4) achievement of cardiorespiratory fitness distribution identical to the most physically active tertile of age-matched individuals from the HUNT study [41]. Additional details are given in ESM Table 9. We were unable to calculate PIFs for muscular strength as no reference distribution was identified. All p values were two-sided and interpreted at the 0.05 level. Analyses were performed in Stata 15.0 (StataCorp, College Station, TX, USA).

In total, 22 studies (representing 18 unique cohorts) on cardiorespiratory fitness [13, 14, 18, 42‐60] and 13 studies on muscular strength [15, 16, 18, 43, 61‐69] were identified for inclusion in the systematic review. The phases of the literature search are shown in Fig. 1. Additional data were retrieved by personal communication to six cohorts for cardiorespiratory fitness [18, 45, 48‐50, 52] and five for muscular strength [16, 18, 63, 65, 68].

Cohort studies reporting estimates for cardiorespiratory fitness ranged in size from 571 to 1,534,425 participants with a median cohort baseline age of 45 years. The cumulative incidence of type 2 diabetes ranged from 0.7% to 26.2% with median follow-up ranging from 3 to 29 years. Nine cohorts used a maximal test while eight used a sub-maximal assessment of fitness. For muscular strength, cohort sizes ranged from 328 to 1,534,425 participants. The median cohort age at baseline was 52 years. The cumulative incidence of type 2 diabetes ranged from 2.1% to 18.5% with median duration of follow-up ranging from 3 to 26 years. Eight studies reported muscle strength normalised to kilogram body weight (two of these after receiving additional data from study authors) while five did not normalise muscular strength. Eleven studies used maximal handgrip strength to define muscular strength while two used a composite index including multiple muscle groups. The median SD of muscular strength as a percentage of the mean was 31% (range: 22–39%). Of the 31 studies considered, only four used a measure other than BMI to control for adiposity. The median NOS score was six for cardiorespiratory fitness and six for muscular strength. Importantly, all studies failed to account for at least four of the seven pre-specified confounding factors with potentially non-trivial impact on the internal validity of risk estimates. An overview of study characteristics is presented in ESM Table 10 and ESM Table 11. Seven studies for cardiorespiratory fitness [54‐60] and two studies for muscular strength [67, 69] presented data that were not harmonisable for inclusion in meta-analysis. Results from these studies generally supported a protective effect of higher cardiorespiratory fitness and muscular strength on risk of type 2 diabetes.

In adiposity-controlled models including 40,286 incident cases of type 2 diabetes in 1,601,490 participants, each 1 MET higher cardiorespiratory fitness was associated with an 8% (95% CI 6%, 10%) reduction in risk of type 2 diabetes [13, 14, 18, 44, 45, 48‐50, 52, 53]. The per 1 MET risk reduction in models omitting adiposity control was 20% (95% CI 14%, 25%) [14, 43‐48, 50‐52]. Study-specific per 1 MET estimates are shown as forests plots with (Fig. 2) and without (ESM Fig. 1) adiposity control.

The non-adiposity-controlled meta-analysis was highly sensitive to the choice of random- or fixed-effects modelling; estimated heterogeneity was substantial in both models (I² 83% and 93%, respectively), whereas visual inspection of the forest plots suggested moderate heterogeneity. Heterogeneity appeared to be explained by the influence of two studies, one with a noticeably narrow 95% CI [18] and one with a substantial effect size and a moderate size CI [45]. Excluding these studies reduced the I² to 48%. The association between adiposity-controlled cardiorespiratory fitness and type 2 diabetes risk modelled using restricted cubic splines [14, 18, 42, 44, 45, 48‐50, 52, 53] is shown in Fig. 3. The model was consistent with a continued linear risk reduction with no statistical support of a non-linear association (p = 0.07) within our data ranging from four to 15 METs. A restricted cubic spline model including non-adiposity-controlled estimates [14, 44, 45, 48, 50‐52] is presented in ESM Fig. 2.

The risk differences associated with a 1 MET higher cardiorespiratory fitness were 87 (95% CI 64, 112) incident diabetes cases per 100,000 people per year in the US population aged 45–64 years for the adiposity-controlled RR and 218 (95% CI 150, 280) for the non-adiposity-controlled RR (risk difference for other age strata available in ESM Table 12). The PIFs (adiposity-controlled) associated with a hypothetical intervention resulting in a 1 MET increase among the least fit 50% of US men and women aged 40–59 years were 4.2% and 3.6%, respectively. PIFs (adiposity-controlled) of achieving the same cardiorespiratory fitness distribution as is observed in a population-based sample of same-aged Norwegians were 18.0% and 20.6% for men and women, respectively (Table 1 [70, 71]); PIFs based on the non-adiposity-controlled estimate are presented in ESM Table 9. The impact of stratification on the pooled RR (adiposity-controlled) is shown in Table 2. Apart from a marked difference in sex-stratified results (only one study included women only), none of our stratification parameters substantially altered the risk reduction.

In adiposity-controlled models including 39,233 incident cases and 1,713,468 participants, each SD higher muscular strength was associated with a 13% (95% CI 6, 19) lower risk of type 2 diabetes [15, 16, 18, 61‐66, 68] (Fig. 4). Pooling the nine available estimates not controlled for adiposity yielded an RR of 0.76 (95% CI 0.64, 0.91) [15, 16, 43, 63‐65, 68] per SD higher muscular strength (ESM Fig. 3). Heterogeneity was substantial in both models (I² 81% and 91%, respectively). The risk difference associated with a 1 SD higher muscular strength (adiposity-controlled RR) was 142 (95% CI 43, 211) new diabetes cases per 100,000 people per year in the US population aged 45–64 years (additional information in ESM Table 12). Studies applying normalisation of muscular strength to body weight yielded, on average, larger effect sizes (0.83 [95% CI 0.79, 0.86]) than those relying on absolute strength (0.95 [95% CI 0.87, 1.04]).

Pooled adiposity-controlled estimates were robust to removal of single estimates with RRs ranging from 0.91 to 0.93 for cardiorespiratory fitness and from 0.86 to 0.88 for muscular strength (presented in ESM Tables 13–16). Egger’s test did not indicate a risk of small-study bias in adiposity-controlled estimates (p = 0.75 and 0.53) or non-adiposity-controlled estimates for muscular fitness (p = 0.83). Some evidence was found in the non-adiposity-controlled cardiorespiratory fitness estimates (p = 0.002). The latter appeared to be explained by the influence of one study which had by far the smallest SE and also reported the largest effect size [14]. Funnel plots are presented in ESM Figs. 4–7.