Impact of different training modalities on glycaemic control and blood lipids in patients with type 2 diabetes

META-ANALYSIS
Impact of different training modalities on glycaemic control
and blood lipids in patients with type 2 diabetes: a systematic
review and network meta-analysis
Lukas Schwingshackl & Benjamin Missbach & Sofia Dias &
Jürgen König & Georg Hoffmann
Received: 13 March 2014 / Accepted: 2 June 2014 / Published online: 5 July 2014
# Springer-Verlag Berlin Heidelberg 2014
Abstract
Aims/hypothesis This study aimed to systematically review
randomised controlled trials comparing the effects of aerobic
exercise training (AET), resistance training (RT) and com-
bined training (CT) on glycaemic control and blood lipids in
patients with type 2 diabetes mellitus.
Methods Searches were performed in MEDLINE, EMBASE
and the Cochrane Library. Inclusion criteria were: type 2
diabetes mellitus, adult, supervised training and a minimum
intervention period of 8 weeks. Pooled effects were calculated
by fixed/random effect pairwise and Bayesian fixed/random
effects network meta-analyses.
Results A total of 14 trials enrolling 915 participants were
included. AET was more effective than RT in improving
H b A 1 c l e v e l s ( m e a n d i ff e r e n c e [ M D ] − 0 . 2 0 %
[−2.2 mmol/mol]; 95% CI −0.32, −0.08; p = 0.0007,
1 0 t r i a l s / 5 1 5 p artic ip ants) a nd fastin g g lu cose
(MD −0.9 mmol/l; 95% CI −1.71, −0.09; p=0.03, 8 trials/245
participants). Compared with AET, CT resulted in a significantly
more pronounced reduction in HbA 1c (MD −0.17%
[−1.87 mmol/mol]; 95% CI −0.31, −0.03; p=0.02, 9 trials/493
participants). Compared with RT, the MD of the change in
HbA 1c (MD −0.62%, [−6.82 mmol/mol]; 95% CI −0.95,
−0.30; p=0.0002, 5 trials/362 participants], fasting glucose
(MD −1.99 mmol/l; 95% CI −3.07, −0.90; p=0.0003, 3 trials/
99 participants) and triacylglycerols (MD −0.28 mmol/l; 95%
CI −0.46, −0.10; p=0.003, 4 trials/213 participants) were all in
favour of CT. The exclusion of trials with a high risk of bias
yielded only non-significant results.
Conclusions/interpretation The present data suggest that CT
might be the most efficacious exercise modality to improve
glycaemic control and blood lipids. Interpretation with respect
to clinical relevance is limited by the low quality of the studies
included and the limited information on the clinically impor-
tant outcomes or adverse effects of exercise.
Keywords Aerobic exercise . Combined training . Network
meta-analysis . Resistance training . Systematic review
Abbreviations
AET Aerobic exercise training
BW Body weight
CT Combined training
DBP Diastolic blood pressure
FG Fasting glucose
MD Mean difference
RT Resistance training
SBP Systolic blood pressure
TC Total cholesterol
TG Triacylglycerols
Introduction
Increased physical activity and improved nutritional habits in
the form of hypocaloric diets (of varying macronutrient com-
positions) are of particular importance to decelerate the man-
ifestations of type 2 diabetes [1–3]. The ADA and the
Electronic supplementary material The online version of this article
(doi:10.1007/s00125-014-3303-z) contains peer-reviewed but unedited
supplementary material, which is available to authorised users.
L. Schwingshackl (*) : B. Missbach : J. König : G. Hoffmann
Department of Nutritional Sciences, Faculty of Life Sciences,
University of Vienna, Althanstraße 14 (UZAII), 1090 Vienna,
Austria
e-mail: [e-mail]
S. Dias
School of Social and Community Medicine, University of Bristol,
Bristol, UK
Diabetologia (2014) 57:1789–1797
DOI 10.1007/s00125-014-3303-z

American College of Sports Medicine have stated that a
combination of resistance training (RT) and aerobic exercise
training (AET) of at least 150 min of moderate-intensity
exercise per week may be more effective in improving
glycaemic control than focusing solely on one single training
modality (evidence category B) [4].
The isolated effects of either RT or AET, or a com-
bination of both (combined training [CT]), on anthropo-
metric, cardiac and metabolic risk factors have been
meta-analysed by Snowling and Hopkins [5] as well as
by Chudyk and Petrella [6]. Both studies reported that
the reduction in HbA 1c and fasting glucose (FG) levels,
systolic blood pressure (SBP), waist circumference,
HDL and triacylglycerols (TG) was more pronounced
following AET and CT compared with RT. In addition,
HbA 1c - and blood pressure-lowering effects of RT were
shown. However, all these systematic reviews included
trials in which training modalities were compared with
the data from a sedentary control group [7, 8].
To date, no systematic review has compared the
direct and indirect effects of these three different train-
ing modalities on the outcomes of glycaemic control
and blood lipids in patients with type 2 diabetes. A
recent pairwise meta-analysis comparing RT (all super-
vised) with AET (not all supervised) exercise in patients
with type 2 diabetes concluded that although differences
in some outcome variables reached statistical signifi-
cance, there was no evidence that they were of clinical
relevance [9]. In a recently published network meta-
analysis, we were able to demonstrate that CT is ranked
as the most likely effective exercise model in the treat-
ment of overweight and obesity [10].
The aim of the present study was to assess the efficacy of
AET, RT and CT on glycaemic control, blood pressure and
blood lipids in patients with type 2 diabetes mellitus in a
systematic review including a pairwise and network meta-
analysis of randomised trials. The importance of supervised
training has been demonstrated by Umpierre et al, who
showed that structured training compared with training advice
only significantly improved glycaemic control in patients with
type 2 diabetes [11]. Based upon these findings, only trials that
conducted an exercise intervention which was guided by
supervised training were enrolled in the present systematic
review and meta-analyses.
Methods
The review was registered in PROSPERO International
Prospective Register of Systematic Reviews (www.crd.york.
ac.uk/prospero/index.asp, identifier CRD42014007502).
However, no study protocol was published before the
initiation of the meta-analysis.
Literature search
Queries of the literature were performed using the electronic
databases MEDLINE (until 2 May 2014), EMBASE (until 2
May 2014) and the Cochrane Central Register of Controlled
Trials (until 2 May 2014) with no restrictions. The following
keywords were used: (‘strength’ OR ‘resistance’ OR ‘aerobic’
OR ‘endurance’ OR ‘combined training’ OR ‘progressive’
OR ‘walking’ OR ‘interval training’ OR ‘weight lifting’)
AND (‘training’ OR ‘exercise’ OR ‘physical activity’) AND
(‘diabetes’ OR ‘glycemic’ OR ‘glycaemia’ OR ‘glycaemic’
OR ‘glycemia’ OR ‘HbA1c’ OR ‘A1c’ OR ‘glycated’ OR
‘glycosylated’ OR ‘glucose’ OR ‘lipids’ OR ‘body weight’
OR ‘blood pressure’) AND (‘randomized controlled trial’ OR
‘randomized’ OR ‘clinical trials as topic’ OR ‘placebo’ OR
‘randomly’ OR ‘trial’) NOT (‘animals’ NOT ‘humans’).
Moreover, the reference lists from the retrieved articles,
systematic reviews and meta-analyses were checked to search
for further relevant studies. This systematic review was planned,
conducted and reported in adherence with standards of quality
for reporting meta-analyses [12]. The entire literature search was
conducted independently by two authors (L. Schwingshackl and
B. Missbach), with disagreements resolved by consensus. The
detailed search strategy for MEDLINE is given in the electronic
supplementary material (ESM Methods).
Eligibility criteria
Studies were included in the meta-analysis if they met all of
the following criteria: (1) a randomised controlled design; (2)
a minimum intervention period of 8 weeks; (3) patients with
type 2 diabetes; (4) patients’ age ≥19 years; (5) a comparison
of either AET vs RT and/or CT vs AET and/or CT vs RT; (6)
an assessment of at least one of the following outcome
markers: HbA1c, blood glucose, body weight (BW), blood
pressure or blood lipids (total cholesterol [TC], LDL, HDL
and TG); (7) the reporting of changes from baseline value
scores with SDs (or data suitable to calculate these variables:
SE and 95% CI); if the SDs of the changes from baseline value
scores were not available, post-intervention values were im-
puted, according to the Cochrane Handbook [13]; (8) training
that was conducted under direct (guided by a physiotherapist
in training classes, hospital gyms, etc.) or partial supervision,
and was not home-based; and (9) the exclusion of studies with
a dietary co-intervention that was not applied in all the inter-
vention groups. All abstracts and full texts were independently
assessed for eligibility by two authors (L. Schwingshackl and
B. Missbach).
Risk of bias assessment
Full copies of the studies were independently assessed by two
authors (L. Schwingshackl and B. Missbach) for
1790 Diabetologia (2014) 57:1789–1797

methodological quality using the risk of bias assessment tool
from the Cochrane Collaboration [13, 14]. The following
sources of bias were detected: selection bias (random sequence
generation and allocation concealment), detection bias (blinding
of outcome assessment), blinding of participants and personnel
(performance bias), attrition bias (incomplete outcome data) and
reporting bias (selective reporting) (ESM Fig. 1).
Data extraction and statistical analysis
The following data were extracted from each study: the first
author’s last name, publication year, study duration, partici-
pant’s sex, age and BMI, sample size, duration of diabetes,
HbA1c at baseline, drug treatment, change of treatment during
the trial, treatment effects, intervention type, dose, intensity and
frequency, and differences in the means of two time points or
post-intervention mean values with corresponding SDs. For
each outcome measure of interest, pairwise and network random
effects meta-analyses were performed in order to determine the
pooled relative effect of each intervention relative to every other
intervention in terms of the mean differences (MDs) between
the changes from baseline value scores (or post-intervention
values) of the different interventions. To process the data for
the meta-analysis, we imputed the data for the changes from
baseline means and their SDs. When the SDs for the changes
from baseline values were not available [15–20], the post-
intervention values with the corresponding SDs were imputed,
according the guidelines of the Cochrane Handbook [13].
Data were pooled if outcomes were reported by at least
three studies. Heterogeneity between trial results was tested
with a Cochran’s Q test. A value for I2 of >50% was consid-
ered to represent substantial heterogeneity [21]. When sub-
stantial heterogeneity was present, the random effects model
was used to estimate MDs with 95% CIs. Forest plots were
generated to illustrate the study-specific effect sizes along with
a 95% CI. To determine the presence of publication bias, the
symmetry of the funnel plots in which mean MDs were
plotted against their corresponding SEs were assessed.
Additionally, Begg’s and Egger’s regression tests were per-
formed to detect small study effects [22, 23].
Separate pairwise meta-analyses were first used to compare
all the interventions. Network meta-analysis was then used to
synthesise all the available evidence [24]. Network meta-
analysis methods are extensions of the standard pairwise
meta-analysis model that enable a simultaneous comparison
of multiple interventions while preserving the internal
randomisation of individual trials. They have the advantage
of adequately accounting for the correlation in relative effect
estimates from three-arm trials as well as providing a single
coherent summary of all the evidence. Random effects net-
work meta-analysis models were used when substantial het-
erogeneity was found in any of the pairwise comparisons for
that outcome. Otherwise, the choice between fixed and
random effects was made by comparing the deviance infor-
mation criteria for each model [24, 25]. The model with the
lowest deviance information criterion was chosen (differences
>3 are considered meaningful). Pooled effect sizes from the
network meta-analyses are presented as posterior medians and
95% credible intervals (i.e. the Bayesian equivalent of CIs) in
the appropriate units, along with the estimated between-study
heterogeneity.
For pairwise meta-analyses, data were analysed using
Review Manager 5.1 software, provided by the Cochrane
Collaboration (http://ims.Cochrane.org/revman). Network
meta-analyses were conducted using Markov chain Monte
Carlo simulation implemented with the open-source software
WinBUGS, version 1.4.3 [26]. The WinBUGS code used is
freely available online [24, 27] (program ‘TSD2-5aRE_
Normal_id.odc’ or ‘TSD2-5aFE_Normal_id.odc’).
Minimally informative normal priors were used for all
treatment effect variables and a uniform prior (0, 150) was
used for the between-study SD (heterogeneity) variable.
Sensitivity to this prior was assessed, but there was no mean-
ingful change in the relative effects or overall conclusions.
Three Markov chain Monte Carlo chains were used to
assess convergence using Brooks–Gelman–Rubin plots and
inspection of the trace plots [28]. Convergence was achieved
after 20,000 iterations for all outcomes. Posterior summaries
were then obtained from a further simulation of 50,000 itera-
tions in each of the three chains (giving 150,000 in total),
resulting in a small Monte Carlo error.
The potential for inconsistency was assessed by inspection
of the available evidence. In case of possible inconsistency,
Bayesian p values for the difference between direct and indi-
rect evidence were calculated, and direct and indirect esti-
mates were compared [29, 30].
Results
Overall, a total of 14 trials (16 reports) extracted from 9,477
articles met the eligibility requirements and were included for
the present systematic review and meta-analysis [15–20,
31–40]. One study was excluded since it was not described
as randomised [41], and two trials provided no information on
whether the AET was supervised [42, 43]. The detailed steps
of the article selection process for the meta-analysis are de-
scribed as a flow diagram in ESM Fig. 2. The studies were
published between 2003 and 2013 and had enrolled a total of
915 participants. The study duration ranged between 2 and
12 months; the patients’ mean age was between 49 and
62.5 years, and their BMI between 27.1 and 43.8 kg/m2 .
Fourteen trials met the objectives for meta-analysis: 10 com-
pared RT vs AET, 9 compared CT vs AET, and 5 compared
CT vs RT (ESM Fig. 3). The general and specific study
Diabetologia (2014) 57:1789–1797 1791