The search yielded 2477 records. Two records were identified through hand-searching the reference lists. Six hundred and nineteen duplicates (619) were removed and the remaining 1860 abstracts were assessed for eligibility (Fig. 1). One thousand seven hundred and ninety four (1794) abstracts were excluded during the primary exclusion round and a further 52 were excluded after assessing the full text articles. The reasons for exclusion of the full articles are shown in the appendix (p 12). Fourteen studies24–37 were included in the qualitative synthesis and thirteen of them in the meta-analysis.24–31,33−37 Maindal et al.32 was not possible to be included in the meta-analysis due to missing values. The author responded to the initial inquiry to provide the requested data but did not follow through. Two reminder email requests were sent to no avail. Three of these studies24,26,28 have been cited in previous nutrition guidelines for diabetes management.38
All but one29 trial were conducted in outpatient settings. Thirteen trials involved a parallel RCT design24–29,31−37 and one trial30 cluster randomization with each cluster representing a community. In total, 3338 subjects with type 2 diabetes, representing a diverse range of participants from eighteen countries from the six inhabited continents, were included (appendix p 20 and 27). Only the “diet” group that was randomly assigned to nutrition therapy (n = 248), and not the “diet plus activity” group (n = 246), was included from the study by Andrews et al.24
Table 1 summarizes the study characteristics. Supplementary tables 1, 2 and 3 (appendix pp 22–30) list details of each study and a summary of participants’ demographics and clinical characteristics. The mean age of participants was 55·8 years. The majority of participants were of white European ethnic background (40%), with one quarter Asian and 12% black African. Approximately half of the participants were males (49%) and half females (51%). One-third of the participants (35%) had completed less than high-school education and half (53%) were employed. Participants were on average diagnosed with type 2 diabetes for six years and 10 weeks and about one in ten was a self-reported smoker. The mean BMI at baseline was 31·5 kg/m2 and the mean HbA1c was 7·89%. The majority of the participants were on oral hypoglycaemic agents (59%), antihypertensives (68%) and lipid modifying medication (51%). No significant differences at baseline between the pooled intervention versus the control group were observed.
For the individual studies, there were few significant differences in baseline characteristics between groups. In the study by Huang et al.,28 patients in the control group had higher BMI and diastolic BP than those of the intervention group at baseline. In two studies,31,37 the intervention group had at least twice as many smokers as the control group. In the study by Eakin et al.27 the control group was of a lower education status. Baseline BMI was much lower in the four Asian studies28,30,36,37 than in the eight studies of non-Asians.24–27, 29,31–35 HbA1c was just above normal at baseline in one study,32 close to normal in another one,24 just < 8% in five studies,27,30,31,34,35 and > 8% in the other seven studies.25,26,28,29,33,36,37
Interventions lasted from four months to three years. In most studies (79%), the participants consulted with the dietitians at least six times. A few studies reported on the duration of each session, but in most this was not clear. Sessions were either face-to-face, telephone, individual, group or a combination of some of the aforementioned. The reported dropout rate varied considerably, from 2% in the study by Andrews et al.,24 to 20% in the study by Huang et al.,28 but with no significant difference between groups. Shahid et al.36 did not report attrition.
Risk of bias was assessed at the study level using the Cochrane Risk of Bias tool.18 In all fourteen studies, subjects were randomly assigned before the intervention. The random-sequence–generation method was stated in all the studies but two.30,36 In the study by Liu et al.,30 two Chinese communities were randomly assigned to either the intervention or a control. This assignment involves high risks of recruitment and selection bias. The study by Shahid et al.36 reported that patients were “randomly distributed” but there was no reference to a methodology. Allocation was not concealed, or it was unclear, in five studies.28–30, 33,36 In one study doctors were unaware of allocation of participants but dietitians and nurses were.24 The blinding of participants and personnel (performance bias) was not possible in the present studies. The blinding of outcome assessors was reported in four studies.24,27,31,35 Handling of missing endpoints (e.g. intention-to-treat, multiple imputation) was not clear in some studies. 28,29,30,37 The reporting in two trials was not consistent with their registration information28,34,37 and two trials were not registered.30,33 Three studies did not report conflict of interest.31,35,36 In one study that reported industry funding, authors stated affiliations with the funding body.34 No other potential sources of bias were identified from the studies (Fig. 2). Conflicts of interest and funding sources are listed in the appendix (p 31). The risk of bias ratings meant the overall rating of the body of evidence was downgraded despite all being RCTs.
All trials reported changes in HbA1c, with nine reporting it as the primary outcome and three27,32,34 as a secondary outcome. Ten-year modelled cardiovascular risk was the primary outcome for Maindal et al.27 and changes in body weight for the other two.32,34 Ten studies reported changes in BMI and weight, twelve reported on blood pressure and eight on low-density lipoprotein (LDL) cholesterol. Less than half of the studies reported changes in energy and nutrient intake, medication use, waist circumference and physical activity. We pooled results for HbA1c, BMI, weight, LDL cholesterol and systolic and diastolic blood pressure (BP) in meta-analyses.
Table 2 shows the summary of findings of the meta-analysis. The GRADE scoring methodology is explained in the appendix (p 11). Figure 2 and supplementary Figs. 2–6 (appendix pp 32–36) show Forest plots of the primary and secondary outcomes respectively. Nutrition intervention delivered by dietitians achieved a -0·47 [-0·92, -0·02] percentage point greater reduction in HbA1c, -0·38 [-0·63, -0·13] kg/m2 lower BMI, -1·49 [-2·14, -0·84] kg greater reduction in weight, -0·15 [-0·33, 0·04] mmol/L lower LDL cholesterol and − 0·75 [-2·45, 0·96] mm Hg and − 1·17 [-4·52, 2·17] mm Hg greater reductions in systolic and diastolic BP respectively, compared to nutrition advice by other healthcare professionals. The reductions in HbA1c, BMI and weight were statistically significant (P = 0·04, P = 0·003 and P < 0·00001 respectively). Heterogeneity between the studies was statistically significant (Q = 985·48, P < 0·00001) and “considerable” in magnitude (I2 = 99%) for the HbA1c, BMI (Q = 53·75, P < 0·00001, I2 = 85%), weight (Q = 45·44, P < 0·00001, I2 = 80%), LDL cholesterol (Q = 206·28, P < 0·00001, I2 = 97%), systolic BP (Q = 76·45, P < 0·00001, I2 = 87%) and diastolic BP (Q = 1686·47, P < 0·00001, I2 = 99%) outcomes. The body of evidence was considered inconsistent due to the “considerable” heterogeneity and the level of evidence was downgraded for all outcomes. A further point was deducted from the evidence score due to indirectness, as some studies had short duration and all of them reported surrogate endpoints.
Pooled participants in the intervention group (n = 1334) achieved more than four times the reduction in HbA1c, i.e. -0·65% (± 0·42), versus − 0·15% (± 0·39) that was achieved by the participants in the control group (n = 1177). In addition, pooled participants in the intervention group (n = 998) reduced their weight on average by about 1·35 kilogram (± 0·75 kg), whereas participants in the control group (n = 830) gained on average 78 grams (± 910 grams) (data not shown). This weight reduction in the intervention group corresponded to approximately 2% of their baseline weight.
We conducted a series of secondary analyses excluding one of all possible combinations of two of, three of, four of, five of, six of and all seven of studies that scored poorly in the ROB assessment.28–30, 33,34,36,37 None of the different combinations changed the direction of associations, but some of them affected the statistical significance of the HbA1c, BMI and weight outcomes (data not shown). When only the six studies24–27, 31,35 that scored higher on ROB (at least five “low risk” fields in total and mandatory “low risk” assessment for randomisation process, detection bias and selective reporting) were analysed, the results remained statistically significant albeit attenuated, e.g. HbA1c was − 0·29 [-0·50, -0·07] %.
We performed another series of secondary analyses, this time including only the eight studies that followed up for at least 12 months.24,25,27,28,30,33−35 Removing the five datasets of the studies that followed up at six months or sooner26,29,31,36,37 attenuated the result for HbA1c (-0·24 [-0·41, -0·07] %) but increased it for weight (-1·62 [-2·73, -0·51] kg). None of these changes were statistically significant and the point estimate gained precision. The directions of association and measured effects were not statistically different for the LDL cholesterol and blood pressure outcomes. On the contrary, the BMI outcome crossed the line of no-effect (data not shown).
The robustness of the estimate was examined by sequentially removing each study and reanalysing the remaining datasets. The estimated effect sizes ranged from − 0·51 to -0·30% for HbA1c; -0·47 to -0·30 kg/m2 for BMI; -1·68 to -1·13 kg for weight; -0·19 to -0·07 mmol/L for LDL cholesterol; -1·50 to -0·34 mm Hg for systolic blood pressure; and − 1·51 for − 0·19 mm Hg for the diastolic blood pressure outcome. Individual removal of any of the following datasets from Coppell et al., Liu et al., Lynch et al., or from Suriyawongpaisal et al. abolished the statistical significance of the result for the HbA1c. Removal of the Shahid et al. dataset rendered the LDL cholesterol result statistically significant. No single study had an impact on the significance of the effect size for the BMI, weight, systolic and diastolic blood pressure outcomes (data not shown).
We also tested whether the fixed-effects model would produce different results to the random-effects one. The fixed-effects model rendered the LDL cholesterol and blood pressure outcomes statistically significant and maintained statistical significance in the HbA1c, weight and BMI outcomes, augmenting the effect size of HbA1c. It resulted in narrower confidence intervals, as expected.
Publication bias was assessed via visual inspection of funnel plots (appendix p 37). The data appeared skewed for the HbA1c and the diastolic BP outcomes. The asymmetry observed for the HbA1c is likely a result of the heterogeneity.39 Whereas the skewed distributions seen for the diastolic BP outcome is largely due to an excess of medium to large studies with negative effects on the outcome. The asymmetry manifested in the weight outcome is likely due to an excess of small to medium studies with positive effects on the outcome. There was no evidence of publication bias for systolic BP. Publication bias could not be assessed for the BMI and LDL cholesterol outcomes as the number of studies is not adequate to distinguish chance from real asymmetry.19