Dietary Glycemic Index and Load and the Risk of Type 2 Diabetes: A Systematic Review and Updated Meta-Analyses of Prospective Cohort Studies

Published meta-analyses indicate significant but inconsistent incident type-2 diabetes (T2D)-dietary glycemic index (GI) and glycemic load (GL) risk ratios or risk relations (RR). It is now over a decade ago that a published meta-analysis used a predefined standard to identify valid studies. Considering valid studies only, and using random effects dose–response meta-analysis (DRM) while withdrawing spurious results (p < 0.05), we ascertained whether these relations would support nutrition guidance, specifically for an RR > 1.20 with a lower 95% confidence limit >1.10 across typical intakes (approximately 10th to 90th percentiles of population intakes). The combined T2D–GI RR was 1.27 (1.15–1.40) (p < 0.001, n = 10 studies) per 10 units GI, while that for the T2D–GL RR was 1.26 (1.15–1.37) (p < 0.001, n = 15) per 80 g/d GL in a 2000 kcal (8400 kJ) diet. The corresponding global DRM using restricted cubic splines were 1.87 (1.56–2.25) (p < 0.001, n = 10) and 1.89 (1.66–2.16) (p < 0.001, n = 15) from 47.6 to 76.1 units GI and 73 to 257 g/d GL in a 2000 kcal diet, respectively. In conclusion, among adults initially in good health, diets higher in GI or GL were robustly associated with incident T2D. Together with mechanistic and other data, this supports that consideration should be given to these dietary risk factors in nutrition advice. Concerning the public health relevance at the global level, our evidence indicates that GI and GL are substantial food markers predicting the development of T2D worldwide, for persons of European ancestry and of East Asian ancestry.


The Literature Search Strategy
The search strategy below for MEDLINE and EMBASE simultaneously was developed with expertise from LitSearch at the Royal Society of Medicine (RSM), London, UK, and was an update on prior searches undertaken. It was last run on ProQuest accessed via the RSM website on the 6 th December 2018..
i. One report provided no quantitative exposure data for GL: Oba et al [25] j. Not reporting on their fully adjusted model for the T2D-GL relations:  Bhupathiraju et al HPFS, NHS I and NHS II [19].
k. Not having the longest duration of follow-up for the T2D-GL relation because:  Salmeron for women (NHS I) at 6 y follow-up 199 [4] , the longer study remaining included being the NHS I at 26 y in Mekary et al [23] which was pre-combined with that of Halton et [15] at 20 y follow-up because the individual results before pre-combining were inconsistent with one another (I 2 =95%).
l. No validation results complete for CORR  Provided no information on the validity (CORR) of their dietary instrument for carbohydrate Rossi et al [26]  Incomplete reportin CORR for only 4 of 15 regional and sex specific cohorts in a multiple regional study [22]. m. Reports on the same model with inconsistent results  halton and Mekary combined To these was added one from two different reports that addressed the same study but had inconsistent dose-response results (I 2 =95%, n=2), [15,23] which pre-combined (Sections 2.1.4).

Newcastle-Ottawa score of study quality (NOS) as used in the present study
While generally accepted that individual study quality should be assessed and reported when conducting systematic reviews, no method has been validated for non-randomized studies such as prospective cohort studies. The value of study quantity assessment remains for the present primarily in providing a measure to which a study has been conducted and reported according to generally recognized practices for studies deemed of high quality. Individual quality items and groups of quality items are generally recognized as potential determinants of a successful study and may correlate with study outcomes, but this should not be expected automatically and there is increasing recognition that study quality score should not be used as if a determinant of a study outcome.
The following reproduces the protocol as encountered [27] with insert in bold italics to adapt it to the present study.
Note: A study can be awarded a maximum of one star (point) for each numbered item within the Selection and Outcome categories. A maximum of two stars (points) can be given for Comparability. This was reduced by one star for studies with invalid dietary instruments (those with an instrument correlation coefficient ≤0.55 for dietary carbohydrate with food records. Selection for healthy persons representative of a community aiming for national (and eventually global ) representation. 1) Representativeness of the exposed cohort a) truly representative of the average __ adult mixed gender or male or female __ in the community ? * b) somewhat representative of the average __ adult mixed gender or male or female_ in the community ?* For example not full age range of the community for which type-2 diabetes is incident. c) selected group of users eg nurses, volunteers d) no description of the derivation of the cohort 2) Selection of the non exposed cohort a) drawn from the same community as the exposed cohort ? * b) drawn from a different source c) no description of the derivation of the non exposed cohort 3) Ascertainment of exposure a) secure record (e.g. surgical records) ?* Dietary instrument used and reported to be validated b) structured interview ?* c) written self report d) no description 4) Demonstration that outcome of interest (type-2 diabetes) was not present at start of study a) yes ?* b) no Comparability 1) Comparability of cohorts on the basis of the design or analysis a) study controls for ___exposure to known non-nutrient risk factors ____age, BMI, smoking, physical activity. * b) study controls for any additional factor ? Exposure to suspected macronutritional risk factors, at least two from intakes of dietary fiber (or cereal fiber) intake, energy intake, fat intake, and alcohol intake.* Outcome 1) Assessment of outcome * a) independent blind assessment ? b) record linkage ? Clinical report * c) self report d) no description 2) Was follow-up long enough for outcomes to occur. a) yes? Select yes if four or more years of follow-up (low to allow duration of follow up to be assessed as a covariate) * b) no 3) Adequacy of follow up of cohorts a) complete follow up -all subjects accounted for ? * b) subjects lost to follow up unlikely to introduce bias -small number lost -_<20%__ or description provided of those lost ?* c) follow up rate _>20%_lost and no description of those lost. d) no statement.

Attributes of studies on the T2D-GI relation
4.1 Extracted data or corresponding values obtained by calculation from extracted data.  [26] refer to [38] who reported on the correlations for polysaccharides and sugars separately but not for carbohydrate, for which an estimate for carbohydrate was used at present where specified among sensitivity analyses. d Sluijs et al 2013 [22] report values for 4 cohorts out of 15 in their multi-regional study. Other values were not verifiable from the citation provided by Slujis et al [22], which was Margetts [39] who reported values from "0.4 to 0.7" without attribution to particular country regions. For regions combined, a correlation was assumed at 0.55 among specified sensitivity analyses. A value of 0.64 was identified in a full paper investigating the validity of the French regional study [40]. Abbreviations: AA, African-American; CI, confidence interval; EA, European-American; FFQ, Food Frequency Questionnaire; GI, Glycemic Index; HPFS, Health Professionalsʹ Follow-up Study; id, identity; m, men; NL, Netherlands; NHS II, Nursesʹ Health Study 2; Qmax, identifies that RR is at the maximum quantile relative to the minimum quantile; T2D, Type 2 diabetes; w, women, UK, United Kingdom. A value of 10 was assumed when interconverting between T2D-GI relations per Q1 to Q5 and per 10 g GI. c All such: Individual study values by region were not reported in the original studies or citations. Values given in brackets are the those reported as combined values for the eight regional studies by country [22].  In the studies of Oba (m & w), FFQs were applied potentially once or twice for the given dietary values even though 3 assessments were made. g All such: Individual study values by region were not reported in the original studies or citations. Values given in brackets are the those reported as combined values for the eight regional studies by country [22]. Abbreviations: AA, African American; DHQ, diet history questionnaire; EA, European American; FFQ, food frequency questionnaire; HPFS, Health Professionalsʹ Follow-up study; id, identity; m, men; NHS II, Nurses' Health Study II; NL, Netherlands; QDQ, Quantitative diet questionnaire; Q+SI, undefined questionnaire plus structured interview; w, women; UK, United Kingdom. Health professionals a In meta-regression analysis, ascertainment was coded as 1 if self-reported, 0.25 if mixed self and clinically-reported (representing 50% unconfirmed T2D half of which was probable T2D), and 0 if clinically-reported.
b Potential scores are from 0 to 9.
c Not including those participants excluded from entry to the study, for which reasons for exclusion were given in the original reports.
d Type 1 diabetes was not excluded but considered only a minor contamination.
e Jointly for Stevens 2012 AA and EA.
f Doctorʹs confirmation sought but percentage confirmed was not reported.
g Only 0.4% of persons self reporting T2D were confirmed, a large proportion (43%) of requests for confirmation were unanswered. h Of 1608 self-reported cases, 896 were confirmed by medical record, the remainder were unconfirmed.
i All such. Information unavailable .
J Clinical records available confirmed 95% of self-reported T2D, but the proportion of participants' medical records available was not reported. k Jointly for Oba 2013 m and w. l Clinical confirmation of 60% of self-reported T2D.
m Data in brackets not assignable by country. NOS scale for the whole study was 6. n Not declared. Abbreviations: AA, African-American; AU, Australian; EA, European-American; HPFS, Health Professionalsʹ Follow-Up Study; id, identity; m, men; NHS II, Nursesʹ Health Study 2; nd, not declared; NOS, Newcastle-Ottawa study quality; NL, Netherlands; T2D, type 2 diabetes; w, women; UK, United Kingdom. Abbreviations: GI, Glycemic Index; GL, Glycemic Load; LCI, lower 95% confidence interval; RR, relative risk at Qmax (highest quantile relative to lowest quantile); T2D, incident type 2 diabetes; UCL, upper 95% confidence interval.  16 Hopping et al 2010 [16], Japanese American men, RR is based on rate ratios.  17 Hopping et al 2010 [16], Japanese American women, RR is based on rate ratios.  18 Hopping et al 2010 [16], Native Hawaiian men, RR is based on rate ratios.  19 Hopping et al 2010 [16], Native Hawaiian women, RR is based on rate ratios. Studies not included [1,[3][4][5][6][7][8] with reasons are reported in section 3 (above) and titled: Explanations for studies not meeting the inclusion/exclusion criteria for GI and GL combined in Figure 1  Other extracted data and author supplied information are given in subsequent footnotes. c

Attributes of studies on the T2D-GL risk relation.
All such in this column in rows for Q1, authors of the original reports provide 95CI values for relative risks from Q1 to Qn defining the relative risk at Q1 as one with zero degrees of freedom, hence no 95CI values are given for Q1. Author response confirmed further information was not available or not readily accessible [31]. j Calculated: Mean of ten energy intake values (6879+6879+7297+7945+8577+8368+7075+7046+7226+8021)÷10 (kJ/d) [31]. k Other extracted data for European Americans: incremental RR per 1sd of energy adjusted GL (mean and 95%CI) 1.13 (1.0 to 1.276) meant that case and control data were not needed to obtain rates of change in RR with GL in the first step of two-step analysis. 1SD of energy adjusted GL was calculated at 62g for the mean energy intake shown and is the combined SD values obtained on pooling means and SDs for quantiles of energy adjusted GL in Tables 1 and 2 of the original publication [37]. l Calculated: The range of GL from quantile 1 to quantile 5 was obtained assuming a normal distribution calculated from study mean and SD for energy adjusted GL intakes in Tables 1 and 2 of the original publication. The study average of glycemic load was derived from the mean of two sets of ten quintiles values [37], thus (144+130+136+148+172+122+141+150+159+160] ÷10. A value for 1SD of energy adjusted GL was calculated at 62g by combining the SD values for each quantile, and accounting for the SD between quantiles. This complex arrangement was used because information on GL intakes by quantile was available not for GL quantiles directly but was available for fiber and glycemic index quantiles, while correspondence with authors was not able to provide answers. m Calculated: Mean of ten energy intake values (1796+1531+1528+1562+1708+ 1566+1647+1658+1673+1581)÷10 [37]. n Hazard ratio for slope (mean and 95%CI) 0.999 (0.966-1.002) per g GL for African-Americans [37] was extracted, which meant that case and control data were not needed to obtain rates of change in RR with GL in the first step of two-step analysis. o Calculated: Study average of glycemic load was derived from the mean of two sets of 5 quintiles values of (165+135+141+151+177+136+156+164+161+151) ÷10 [37]. p Calculated: Mean of ten energy intake values (1606+1654+1674+1587+1483 +1780+1456+1485+1551+1740)÷10 from reference [37]. q Calculated: Total number of participants (91249) divided by the number of quantiles (5), then less the number of cases tabulated. r Calculated: Using glycemic load (g/d) and glycemic index to calculate carbohydrate intake (g/d), followed by use of carbohydrate intake per unit energy intake (kcal/100kcal energy) to calculate energy intake [20]. s Data provided by correspondence with the first author of the original report [35] who kindly re-analyzed their data with GL adjusted for energy intake by the residual method.  [32], a glucose reference standard was assumed. This appears corroborated by a value of 86 for the same community at a time when white bread was usually a standard [33]. Two corresponding authors were not available to report differently. ac Calculated: Based on the reported fat and carbohydrate intakes [32], calorie conversion factors of 9 and 3.75 kcal/g for fat and carbohydrate as monosaccharide respectively and 14.8% energy as protein average across sexes and tertiles for this population [34]. ad Calculated: Based on reported values of GL (g/d) [17] of 145 sd 32 for men, and 114 sd 23 in women, a normal distribution and the fraction of men in the population of 0.46 being applied to all quantiles.
ae Case and control data were not needed when obtaining the rate of change in RR with GL in the first step of two-step analysis because the rate estimate is based on only one quantile versus referent. Case and control data were only needed when there was multiple data within the study when the case and control data help account for non-independence of observations from the same study [21]. af Calculated from values for each quantile in men and women separately and the fraction of the population that were men, (0.46x(1723+1732+1726+1727+1690]÷5) +(1-0.46)x(1288+1336+1326+1291+1268)÷5. ag By correspondence, the first author of the original report [2] indicates that GL was adjusted for energy intake in men and women separately, with means of 2017 kcal/d in men and 1608 kcal/d in women, with a combined sex mean of 1835 kcal/d. Correspondence confirms GL values were based on the glucose standard, and that all non-European American participants were African-American.  [15]. am Authors explained by correspondence that the published and author provided values of GL for this study (shown above) an Calculated: Number participants less the number of cases, by quantile, data supplied by authors. Values agrees to 1 in 3000 with values calculated as the total number of participants divided by the number of quantiles, then less the number of cases by quantile for the published data [16,37]. ao Case and non-case data was not used because the authors supplied rate information: RR was reported to increase by 1.27 (95%CI: 1.11,1.44) per 1SD rise in reported GL (g/2053kcal) of 21.2 g [21]. This information was re-expressed per 80g GL in 2000kcal. Operationally this was via lnRR rise in glycemic load. ap Data not used in the two-step analysis, but approximated for the meta-analysis of rise in lnRR from the lowest to highest quantile. Data was calculated from information in footnotes 'aq' & 'as'. aq The median glycemic load for quantile 5 was approximated using the reported glycemic load of 117.9g and its SD 21.2 g [21].
Using these values a normal distribution assumed and was simulated for 100000 observations, divided into quintiles, and the median for the fifth quintile obtained.  [15]. a glucose reference standard was evident, as in the prior study from this group at 20y follow-up. bb Reported in published correspondence [23]. Abbreviations: G, glucose; RR, relative risk; WB, white bread.  1 a Further to Tables S10 and S11.
b Correlations were for carbohydrate intake, and are reproduced either from the citation or from its referenced validation study.
Values are after adjustment for energy intake (unless specified differently) and de-attenuation (unless also accompanied by bracketed values, when values in brackets indicated approximate de-attenuated values obtained as described in the main article. The correlation shown is for validation of one application of the instrument. To aid comparability between studies, correlations obtained by repeated measures were not used. c As discussed [35], a discrepancy appears between the published validation of the instrument, which was on a population external to the population sampled for the cohort study, and the reproducibility of the instrument in a sample of the cohort studied. Within the study the FFQ showed "fair" to "moderate" agreement-interpretable from tables of kappa as 0.21-0.40 and 0.41 to 0.60 respectively, for which the mid-range of 0.41 was used as a crude estimate. Adjustments to approximate an energy-adjusted de-attenuated value suggest a value of approx. 0.56 compared with the questionnaires validation, which gave 0.78 but in the different population. d Crude value as reported in the validation publication, in which the authors claim an energy adjustment did not change the result appreciably. Value in parenthesis is after approximate adjustment at present for de-attenuation. e A value for the mixed sex population was the average of values for men (0.73) and women (0.51).  a Further to Tables S10, S11 and S12.
b Calculated values, energy adjusted for glycemic load.
c The Newcastle-Ottawa observational study quality scale ranges from 0 to 9 representing a minimum to maximum quality [37].      [30]. Abbreviations: CI confidence interval; P, probability; RR, relative risk; I 2 , inconsistency, which is ratio of among-studies variance to the sum of among-studies and within-studies variances; Q, quantile of glycemic index. The dietary instrument correlation was not reported for all regional studies combined in this citation and references failed to provide sufficient insight. It is known that some provided correlation coefficients for carbohydrate <0.5 and some >0.5. Assuming a value of 0.55 allowed inclusion of the study to the meta-analytical model and showed show that it was not necessarily outlying.  Table S13 footnote a therein. b

Supplemental analyses on the T2D-GL risk relation
As in Table S13 footnote b therein.
c As in Table S13 footnote c therein.
d As in Table S13 footnote d therein.
e As in Table S13 footnote e therein.
f As in Table S13 footnote f therein. 9. Outlying studies in the T2D-GI & GL risk relations: statistical significance and possible cause.   Early report suggested the T2D-GI relation can be confounded by certain foods with specific associations with incident T2D [44]. Simila et al 2011 [44] found a association between the RR for incident T2D and dietary GI of 1.32 when they excluded beer and milk from their calculation of the dietary GI values, which was an RR expected from the present meta-analyses .
By contrast RR was lower at 1.06 when confounded by milk and/or beer [44]. Aside from that, the low T2D-GI RR became inlying when the analytical model included the average sampled population alcohol consumption was included as a covariate (centered on 7 g/d) alongside CORR (centered on 0.7), ethnicity (centered on 0) for European-American vs other ethnicities included), and duration of follow-up (centered on 10 y). b.
van Woudenbergh (2011) [34] noted that the range of GI values across the quantiles for their study was narrow (approx. 6 GI units), perhaps too narrow to observe a reliable result. A definitive explanation was not available at this time.
c. Meyer et al [31] reported a high risk of misclassification of both foods and incident diabetes. Thus validation of the dietary instrument for carbohydrate gave a low value of 0.45. Meanwhile incident type-2 diabetes was self-reported with potentially only 66% of cases validated by medical record. The study became inlying when CORR was a covariate in the analytical model