A systematic review on the effect of sweeteners on glycemic response and clinically relevant outcomes

We did a comprehensive search designed by a MLIS-trained librarian to identify all randomized controlled trials (RCTs) comparing sweeteners in generally healthy, overweight/obese and/or diabetic participants. We included only trials published in English as full peer-reviewed manuscripts. MEDLINE (1950 to January 13, 2011), EMBASE (1980 to January 13, 2011), CENTRAL (January 13, 2011), and CAB (January 13, 2011) were searched. No existing systematic reviews were found. The specific strategies used are provided in Additional File 1. The citations and abstracts were screened by two reviewers to identify pertinent trials. Any study considered potentially relevant by one or both reviewers was retrieved for further consideration.

We considered non-caloric sweeteners to include high-intensity caloric sweeteners that are functionally non-caloric simply due to extremely low doses (for example, aspartame). Each potentially relevant study was independently assessed by two reviewers for inclusion in the review using predetermined eligibility criteria. Disagreements were resolved by consultation with a third party. Trials with healthy, overweight/obese, and/or diabetic adult (≥ 16 years old) participants meeting the following criteria were eligible for inclusion: parallel or crossover RCTs; weight change, energy intake, lipids, glycated hemoglobin (HbA1C), or insulin resistance were reported; had at least two groups comparing different sweeteners (for example, glucose, fructose, sucrose, other saccharides, sugar alcohols, non-caloric sweeteners: aspartame, saccharin, stevioside, sucralose); and where follow-up was at least one week in duration (see the Box in Additional File 1 for study selection summary). RCTs measuring 2-hour blood (serum or plasma) glucose responses in similar populations without the follow-up requirement were also reviewed. All outcomes selected for study (including weight change) are reversible and thus (providing that order was randomly assigned), a cross-over design should be appropriate. Trials with less than ten participants per group were excluded to improve the efficiency of the work without an appreciable loss of power, and with the possible elimination of some small study bias. Trials aimed at evaluating exercise performance or memory enhancement were excluded. Trials with placebo controls were also excluded as we aimed to investigate comparative effectiveness of different sweeteners, as opposed to exploring the implications of avoiding sweeteners altogether.

A standardized data extraction method was performed by a single reviewer. A second reviewer checked the extracted data for accuracy. The following properties of each trial were recorded in a database: trial characteristics (country, design, sample size, duration of follow-up); participants (age, gender, co-morbidity (obesity, diabetes mellitus - type 1 and 2), baseline body mass index (BMI), diabetic therapy (insulin, oral antihyperglycemic agents, diet, and so on); sweetener characteristics (type, quantity, schedule); diet (that is, daily caloric content by macronutrient/fiber content); and outcomes. Outcomes included weight change (absolute, BMI), energy intake, lipid measures (total cholesterol, triglycerides, high density lipoprotein (HDL), low density lipoprotein (LDL)), HbA1C, insulin resistance (for example, Homeostatic Model Assessment (HOMA) index), and 2-hour blood glucose (with or without meals).

Risk of bias was assessed using items known to be associated with the magnitude of results (that is, method of randomization, double-blinding, description of withdrawals/dropouts, and allocation concealment) [12, 13]. Source of funding was also extracted given its potential to introduce bias [14].

Data were analyzed using Stata 11.1 (http://​www.​stata.​com). Missing standard deviations (SDs) were imputed using the maximum value reported in any included study [15]. Missing correlations for change from baseline and for differences between crossover trial periods were assigned a value of 0.63, the maximum reported value in the included studies. Changes from baseline means were used in place of final means in parallel randomized trials. For weight change, the baseline value prior to the immediate period was used. The mean difference (MD) was used to summarize outcomes. Due to expected diversity between studies, we decided a priori to combine results using a random effects model (Stata command: metan). Additionally we planned to examine the association between certain variables (population, dose, diet, age, gender, and bias criteria) and the effect of specific sweeteners on outcomes, and publication bias with weighted regression [16], however the available comparisons were too sparse to pool trials with outcomes of one week or less. For 2-hour responses, we pooled comparisons by type of sweetener and ordered the matrix tables by expected order of efficacy [17] (that is, non-caloric sweeteners, sugar alcohols, other saccharides, fructose, sucrose, and glucose). Statistical heterogeneity was quantified using the τ² statistic (between-study variance) [18]. Furthermore, we explored comparative effectiveness of sweeteners on 2-hour responses using network meta-analysis [19] (specifically, Markov chain Monte Carlo [MCMC] methods within a Bayesian framework) in WinBugs (http://​www.​mrc-bsu.​cam.​ac.​uk/​bugs; code was obtained from Ades et al. [20]). Network analysis extends meta-analysis from simply pooling directly compared treatments (direct evidence) to pooling data from studies not directly compared but linked via one or more common comparators (indirect evidence) by assuming consistency of the evidence [19]. Therefore, this technique facilitates the comparison of any two sweeteners not directly compared in any one study. We used non-informative prior distributions: uniform for the between-study variance (range 0 to 20) and Gaussian for the other parameters (mean 0 and variance 10,000). All chains were run for 10,000 iterations after 1,000 burn-in iterations. Convergence of the MCMC algorithm was assessed using autocorrelation plots. By-population results were generated. Inconsistency in the network (disagreement between direct and mixed evidence) was measured using back-calculations [21]. Ninety-five present Bayesian credible intervals are reported.

The searches identified 3,666 unique records with no trials found outside the main literature searches. After initial screening, 491 articles were retrieved for detailed evaluation (Figure 1) and of these 440 articles were excluded resulting in 53 trials (from 51 publications) that met the selection criteria. Disagreements about the inclusion of studies occurred in 11% of the articles (kappa = 0.71). Fourteen were ultimately included. The remaining were excluded for the following reasons: thirteen with no relevant control group, nine with no relevant population, five with no relevant intervention group, four due to study design, four for small sample size, and one for no usable data. The sweeteners studied in eligible trials are described in Table 1.

Of the forty included trials with 2-hour response data (703 participants; Table 2), three trials compared a non-caloric sweetener (aspartame [22, 23], or sucralose [24]) to a saccharide (fructose [24] or sucrose [22, 23]); one trial compared a non-caloric to another non-caloric (aspartame versus saccharin [25]); four trials compared a sugar alcohol or a malt containing a sugar alcohol (sorbitol [26], xylitol [26, 27], maltitol [26], Lycasin [26, 28], or a hydrogenated starch hydrolysate (HSH) [29]) to a saccharide (glucose [26, 28, 29] or sucrose [27]); and thirty-two trials compared a saccharide to another saccharide (glucose [30‐51], fructose [31‐34, 36, 38‐41, 44‐47, 49, 50], mixtures of glucose and fructose [that is, sucrose [34, 37, 39, 42, 43, 48‐57], HFCS [42, 54, 55, 58], honey [48, 57, 59], glucose/fructose equivalent honey [59]], isomaltulose [52, 56], maltose [53], sucromalt [58], trehalose [30], or a mixture of trehalose and fructose [30]). Approximately half of the doses for saccharides were less than the 60 g/day recommended for diabetic patients on a 2,000 kcal diet; the remainder exceeded 60 g/day (typically 75 g). All of the doses for sugar alcohols exceeded the 10 g/day recommendation (range 20 to 50 g), which is aimed at limiting gastrointestinal symptoms. None of the four non-caloric sweetener groups were above Acceptable Daily Intake (ADI) values.

Twelve trials included diabetic populations (range mean BMI 23 to 32 kg/m²) [25, 27‐29, 36, 42, 45, 48‐50, 56, 57], five trials exclusively studied overweight or obese individuals (range mean BMI 28 to 35 kg/m²) [30, 31, 41, 45, 52], and thirty-five trials included generally healthy individuals (range mean BMI 20 to 26 kg/m²). Median mean age was 35 years (range 22 to 72 years) and median sex distribution was 47% women.

Sample size ranged from 10 to 37 (median 17), three studies (8%) had sample sizes ≥ 30 per group and all were randomized crossover trials. The median Jadad score was 1 (range 1 to 4); no studies reported concealing treatment allocation.

Table 3 reports the results of the direct meta-analysis for all populations in the lower triangle and the mixed evidence from the Bayesian network (Figure 2) in the upper triangle. The network included 36 trials and 610 participants. The direct evidence from all nine comparisons was consistent with the mixed evidence from the network. There was large heterogeneity between trials (I²'s ≥ 77%) for three of seven multi-study direct evidence comparisons. Two of the heterogeneous comparisons included a variety of sweeteners (that is, multiple sugar alcohols (τ² = 9.05 (95% CI 2.94,32)), or multiple other sugars (τ² = 1.72 (0.37,1.48))) within one category. In the fructose versus glucose comparison, six trials were responsible for the heterogeneity (τ² = 1.40 (0.68,1.50)). Three [36, 45, 50] were subgroups of diabetic participants; they increased the magnitude of the mean difference. The other three trials [32, 33, 46] showed important differences prior to the 2-hour time point (data not shown) but at two hours showed little or no difference between sweeteners. The single estimate of heterogeneity (τ²) for the network meta-analysis was 0.65 (95% CI 0.35,1.10).

Reporting the mixed evidence, two comparisons: fructose versus sucrose (MD -1.12 mmol/L (-1.95,-0.27)), and fructose versus glucose (-1.56 mmol/L (-2.18,-1.02)) were statistically significant, all favoring fructose, but neither of the confidence limits excluded the possibility of non-clinically relevant differences (< 1·15 mmol/L - calculation based on a clinical important difference of 1% for HbA1C) [60]. The weighted regression test for publication bias was not significant.

In the subnetwork of 31 trials enrolling participants without diabetes (446 participants; τ² = 3.66 (1.66,7.31); Appendix Table 1 in Additional File 1), the direct evidence from all 8 comparisons was consistent with the mixed evidence from the network. The heterogeneity although reduced remained large between trials (I²'s ≥ 60%) in both of the remaining multi-study direct evidence comparisons. Using the mixed evidence, three comparisons: fructose versus sucrose (-0.54 mmol/L (-1.06,-0.03)), fructose versus glucose (-0.89 mmol/L (-1.21,-0.59)), and fructose versus other sugars (-0.85 mmol/L (-1.47,-0.21)) were statistically significant, all favoring fructose, but none of the confidence limits excluded the possibility of non-clinically relevant differences.

In the subnetwork of ten trials enrolling participants with diabetes (152 participants; Appendix Table 2 in Additional File 1), the direct evidence from all six comparisons was consistent with the mixed evidence from the network. Note, this network did not include non-caloric sweeteners. Because the estimate of τ² (224 (0.14,139)) did not converge, we report our findings from the direct evidence. Three direct comparisons were significant and found clinically relevant differences between agents over the entire confidence interval span: fructose versus glucose in 5 trials with 52 participants (-4·81 mmol/L (-6·34,-3·29), I² = 0%, τ² = 0 (0,7.47)), HSH versus glucose in 1 trial [29] with 12 participants (-6·19 mmol/L (-9·78,-2·60)) and isomaltulose versus sucrose in 1 trial [52] with 20 participants (-3·44 mmol/L (-5·31,-1·56)).

Of the 13 trials (412 participants; Table 4), 3 trials compared a non-caloric sweetener (aspartame [61], cyclamate [62], or a mixture [63]) to sucrose, and 10 trials compared a saccharide to a different saccharide (glucose [64‐66], fructose [64, 65, 67], mixtures of glucose and fructose [that is, sucrose [66‐72] or honey [69]], FOS [71‐73], a mixture of isomaltulose and sucrose [68], or tagatose [70]). No trials evaluated stevioside. Seven trials did not give any daily diet recommendations; one FOS trial recommended low-fiber intake [72]; one restricted added sweeteners to the assigned sweetener [62]; three trials restricted total energy levels and composition of macronutrients (55% carbohydrate, 30% fat, 15% protein) [64, 65, 67]; and another restricted total energy levels and the composition to the assigned sweetener plus calcium caseinate [66]. With three exceptions [63, 64, 66], the doses of sweeteners were at or below current clinical practice guideline (CPG) recommendations (10% of total energy intake (for example, 60 g of sucrose in a 2000 kcal diet) although only three trials [64, 65, 67] restricted overall energy intake, therefore further sweetener consumption may have exceeded current recommendations. One trial [63] prescribed sweeteners (simple carbohydrates) at 25% of total energy intake - the American Diabetes Association (ADA) 2004 recommended maximum. The earliest trial [66] prescribed sweeteners at 87% of total energy intake - they were differentiating energy availability from energy content.

Four trials were in diabetic populations [62, 67, 71, 73], seven trials were in generally healthy populations [61, 64‐66, 68, 70, 72] and two trials were in overweight/obese [69] or overweight [63] populations. Mean BMI levels ranged from 21 to 31 kg/m². Median mean age was 35 years and median sex distribution was 54% women.

Sample size ranged from 10 to 133 (median 20), 1 had a sample size ≥ 30 per group and duration of follow-up ranged from 1 to 12 weeks (median 4 weeks). Ten were crossover trials [62, 64‐67, 70‐73] and four were parallel trials [61, 63, 68, 69]. Jadad scores ranged from 1 to 2 (median 1). Twelve of thirteen trials did not report whether or how treatment assignment was concealed. One used alternating assignments according to body weight [70].

Two trials reported change in BMI (Table 5). The 4-week trial in healthy participants [61] did not find a significant loss in BMI in non-caloric sweetener recipients (-0.3 kg/m² (-1.1,0.5), 133 participants). The trial in overweight participants [63] found a significantly greater loss in BMI over ten weeks of follow-up in participants consuming the non-caloric sweetener (-0.9 kg/m² (-1.5,-0.4), 41 participants). Two trials reported absolute change in weight. One crossover trial was done in type 1 diabetic participants and found no difference in weight loss between groups over four weeks (0.8 kg (-3.3,4.9), ten participants [62]). The other trial in overweight participants [63] found significantly greater weight loss over 10 weeks in the non-caloric sweetener group (-2.6 kg (-3.7,-1.5), 41 participants).

Two trials reported energy intake; both reported a significant effect of non-caloric sweeteners. The 4-week trial in generally healthy participants [61] found a significantly reduced intake of calories in non-caloric sweetener participants (-283 kcal (-414,-153), 133 participants).The trial in overweight participants [63] also found significantly less energy intake (over one day) in the non-caloric sweetener group after ten weeks of follow-up (-491 kcal (-806,-177), 41 participants).

Available trials found no effect of sweetener type on HbA1C (one trial: -0.02% over four weeks (-0.40,0.30), ten participants [62]) or the HOMA index (one trial: -0.20 over ten weeks (-0.58,0.18), forty-one participants [63]). The trial in ten type 1 diabetic participants [62] found no effect on total cholesterol, HDL cholesterol, or triglycerides over the course of four weeks; the other trial in forty-one overweight participants [63] found no effect on triglycerides over the course of ten weeks.

Two trials reported change in BMI (Table 6); one comparing honey to sucrose in overweight/obese participants over 4 weeks of follow-up [69]; the other comparing a mixture of isomaltulose and sucrose to sucrose over 12 weeks of follow-up [68] in healthy participants. Neither found a significant difference between sweeteners. One trial compared FOS to glucose [73] (three weeks in twenty diabetic participants) and one trial compared FOS to sucrose [72] (four weeks in twelve healthy participants), respectively. Neither found a difference in absolute weight change. Five other trials done in varying populations (including overweight/obese [69] or healthy populations [64‐66, 68]) found no differences in change in absolute weight between sweeteners. Two trials reported energy intake (FOS compared with glucose [73] and sucrose [72] respectively, but neither found a significant difference.

Two trials (one comparing FOS to sucrose [71] and one comparing isomaltulose/sucrose to sucrose [68]) found no significant effect on HbA1C. However, the latter [68] found a significant decrease in the HOMA index among isomaltulose/sucrose recipients (-0.44 (-0.76,-0.12)).

Seven trials reported change in total cholesterol. The pooled result of two trials [71, 72] comparing FOS to sucrose was statistically significant (0.26 mmol/L (0.03,0.48), I² = 0%, τ² = 0 (0,0.01)), although this conclusion was based on a total of only twenty-two participants. One trial comparing isomaltulose and sucrose to sucrose (50 healthy participants over 12 weeks) [68] found a significantly smaller increase in total cholesterol for the isomaltulose/sucrose group (-0.10 mmol/L (-0.17,-0.02)). No trials found an effect of sweetener type on LDL cholesterol or HDL cholesterol. The trial comparing isomaltulose and sucrose to sucrose [68] also found a significant effect on triglycerides (-0.27 mmol/L (-0.44,-0.10), 0.11 decrease versus 0.16 mmol/L increase). However, four trials studying other combinations of sweeteners [69, 71‐73] found no effect of sweetener choice on triglyceride levels.

Head-to-head comparisons between saccharides did not identify any statistically significant differences. The confidence limits of these results either included minimally important differences or the group sizes were too small (< 30) to have good estimates of standard deviation [74]. The one exception was the comparison between sucrose and FOS, which suggested that total cholesterol was reduced to a greater extent with sucrose than with FOS. However, the confidence intervals for this analysis included values that were not clinically relevant (0.03 to 0.59 mmol/L). There was no evidence that HFCS or fructose increased levels of cholesterol relative to other sweeteners.

Although we found that fructose reduced 2-hour blood glucose concentrations by 4.81 mmol/L compared to glucose in diabetic participants, data comparing non-caloric and sugar alcohols to the more commonly used sucrose or HFCS were inconclusive. Contrary to perception and current recommendations, no substantive evidence describing important long-term benefits of hypocaloric sweeteners for diabetic patients were identified. Also, despite popular belief, no high-quality RCT evidence was found indicating that fructose causes or exacerbates hypertriglyceridemia [6].

Although the identified trials were numerous, they were very small and largely short-term. We found 13 trials with participant follow-up longer than 1 week and group sizes ≥ 10: 3 that compared non-caloric sweeteners to sucrose, and 10 that were head-to-head comparisons of saccharides. Ten of 13 trials had a Jadad score of 1 and none adequately concealed treatment assignment prior to assignment. Although blinding the participants would have been impossible in many of the trials due to taste differences between sweeteners [63], the reporting of important design descriptors were largely absent, indicating a substantial risk for bias [12, 13]. The longest trial was only 10 weeks - not long enough to determine whether substituting a non-caloric sweetener for a caloric sweetener is sustainable in daily practice. To detect an important reduction in weight over at least one year such as 2.5 kg/m² (less than 0.05 kg/m²/week) in a RCT would require a minimum of 85 participants (assumptions: 25% loss-to-follow-up, α = 0.05, power = 90%, SD = 3 kg/m²).

Our network meta-analysis had several limitations: 1) the sugar alcohol and other sugar categories contained multiple sweeteners that are likely to have different blood glucose profiles thereby inducing heterogeneity, 2) power to detect inconsistency is limited by the number of trials included in each test, and 3) the back-calculation method used to detect inconsistency involved multiple tests thereby increasing the false-positive rate. However, we did not detect any inconsistency.

Another limitation was that only three studies restricted the total energy consumed by each participant. Therefore, participants may have supplemented energy lost with non-caloric sweeteners with other food products - sweetened or otherwise. However, it may be argued that this is a strength of the trials - in that they reflect what happens in real world self-management diet practices. Lastly, and perhaps most importantly, all studies were small, thereby underestimating standard deviation and as a result underestimating confidence interval widths and increasing the likelihood of false-positive findings [74]. Despite this, the confidence intervals for many analyses were wide and did not exclude a minimally important difference. Small study bias (or publication bias) may also play a role in our findings concerning longer-term outcomes.

In theory, substituting non-caloric and lower caloric sweeteners for simple sugars should reduce energy intake and thereby the risk of obesity and its consequences. However, there are a number of reasons why increasing use of non-caloric and lower caloric sweeteners might not lead to the expected improvements in energy regulation. First, use of hypocaloric sweeteners might not induce weight loss even in the short term. For example, if reductions in calories due to sweeteners are offset by increases in caloric intake from other sources [75, 76], or offset by decreases in caloric expenditure [77, 78]. Although our data suggest that non-caloric sweeteners may lead to clinically relevant weight loss through reduced energy consumption, this conclusion was driven by a single trial with a total of 41 participants. Unlike caloric sweeteners (which may partially compensate added calories with reduced energy intake from other sources) [79], non-caloric sweeteners are not known to suppress appetite, and therefore would not reduce the motivation to eat. Furthermore, it has been suggested that the psychobiological signals with non-caloric sweeteners may directly influence physiological regulatory mechanisms and thus further reduce their potential for reducing net energy intake [75, 80]. Second, if calorie reduction is not maintained, short-term reductions in weight due to the use of hypocaloric sweeteners might not be sustained. Third, it is possible though speculative that any health benefits due to weight loss from non-caloric sweeteners might be wholly or partially offset by currently unrecognized adverse events due to their use. The lack of data on the long-term benefits of non-caloric sweeteners means that it is currently impossible to determine whether these substances will improve public health.