The impact of glycaemic load on cognitive performance: A meta-analysis and guiding principles for future research

The effect of breakfast glycaemic load (GL) on cognition was systematically examined. Randomised and non-randomised controlled trials were identified using PubMed, Scopus, and Cochrane Library (up to May 2022). 15 studies involving adults (aged 20 - 80 years) were included. Studies had a low risk, or some concerns, of bias. A random-effects meta-analysis model revealed no effect of GL on cognition up to 119minutes post-consumption. However, after 120minutes, immediate episodic memory scores were better following a low-GL compared to a high-GL (SMD = 0.16, 95% confidence interval [CI] = -0.00-0.32, p = 0.05, I2 = 0%). Subgroup analyses indicated that the benefit was greater in younger adults (<35 years) and those with better GT. A qualitative synthesis of 16 studies involving children and adolescents (aged 5 - 17 years) suggested that a low-GL breakfast may also benefit episodic memory and attention after 120minutes. Methodological practises were identified which could explain a failure to detect benefits in some studies. Consequently, guiding principles were developed to optimise future study design.


Introduction
Several reviews have concluded that breakfast benefits various aspects of cognition in adults, children, and adolescents (Adolphus et al., 2016;Anderson et al., 2020;Galioto and Spitznagel, 2016;Guyatt et al., 2008;Hoyland et al., 2009). However, a more important question has not received the same attention: what type of breakfast optimally enhances acute cognitive performance? Numerous studies have examined this question by manipulating the glycaemic index (GI) of breakfast (Philippou and Constantinou, 2014). GI is a measure of the speed and duration of the increase in blood glucose that results from consuming a set weight of carbohydrate. A related concept is GL, which considers both the GI and the amount of carbohydrate supplied in a food item or meal. GL provides an overall measure of the total glycaemic impact of a specific portion of food and more strongly predicts an individual's glycaemic response than GI (Barclay et al., 2005). For these reasons, the present review focused on the cognitive effects of breakfast GL rather than GI.
It has been hypothesised that a low GL (LGL) breakfast may benefit cognitive performance two to three hours after consumption, reflecting a continuous source of glucose for the brain (Benton et al., 2003;Cooper et al., 2012;Young and Benton, 2014). Conversely, a high GL (HGL) breakfast produces a rapid rise and drop in blood glucose levels which may disrupt cognitive performance, particularly later in the morning as glucose concentrations fall (Nilsson et al., 2009;Young and Benton, 2014). Although this suggestion is conceptually appealing, results have been inconsistent. Some studies have reported that a LGL breakfast benefitted episodic memory, working memory, and attention at various times throughout the postprandial period (Benton et al., 2003;Cooper et al., 2012;Ingwersen et al., 2007;Mahoney et al., 2005;Nabb and Benton, 2006a;Nilsson et al., 2012;Wesnes et al., 2003;Benton, 2014, 2015), whereas other studies have reported that a HGL breakfast benefitted cognitive performance (Dye et al., 2010;Nabb and Benton, 2006a;Smith and Foster, 2008;Young and Benton, 2014).
To the best of our knowledge, only one review has examined the effect of breakfast GL on acute cognitive performance (Gilsenan et al., 2009). The authors concluded that there was insufficient evidence to support an effect of breakfast GL on cognitive performance in children, adolescents, and adults. However, several studies have since been published, potentially enabling a quantitative synthesis of the literature (Anderson et al., 2020(Anderson et al., , 2018(Anderson et al., , 2021Deng et al., 2021;Lee et al., 2019;Sanchez-Aguadero et al., 2020;van der Zwaluw et al., 2014;Benton, 2014, 2015). Furthermore, it is plausible that inconsistencies in the literature are due to methodological factors, including differences in the time that cognitive performance was assessed during the postprandial period, sample age, task domain, and glucose tolerance (GT). The impact of these factors on the cognitive effects of breakfast GL has not yet been systematically explored via meta-analysis.
Therefore, we performed an up-to-date systematic review and metaanalysis of the impact of breakfast GL on cognition in children, adolescents, and adults. The primary aim of this meta-analysis was to assess the influence of methodological factors including the timing of testing, sample age, and GT. The secondary aim was to create a series of guiding principles that outline variables that may need to be considered when designing studies in future. It is hoped that the identification of these variables will facilitate a better understanding of the relationship between breakfast GL and cognitive performance.

Method
This systematic review and meta-analyses were conducted in accordance with the Preferred Reporting Items for Systematic reviews and Meta-Analysis (PRISMA) guidelines. PRISMA 2020 checklists are provided in Supplementary Document 1 (Table S1 & S2). A study protocol was registered with PROSPERO (CRD42021229575).

Search strategy and selection criteria
A systematic search for studies published up to 31st May 2022 was conducted using PubMed, Scopus, and Cochrane Library. To identify relevant publications, the following search terms were used: 'cognitive function' or 'cognition' or 'cognitive performance' combined with 'glycaemic index' or 'glycaemic load' or 'breakfast' or 'carbohydrate' or 'glucose' or 'sucrose' or 'isomaltulose'. The search was restricted to English-language articles only, and both British and American spellings of key search terms were used. Reference lists from articles and reviews identified during the electronic search were checked for additional studies. We also used the British Library of Electronic Theses Online Service (http://ethos.bl.uk) to identify unpublished studies. Titles and abstracts were read to check for duplicates and to determine whether the study potentially met the inclusion criteria. Studies that did not fulfil the inclusion criteria or were clearly irrelevant to the review were eliminated. The remaining articles were read to establish their suitability. The systematic search was conducted independently by two authors (C.G and H.Y). Any disagreements were resolved by discussion.
As postprandial glycaemia has a diurnal rhythm, which may influence the cognitive effects of GL, we chose to focus exclusively on studies that manipulated the GL of breakfast (Gibbs et al., 2014;la Fleur et al., 2001). Breakfast was defined as the first meal or drink of the day, consumed between 6 am and 10 am, after an overnight fast. Studies were included if they met the following criteria: (A) randomised controlled trials (RCTs) or non-RCTs (B) studies that investigated the acute cognitive effects of variations in breakfast GL or provided adequate information from which GL could be calculated, (C) published or unpublished studies, (D) studies that used objective measures of cognition, and (E) studies involving children/adolescents (5-17 years) or adults (> 18 years) who were healthy (i.e., no diagnosis of disease) or had impaired glucose tolerance (IGT) or type 2 diabetes mellitus (T2DM). The exclusion criteria were as follows: (A) studies that compared food/drink intake with water, artificial sweeteners, or food/drink omission, (B) studies that compared the chronic cognitive effects of dietary GL or GI, or (C) studies that manipulated the GL or GI of nutritional interventions consumed after breakfast time (e.g., an afternoon snack).

Data extraction
Two authors independently extracted the following information using a standardised data spreadsheet: first author name, year of publication, participant characteristics (age, gender, & GT status), study characteristics (sample size, type of design, counterbalancing, randomisation, blinding, length of overnight fast, adjustment for confounding factors, length of washout period between test sessions, number of withdrawals, & control for previous days meal consumption/physical activity levels), characteristics of breakfast interventions (GL, GI, macronutrient content, & energy content), timing of blood glucose/ cognitive tests, type of cognitive domain/subdomain assessed, name of cognitive test, and results of study.
The GL of a breakfast intervention was calculated using the reported GI value multiplied by the amount of available carbohydrate per serving and divided by 100. If the GI of a meal or drink was not reported, it was estimated using values provided by Atkinson et al. (2021) or Sydney Universities Glycaemic Index Research Service (SUGiRS). The two breakfast interventions with the largest difference in GL were categorised as the HGL or LGL interventions. Remaining meals or drinks were categorised as MGL interventions.
For the meta-analysis, means and standard deviations (SDs) of each cognitive outcome, at each postprandial time point, after the LGL and HGL breakfast intervention were extracted. Sample sizes were recorded and, where possible, both adjusted and unadjusted means were extracted. Authors were contacted when data were missing or only change scores were reported. When an author did not respond, data were estimated from figures when available, or the study was not included in the meta-analysis (see Section 3.2.3.).

Organisation process
Using the framework described by Harvey (2019), data were first categorised into one of the following cognitive domains and subdomains: memory (episodic, working, visuospatial, & semantic), attention (selective & sustained), processing speed, executive function (reasoning, problem solving & inhibitory control), and psychomotor function. Next, as cognitive performance was measured at various time points throughout the postprandial period, data were further categorised into one of three time windows: early postprandial period (testing commenced between 10 and 59 min), mid postprandial period (60 -119 min), or late postprandial period (120 min or later). These time windows were chosen to reflect specific points in the typical postprandial glycaemic response.

Risk of bias and certainty of evidence
Risk of bias (RoB) was assessed by two independent authors using The Cochrane RoB 2 tool (Sterne et al., 2019) for crossover and parallel trials. Disagreements were resolved through discussion. The following sources of bias were assessed: randomisation process (selection bias), deviations from intended interventions (performance bias), missing outcome data (attrition bias), measurement of outcome (measurement bias), selection of the reported results (reporting bias), and overall bias. Studies were classified as either 'low risk of bias', 'high risk of bias', or 'some concerns of bias'. To obtain additional information, theses and study protocols were checked and study authors were contacted where possible.
Certainty of evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach (Guyatt et al., 2008). Each cognitive subdomain was assessed based on risk of bias, inconsistency, indirectness, imprecision, and publication bias. There were four possible outcomes: very low, low, moderate, or high.

Data handling and statistical analysis
Meta-analyses were conducted using a generic inverse variance method in Review Manager 5.3 (RevMan) [Computer programme]. Version 5.3. Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2014. All analyses used a random effects model. Effect sizes reflected the standardised mean difference (SMD) between the HGL and LGL breakfast interventions, with 95% confidence intervals (CI). A minimum of three studies per analysis were required. Using the guidelines reported by Cohen (2013), SMDs were interpreted as trivial (< 0.2), small (between 0.2 & 0.6), moderate (between 0.6 & 1.2), or large (between 1.2 & 2.0). A p value of < 0.05 was considered significant, whilst a p value between 0.06 and 0.1 was considered a trend. Heterogeneity was examined using the l 2 statistic, a value above 50% indicated substantial heterogeneity which required exploration.
Studies involving adults were analysed separately from studies involving children and adolescents. To examine the influence of time, three separate meta-analyses were performed for each postprandial time window where possible (early postprandial period = 10 -59 min, midpostprandial period = 60 -119 min, or late postprandial period = 120 min or later). Two a priori subgroup analyses were also performed where data were available. To examine the influence of age in adults, studies were categorised into either the 'younger' or 'older' subgroup based on whether the mean age of the sample was below or above 35 years of age. This cut-off value was chosen because it achieved the most equal subgroup sample sizes, however there is also evidence that certain aspects of cognitive function peak at approximately 35 years of age (Hartshorne and Germine, 2015;Strittmatter et al., 2020). To examine the influence of GT in adults, participants were classified as having 'better' or 'poorer' GT if fasting glucose was below or above 6.1 mmol/L and/or two-hour glucose was below or above 7 mmol/L. Due to an insufficient number of studies, the influence of age and GT was not examined in children and adolescents. For the same reason, the following post-hoc subgroup analyses were not performed: (A) use of dairy (no dairy vs. small quantity of dairy vs. large quantity of dairy) or (B) difference in GL between nutritional interventions (smaller difference vs. larger difference).
Sensitivity analyses were conducted using the leave-one-out method to determine the influence of each individual study on the pooled effect size and p value (Supplementary Document 1, Table S5). We also examined the impact of removing studies that did not match both the macronutrient and energy content of breakfast interventions and compared effect sizes when unadjusted or adjusted means were included in the analysis. Publication bias was assessed by visual inspection of funnel plots; a minimum of 10 studies were required. The Cochrane Handbook for Systematic Reviews of Interventions states that change scores and post-scores should not be analysed as SMDs together. Therefore, separate analyses were performed when necessary. In cases where the same cognitive domain was measured twice within the same postprandial time window, the measurement taken at the time closest to the other studies was used.

Study selection and characteristics
As shown in Fig. 1, 2381 publications were initially identified, of which 45 were potentially eligible. These studies were read fully, and a further 12 studies were excluded. A total of 33 studies met the inclusion criteria, 17 involving adult participants (Table 1) and 16 involving children/adolescents (Table 2). Data were obtained from 15 adult studies for measures of episodic memory, working memory, or attention. We were unable to extract suitable data from Dye et al. (2010) and Kaplan et al. (2000) hence these studies were not included in the meta-analysis. Several other cognitive domains/subdomains were assessed, but these were not quantitatively analysed due to limited data. A detailed discussion of these cognitive outcomes can be found in Supplementary Document 2.
Data were obtained from 11 child/adolescent studies for measures of episodic memory, working memory, or attention. However, for several reasons, a meta-analysis of the data was considered inappropriate. Firstly, we were unable to obtain data from five studies (Benton et al., 2007b;Cooper et al., 2012Cooper et al., , 2015Lee et al., 2019;Taib et al., 2012). In some cases, data were available but only for certain cognitive subdomains (Ingwersen et al., 2007;Mahoney et al., 2005;Wesnes et al., 2003). Secondly, three studies only reported change scores (Brindal et al., 2012(Brindal et al., , 2013Wesnes et al., 2003). As change scores and post-scores were analysed separately, this reduced the number of studies included in each analysis, often to the point that there were no longer enough studies to perform a meta-analysis (i.e., less than 3 studies). For example, there were not enough studies to analyse the effect of breakfast GL on immediate episodic memory during the early and late postprandial period and delayed episodic memory during the early and mid-postprandial period. Consequently, we chose to only perform a systematic review of the literature, which can be found in Section 3.3.3. For brevity, only the aforementioned cognitive measures are discussed. However, a detailed discussion of the effect of breakfast GL on other aspects of cognition, such as executive function and visuospatial memory, can be found in Supplementary Document 2.

Risk of bias and certainty of evidence
The results of the risk of bias assessment are summarised in Supplementary Document 1 (Table S3). Overall, five studies showed a low risk of bias, and 12 studies showed some concerns of bias. Studies were generally judged with some concerns of bias for the following reasons: (A) the method of randomisation and/or allocation concealment was not reported, (B) a pre-registered study protocol was not found; therefore, it is unclear whether statistical analysis plans were changed and/or whether certain cognitive outcomes were selectively reported, and (C) it was unclear whether the researcher(s) administering the cognitive tasks were aware of which breakfast intervention was consumed. For the 'deviations from intended interventions' domain, studies that used a crossover design were judged as having a high risk of bias if participants were clearly not blinded.
The certainty of evidence assessment is reported in Supplementary Document 1 (Table S4). Assessments ranged from very low (working memory & attention) to low (episodic memory). The main issues were risk of bias, as discussed above, and imprecision.

Study characteristics
All studies were randomised, of which four used a parallel design and 13 used a crossover design. Mean age ranged from 20.36 years (Nabb and Benton, 2006a) to 78 years (van der Zwaluw et al., 2014), and sample sizes ranged from 18 (Lamport et al., 2014) to 189 participants (Nabb and Benton, 2006a). The majority of studies recruited healthy participants, several of which examined the influence of GT on the relationship between breakfast GL and cognition (Anderson et al., 2018(Anderson et al., , 2021Kaplan et al., 2000;Benton, 2006a, 2006b;Nilsson et al., 2009Nilsson et al., , 2012van der Zwaluw et al., 2014;Young and Benton, 2014). Two studies recruited participants with clinically diagnosed T2DM (Lamport et al., 2013a;Papanikolaou et al., 2006), and one study recruited participants with IGT (Lamport et al., 2014).

The influence of individual differences.
Individual differences in GT were related to the effect of breakfast GL on delayed episodic memory, but only during the late postprandial period (150-220 min; Fig. 5). In the 'better' GT subgroup, there was a trend towards better delayed episodic memory following the consumption of a LGL breakfast compared to a HGL breakfast (SMD = 0.22, 95% CI = − 0.01 to 0.44, p = 0.06, I 2 = 0%), whereas no trend was observed in the 'poorer' GT subgroup (SMD = 0.21, 95% CI = − 0.12 to 0.54, p = 0.20, I 2 = 0%).
Subgroup analyses by age could not be performed during the early postprandial period due to an insufficient number of studies. Age was related to the effect of GL on delayed episodic memory, but only during the mid-postprandial period (Fig. 6). Specifically, performance was significantly better after a LGL breakfast compared to a HGL breakfast in the 'younger' subgroup (SMD = 0.18, 95% CI = 0.01 -0.35, p = 0.04, I 2 = 0%), but not the 'older' subgroup (SMD = − 0.00, 95% CI = − 0.21 to 0.21, p = 0.99, I 2 = 0%). To check whether this was due to the inclusion of participants with T2DM in the 'older' subgroup, a separate analysis was performed with data from these participants excluded. There was still no significant effect of breakfast GL in the 'older' subgroup (SMD = − 0.08, 95% CI = − 0.33 to 0.17, p = 0.53, I 2 = 0%).

The influence of individual differences.
Individual differences in GT and age were not related to the effect of breakfast GL on accuracy of working memory.

Speed of attention 3.2.3.4.1. The influence of the timing of testing.
For brevity, forest plots, effect sizes, and 95% CI are included in Supplementary Document 1. The Rapid Information Processing Task (RIPT) was used in three studies Benton, 2006a, 2006b;Young and Benton, 2014). This task detects changes in performance over time by measuring attention every minute for five minutes. To reduce the number of data points taken from these studies, data from the first and fifth minute were extracted, and separate analyses were performed for each minute. When one-minute RIPT reaction time scores were included in the analysis (Fig. S2), there was no significant effect of breakfast GL during the early (15 -35 min; p = 0.52), mid (60 -105 min; p = 0.77), or late postprandial period (120 -195 min; p = 0.21). Similarly, when five-minute RIPT reaction time scores were included in the analysis (Fig. S3), there was no effect of breakfast GL during the early (15 -35 min; p = 0.57), mid (60 -105 min; p = 0.64), or late postprandial period (120 -195 min; p = 0.26). Heterogeneity was not substantial for any analysis.

The influence of individual differences.
Individual differences in GT and age were not related to the effect of breakfast GL on speed of attention.
3.2.3.5.2. The influence of individual differences. Individual differences in GT and age were not related to the effect of breakfast GL on accuracy of attention.

Sensitivity analyses.
The results of the leave-one-out analysis are reported in Supplementary Document 1 (Table S5). For brevity, the impact of individually removing studies on significant/trending analyses will only be discussed here. With regards to immediate episodic memory (late postprandial analysis), removal of 13 out of 18 studies reduced the pooled effect size and resulted in the analysis no longer being significant. Removal of Benton et al. (2003) most strongly affected the pooled effect size and significance of the analysis, followed by the better GT data from Young and Benton (2014) and Sanchez-Aguadero et al. (2020). Conversely, removal of some data from Nabb and Benton (2006a) and Lamport et al. (2014) reduced the significance value and increased the pooled effect size (Table S5). With regards to delayed episodic memory (late postprandial analysis), removal of Benton et al. (2003) also produced the largest change in significance and pooled effect size, followed by the better GT data from Young and Benton (2014) and Sanchez-Aguadero et al. (2020).
In most cases, removal of studies that did not match the macronutrient and energy composition of breakfast interventions lowered p values and increased effect sizes (data not reported). For accuracy of attention scores, removal of these studies revealed a trend towards a beneficial effect of a LGL breakfast during the late postprandial period (SMD = 0.23, 95% CI = − 0.01 to 0.48, p = 0.06, I 2 = 0%). Adjusted and unadjusted means were obtained from one study (Sanchez-Aguadero et al., 2020). The inclusion of adjusted or unadjusted means did not influence effect sizes (data not reported).

Publication bias.
Funnel plots were generated for immediate episodic memory, delayed episodic memory, accuracy of attention, and speed of attention (Supplementary Document 1). Due to an insufficient number of studies, we could not generate funnel plots for accuracy of working memory scores. There was some degree of asymmetry for accuracy of attention (early & late postprandial period), speed of attention (early & late postprandial period) and delayed episodic memory (early postprandial period).

Summary of results.
• The influence of the timing of testing: immediate episodic memory was significantly better after a LGL breakfast, compared to a HGL breakfast, but only during the late postprandial period (120 -195 min). There was a similar non-significant trend for delayed episodic memory, whereby performance was better during the late postprandial period (150 -220 min) following a LGL breakfast relative to a HGL breakfast.
• The influence of individual differences in GT: during the late postprandial period, immediate episodic memory was significantly better Note. 1 = GL values reported by study authors, 2 = GL values calculated using reported GI values multiplied by amount of CHO or available CHO, 3 = GL values calculated using published GI tables. * = estimated GL values are not consistent with participants glycaemic responses. BS = between-subjects design, WS = within-subjects design, 00 = baseline, GL = glycaemic load, LGL = low GL, HGL = high GL, MGL = medium GL, GI = glycaemic index, CHO = carbohydrate, PRO = protein, GT = glucose tolerance, BMI = body mass index, BGLs = blood glucose levels, GUM = growing up milk, WLR = word list recall, RT = reaction times, SRT = simple RT, CRT = choice RT, CDR = cognitive drug research, CPT = continuous performance task, SCPT = standard CPT, RMCPT = running memory CPT, CVLT = California Verbal Learning Test, RAVLT = Rey Auditory Verbal Learning Test, WISC = Wechsler Intelligence Scale for Children.

Fig. 2.
Forest plot of the effect of glycaemic load on immediate episodic memory for each postprandial time window. NGT = normal glucose tolerance, T2DM = type 2 diabetes mellitus, IGT = impaired glucose tolerance, HWC = high waist circumference, LWC = low waist circumference, BGT = better glucose tolerance, PGT = poorer glucose tolerance, LC = low carbohydrate, HC = high carbohydrate, LP = low protein, HP = high protein, LF = low fat, and HF = high fat. Note that Nabb and Benton (2006b) manipulated the carbohydrate and fibre content of meals hence LF = low fibre, HF = high fibre, and MF = medium fibre.
following a LGL breakfast, but only in those with 'better' GT. No differences were observed in those with 'poorer' GT. A similar nonsignificant trend was observed for delayed episodic memory, whereby performance was better after a LGL breakfast, in those with 'better' GT, during the late postprandial period. No trend was observed in the 'poorer' GT subgroup.
• The influence of individual differences in age: during the midpostprandial period (62 -119 min), delayed episodic memory was significantly better following a LGL breakfast, but only in the 'younger' subgroup. No differences were observed in the 'older' subgroup.
• There was no effect of breakfast GL on accuracy of working memory, accuracy of attention, or speed of attention.

Studies involving children or adolescents
As mentioned above, data from studies conducted in children were not deemed suitable for meta-analysis. However, children are an important population in which to consider the cognitive consequences of the GL of breakfast. Therefore, we qualitatively reviewed the literature to identify promising avenues for future research as well as factors to be considered in study design.

Risk of bias
The results of the risk of bias assessment are summarised in Supplementary Document 1 (Table S6). Overall, four studies showed a low risk of bias and 12 studies showed some concerns of bias. Studies were generally judged with some concerns of bias for the same reasons reported in Section 3.2.1. For the 'bias arising from period and carryover effects' domain, studies were judged as having some concerns of bias or a high risk of bias if the study was unbalanced and/or participants were tested on consecutive days. For the 'deviations from intended outcome' domain, studies were judged as showing a high risk of bias if participants were clearly not blinded.

Study characteristics
Participants mean age ranged from 5.58 years (Taib et al., 2012) to 15.65 years (Smith and Foster, 2008), and sample sizes ranged from 19 (Benton et al., 2007b) to 84 participants (Anderson et al., 2020). One study was unpublished (Ingwersen, 2011). Two studies used a parallel design, two used a mixed-subjects design, and 12 used a crossover design. All studies were randomised, other than both studies by Mahoney et al. (2005).

Episodic memory.
Twelve studies assessed episodic memory. As shown in Table 2, three studies reported no effect of breakfast GL during the early postprandial period (Lee et al., 2019;Smith and Foster, 2008;Wesnes et al., 2003). In contrast, Ingwersen et al. (2007) reported that episodic memory scores were higher 10 min after consuming a LGL breakfast compared to HGL breakfast.
During the mid-postprandial period, 10 studies reported no effect of breakfast GL (Brindal et al., 2012(Brindal et al., , 2013Ingwersen et al., 2007;Lee et al., 2019;Mahoney et al., 2005;Micha et al., 2011;Taib et al., 2012;Wesnes et al., 2003;Young and Benton, 2015). Smith and Foster (2008) were the only authors to report an effect of breakfast GL within this time window (100 min post-breakfast), whereby delayed episodic memory was better in the HGL group compared to the LGL group. The authors suggested that the beneficial effect of a HGL breakfast on delayed memory may be due to the increased availability of glucose to the brain during encoding. However, there were no significant differences in blood glucose levels between the two groups at any postprandial time point, suggesting that differences in glycaemia may not account for differences in cognitive performance.
Nine studies measured episodic memory during the late postprandial period (120 min or later), of which five reported that breakfast GL did not influence performance (Brindal et al., 2012(Brindal et al., , 2013Lee et al., 2019;Micha et al., 2011;Taib et al., 2012). Using an ANOVA, Benton et al. (2007b) also found that breakfast GL did not influence the memory of young children 140 min post-breakfast. However, there was a significant negative correlation between GL and immediate episodic memory scores. Furthermore, a lower GL predicted better performance, whereas the amount of protein, fat, and carbohydrate did not. Young and Benton (2015) also recruited a sample of children. Although immediate recall was similar 60 min post-breakfast, performance was significantly better during the second test battery (180 min) following the consumption of a LGL breakfast. Furthermore, performance only declined from the first test battery to the second test battery after the consumption of a HGL breakfast. Ingwersen et al. (2007) also found no effect of breakfast GL during the mid-postprandial period (70 min). However, during the late  postprandial period (130 min), task accuracy was better after consuming a LGL rather than HGL breakfast. Similarly, Wesnes et al. (2003) found that a LGL breakfast benefited performance, but only during the late postprandial period. Immediate recall declined by 27% during the final test battery (210 min) after consuming a HGL breakfast but improved by 3 -5% after consuming a LGL breakfast. The ratio of significant to non-significant findings during the early, mid, or late postprandial period are shown in Table 3.
Three studies reported significant effects of breakfast GL that were not time dependent. Taib et al. (2012) reported an overall decline in   6. Forest plot of the effect of glycaemic load on delayed episodic memory during the mid-postprandial period in studies with a mean age above or below 34 years. NGT = normal glucose tolerance, T2DM = type 2 diabetes mellitus, IGT = impaired glucose tolerance, HWC = high waist circumference, LWC = low waist circumference, BGT = better glucose tolerance, PGT = poorer glucose tolerance, LC = low carbohydrate, HC = high carbohydrate, LP = low protein, HP = high protein, LF = low fat, and HF = high fat. Note that Nabb and Benton (2006b) manipulated the carbohydrate and fibre content of meals hence LF = low fibre, HF = high fibre, and MF = medium fibre. delayed memory, across the morning, after the consumption of all drinks other than an isomaltulose-sweetened drink (lowest GL). Interestingly, overall task speed improved across the morning after the consumption of a glucose drink (highest GL) compared to reformulated growing-up milk. In contrast, Wesnes et al. (2003) and Ingwersen et al. (2007) reported that task speed was not influenced by breakfast GL. Brindal et al. (2013) reported that the effect of breakfast GL interacted with participant's biological sex. Relative to baseline, females recalled more words overall after consuming a LGL or MGL breakfast compared to a HGL breakfast. There was no effect of GL in male children. No effect of time was also reported by Lee et al. (2019). However, participants recalled more words overall after consuming a meal with a lower estimated GL (French fries) compared to a higher estimated GL (mashed potatoes or white rice). As glycaemic responses were higher after the consumption of French fries than mashed potatoes, it suggests that estimated GL values are incorrect. Therefore, these findings should be interpreted with caution.
Both studies by Mahoney et al. (2005) found an effect of breakfast GL during the mid-postprandial period. Older (9 -11 years) and younger (6 -8 years) female children performed better 60 min after consuming a LGL breakfast compared to a HGL breakfast. Breakfast GL did not significantly influence the performance of male children. Although no effect of time was reported, Anderson et al. (2020) also reported an interaction between biological sex and breakfast GL. Overall task performance was better, in females, after consuming a HGL drink compared to a LGL drink. The opposite pattern occurred in males, but this was not significant. The reason for these conflicting findings is unclear -participants in Mahoney et al. (2005) and Anderson et al. (2020) were of a similar age and there was a similar difference in GL between the LGL and HGL breakfast interventions.
One study reported that a HGI breakfast predicted better working memory performance during the late postprandial period (Micha et al., 2011). Two studies reported that a LGL breakfast was associated with better performance during the late postprandial period (Cooper et al., 2012(Cooper et al., , 2015. Although reaction times were similar 30 min post-breakfast, there was a greater improvement in task speed 120 min after consuming a LGL breakfast compared to a HGL breakfast (Cooper et al., 2012). On the complex levels of this task, accuracy scores were maintained across the morning following a LGL breakfast but declined following a HGL breakfast. In a later study, Cooper et al. (2015) reported a similar finding, whereby reaction times improved across the morning (from 30 to 120 min) after consuming a LGL meal, regardless of whether participants exercised. However, performance only improved after a HGL meal if participants exercised. Taib et al. (2012) reported that a LGL breakfast benefitted overall task performance. Relative to baseline, numeric working memory scores declined across the morning in all drink conditions. However, the decline was significantly smaller after consuming an isomaltulose-sweetened drink (lowest GL) compared to a glucose drink (highest GL) or reformulated growing-up milk. As shown in Table 2, the GL of the isomaltulose-sweetened milk and reformulated milk were very similar and so it is unlikely that differences in GL accounted for this finding. The authors also reported that a HGL drink benefitted spatial working memory. Specifically, the overall decline across the morning was significantly smaller after the consumption of glucose compared to all three LGL drinks.
Two studies reported significant effects of breakfast GL that were not time dependent. Taib et al. (2012) reported that performance declined across the morning. However, the overall decline was significantly greater following the consumption of standard growing-up milk compared to isomaltulose-sweetened growing-up milk. Furthermore, at 180 min, there was a non-significant improvement in performance, but only after the consumption of isomaltulose-sweetened milk. Although there was no specific effect of time, Anderson et al. (2020) reported an interaction between breakfast GL and biological sex. Responses were faster overall after the consumption of a LGL drink, compared to a HGL drink, but only in female participants. The opposite pattern emerged for males, but this was non-significant.
During the mid-postprandial period, Mahoney et al. (2005) reported an interaction between age and breakfast GL. Younger children (6 -8 years) made more correct responses on an auditory attention task 60 min after consuming a LGL breakfast rather than a HGL breakfast. In contrast, no differences in performance were found in older children (9 -11 years). Furthermore, breakfast GL did not influence visual attention performance. Ingwersen et al. (2007) also examined the effect of age Table 3 The ratio of significant to non-significant findings reported by studies involving children and adolescents. using the same age groups, however no age effects were observed. Three studies found a beneficial effect of a LGL breakfast during the late postprandial period. In Cooper et al. (2012), accuracy scores on the more complex trials were similar during the first test battery (30 min). However, during the second test battery (120 min), scores were better maintained after consuming a LGL breakfast compared to a HGL breakfast. Similarly, Ingwersen et al. (2007) reported that there was a sharp decline in performance 130 min after consuming a HGL breakfast, whereas performance was maintained at this time after consuming a LGL breakfast. Although no effect of breakfast GL was reported using an ANOVA, Benton et al. (2007b) found that the number of lapses of attention (difficult trials only) correlated positively with GL. A lower GL also predicted better sustained attention (difficult trials only).

Summary of results
• There was very limited evidence to suggest that breakfast GL influenced cognitive performance within the first hour after consuming breakfast. • There was some evidence to suggest that a LGL breakfast may protect against a decline in episodic memory, accuracy of attention, and speed of attention during the late postprandial period (120 min postbreakfast or later). • Task difficulty, age, and biological sex might influence the relationship between breakfast GL and cognitive performance.

Adult meta-analysis
Studies comparing HGL and LGL breakfasts have produced mixed results. An obvious hypothesis is that certain methodological factors determine the outcome. Therefore, for the first time, this meta-analysis considered a range of possible factors that may influence the response to the glycaemic influence of breakfast.
As summarised in Section 2.4.8, there were several lines of evidence to suggest that a LGL breakfast benefits cognitive performance in a timedependent manner. During the late postprandial period (120 min or later), a LGL breakfast was significantly associated with better immediate episodic memory (Fig. 2). A similar non-significant trend was also observed for delayed episodic memory during the late postprandial period (150 − 220 min; Fig. 4). In addition, the beneficial effect of a LGL breakfast on immediate and delayed episodic memory was influenced by individual differences in age (Fig. 6) and GT (Figs. 3 & 5). However, the sensitivity analysis showed that the removal of most studies, particularly Benton et al. (2003), Benton (2014), andSanchez-Aguadero et al. (2020), reduced the effect size of the analyses shown in Figs. 2 and 4, highlighting the need for more research in this area to confirm or refute these conclusions.
There was no effect of breakfast GL on speed of attention, accuracy of attention, or accuracy of working memory across all three postprandial time windows. These results suggest that the effect of breakfast GL may be subdomain-specific. In line with this, previous reviews have reported that episodic memory is more responsive to breakfast manipulations than other cognitive domains and subdomains (Galioto and Spitznagel, 2016;Hoyland et al., 2008;Smith et al., 2011;Wasyluk et al., 2019). However, it is important to note that fewer studies assessed working memory than episodic memory in the present review. In addition, a wider variety of tests were used to measure attention and working memory, potentially influencing the findings.
Nonetheless, these findings are consistent with the hypothesis that a LGL breakfast, which provides a steady and continuous supply of glucose to the brain, may be more advantageous for acute cognitive performance than a HGL breakfast (Nilsson et al., 2012;Papanikolaou et al., 2006;Young and Benton, 2014). It is difficult to assess the validity of this hypothesis because many studies did not measure peripheral glucose levels. Furthermore, in those studies that did, some reported cognitive differences when glucose levels were similar (Benton et al., 2003;Nilsson et al., 2012), and others reported no cognitive differences when glucose levels were markedly different (Kaplan et al., 2000;Lamport et al., 2013a;Sanchez-Aguadero et al., 2020). Although there is a positive correlation between blood glucose and extracellular glucose (Rostami and Bellander, 2011;van de Ven et al., 2012), the concentration of glucose in the brain is approximately 20-30% of that in the blood (Béland-Millar et al., 2017), and there is a time lag of up to 30 min between changes in blood glucose and changes in extracellular glucose (Abi-Saab et al., 2002;Gruetter et al., 1998). Cognitive demand can also influence extracellular glucose levels (McNay et al., 2000). Therefore, a failure to observe concurrent cognitive and glycaemic differences does not necessarily disprove this hypothesis. Perhaps the beneficial effect of a LGL relative to HGL breakfast is not directly related to glycaemia, but rather associated aspects of metabolism that appear over time, including changes to concentrations of cortisol, insulin, glucagon, glucagon-like pepetide-1, acetylcholine, glutamate, or serotonin (Adolphus et al., 2016;Hoyland et al., 2009).
Previous reviews have suggested that a LGL breakfast may be particularly advantageous to vulnerable groups, including older adults or individuals with poorer glucoregulation (Galioto and Spitznagel, 2016;Lamport et al., 2009;Sünram-Lea and Owen, 2017). In contrast, subgroup analyses indicated that a LGL breakfast may exert a greater effect in younger adults or those with 'better' GT. Specifically, a LGL breakfast benefitted immediate episodic memory, during the late postprandial period (120 min or later), in the 'better' but not the 'poorer' GT subgroup. A similar non-significant trend was observed for delayed episodic memory during the late postprandial period. Subgroup analyses also showed that a LGL breakfast benefitted delayed episodic memory, during the mid-postprandial period (62 -119 min), in the 'younger' subgroup but not the 'older' subgroup. The finding that age effects were only observed during the mid-postprandial period is surprising, as all other effects were observed during the late postprandial period.
As the mechanisms underlying the acute cognitive effects of GL are currently unclear, it is difficult to speculate why these subgroup effects occurred. However, it is plausible that such mechanisms are hampered in those with 'poorer' GT or those aged above 35 years old. For example, it has been suggested that the beneficial effect of a LGL breakfast may be due to the generation of a smoother postprandial insulin profile (Benton et al., 2003). Glucose intolerance is associated with impaired insulin action and secretion (Abdul-Ghani et al., 2006) and endothelial dysfunction (Convit, 2005). The transport of insulin and glucose across the blood brain barrier, and between intracellular and extracellular fluid in the brain, is thus hindered in glucose intolerant individuals (Lamport et al., 2013a;Young and Benton, 2014). This may result in an insensitivity to the cognitive effects of GL. The prevalence of glucose intolerance and endothelial dysfunction also increases with age, which may also contribute to age effects.
Consistent with this suggestion, a double-blind, well-controlled study by Young and Benton (2014) reported that a LGL breakfast benefitted episodic and working memory in middle aged and older adults (45 -80 years old) with 'better' but not 'poorer' GT. The picture is, however, complicated as a beneficial effect of a LGL breakfast on episodic memory has been demonstrated in older adults with T2DM (Papanikolaou et al., 2006) and IGT (Lamport et al., 2014), and young healthy adults with 'poorer' GT (Nabb and Benton, 2006b). However, it is unclear whether GT interacted with the effects of breakfast GL as Papanikolaou et al. (2006) did not recruit a healthy control group. Furthermore, most participants in Papanikolaou et al. (2006) were treated with metformin or sulphonylureas. These medications improve GT by enhancing insulin secretion and suppressing hepatic glucose production, which may have interacted with the effects of breakfast GL.
Alternatively, the absence of a significant beneficial effect in the 'older' subgroup may be due to the inclusion of participants with a wide range of ages. Specifically, the mean age of the 'older' subgroup ranged from 36.6 years (Lamport et al., 2014) to 65 years (Papanikolaou et al., 2006), whereas the mean age of the 'younger' subgroup ranged from 20.36 years (Nabb and Benton, 2006a) to 28.1 years (Sanchez-Aguadero et al., 2020). It is plausible that the wide age range of the 'older' subgroup introduced variability, which can reduce statistical power and mask significant effects (Jiang et al., 2010;Netz et al., 2019). For example, advancing age is associated with increased interindividual differences in baseline nutritional status, GT, and cognitive, physical, and sensory function (Asamane et al., 2020;Ferrucci and Kuchel, 2021;Mungas et al., 2010;Zulman et al., 2011). Further research aimed at understanding individual differences in the response to GL is clearly warranted.

Child and adolescent qualitative analysis
Due to a lack of data, and the use of different types of scores, a metaanalysis of the effect of breakfast GL in children and adolescents was not possible. Instead, we performed a systematic review of 16 studies. Although the cognitive effects of manipulating breakfast GL have been discussed in previous systematic reviews, the influence of the timing of testing was only briefly considered (Adolphus et al., 2016;Á lvarez-Bueno et al., 2019;Hoyland et al., 2009). In contrast, we systematically examined the effect of breakfast GL in relation to the timing of testing.
There was no clear and robust effect of breakfast GL on episodic memory, working memory, and attention during the early, mid, or late postprandial period. There was some evidence to suggest that a LGL breakfast benefited episodic memory, particularly during the late postprandial period (120 min post-breakfast or later), which is consistent with the results of the present meta-analysis. No study reported an effect of breakfast GL on accuracy and speed of attention during the early postprandial period, and no study reported that a HGL breakfast positively influenced attention. The effect of breakfast GL on working memory was more heterogeneous, with some studies reporting an advantage of a LGL breakfast (Cooper et al., 2012(Cooper et al., , 2015Mahoney et al., 2005;Taib et al., 2012), and other studies reporting an advantage of a HGL breakfast (Anderson et al., 2020;Micha et al., 2011;Taib et al., 2012).
A recent meta-analysis assessed the effect of breakfast GI, rather than GL, on the cognitive performance of children and adolescents. Á lvarez- Bueno et al. (2019) reported that breakfast GI did not influence immediate memory, delayed memory, and attention. However, subgroup analyses revealed that delayed memory scores were significantly higher after a LGI breakfast, compared to HGI breakfast, in children but not adolescents. Á lvarez- Bueno et al. (2019) were able to perform a meta-analysis of the effect of breakfast GI because they took a less rigorous approach to the meta-analysis than we aimed to perform. For example, change scores and post-scores were analysed together, subdomains of memory were analysed as one group (e.g., working memory & episodic memory), and the effect of the timing of testing was not examined. It is unclear how these factors influenced their conclusions, however, our qualitative synthesis indicated that they might be important.
Overall, there was some evidence to suggest that a LGL breakfast exerted a positive effect on episodic memory and accuracy of attention during the late postprandial period. This pattern of results is illustrated in Table 3. However, the quality of evidence was mixed, with most studies showing some concerns of bias. Furthermore, a range of experimental methods were used, giving rise to different methodological limitations. For example, two studies were not randomised, and most studies administered breakfast interventions that were not matched for macronutrient or energy content. A detailed discussion of these methodological limitations can be found in Section 4.3.

Guiding principles
This review highlighted considerable methodological variability between studies that may have contributed to the inconsistent literature. Consequently, a series of guiding principles were developed to guide study design and, in turn, facilitate a better understanding of the relationship between breakfast GL and acute cognitive performance (Fig. 7).

Sample heterogeneity
Research has repeatedly shown that participants respond in different ways to the same nutritional interventions (Lampe et al., 2013). As such, an analysis of the average cognitive response to variations in breakfast GL may fail to reveal the range of responses produced (Blundell et al., 2010). Although there are many other relevant factors, this section discusses the importance of considering age, biological sex, and body mass index (BMI), whereas GT is discussed separately in Section 4.3.2.
The meta-analysis revealed that age influenced the beneficial effect of a LGL breakfast on delayed episodic memory in adults. Therefore, null findings in studies with large age ranges may reflect a failure to test for age effects rather than no effect of breakfast GL. Similarly, some studies analysed children and adolescents as one group (Cooper et al., 2015;Wesnes et al., 2003). However, children and adolescents should ideally be analysed separately given the abundance of metabolic, behavioural, and hormonal changes associated with puberty (Kawakubo et al., 2011). The rate of cerebral glucose utilisation is also higher in young children (4 -10 years) compared to adolescents (Chugani, 1998), potentially making young children more susceptible to changes in postprandial glycaemia. Indeed, Á lvarez- Bueno et al. (2019) reported that delayed episodic memory scores were significantly higher after a LGI breakfast, compared to HGI breakfast, in children but not adolescents. More work is clearly needed to determine the influence of age, ideally by directly comparing different age groups within the same study.
As shown in Table 1, few studies assessed whether measures of obesity influenced the relationship between breakfast GL and cognitive function in adults. This is surprising given the strong association between obesity, cognitive dysfunction, and GT throughout adulthood (Fellows and Schmitter-Edgecombe, 2018). Although a larger number of studies involving children and adolescents included BMI as a covariate, there was no evidence to suggest that BMI moderated the effect of breakfast GL on cognitive performance (see Table 2). However, other than Anderson et al. (2020), participants BMIs were within the normal, healthy range. As glycaemic responses to HGI meals are greater in overweight adolescents and adults, compared to normal weight individuals (Yalçın et al., 2017;Zakrzewski and Tolfrey, 2012), it is possible that BMI or other measures of obesity exert a moderating effect on the relationship between GL and cognition.
We were unable to perform subgroup analyses according to biological sex. However, several studies have reported that female children (ranging from 6 to 12 years of age) were more susceptible to variations in breakfast GL than male children (Anderson et al., 2020;Brindal et al., 2013;Mahoney et al., 2005). This finding may be due to sex differences in insulin sensitivity and cortisol levels, both of which have been suggested to underlie the effect of GL on acute cognitive performance (Cooper et al., 2012;Lamport et al., 2013a). In adults, no study reported that sex moderated the cognitive effects of breakfast GL. It has previously been reported that the consumption of glucose enhanced episodic memory in older men but not older women (Craft et al., 1994). Therefore, further exploration of this factor is warranted.

Individual differences in glucose tolerance
The importance of investigating the influence of GT was highlighted in the present meta-analysis. Studies have also reported that GT moderated the effect of breakfast GL on attention, inhibitory control, working memory, and visuospatial memory (Anderson et al., 2018;Lamport et al., 2014;Benton, 2006a, 2006b;Young and Benton, 2014). However, despite its obvious importance, many studies did not measure GT. In those studies that did, some defined 'poorer' GT using the WHO criteria for IGT or T2DM (Lamport et al., 2014(Lamport et al., , 2013a, whereas others used the median split of the sample (Brindal et al., 2013;Nilsson et al., 2009;van der Zwaluw et al., 2014). Although the latter approach can provide valuable information, definitions of 'poorer' or 'better' GT are sample dependent which limits comparisons between studies. As such, researchers may benefit from using criteria in line with the WHO for impaired fasting glucose, IGT, or T2DM.
Various measures of GT were also used including fasting glucose, 2-h glucose, or change in glucose levels from baseline to 30 min. To categorise participants as having 'poorer' or 'better' GT, the present metaanalysis used the following definition: fasting glucose below or above 6.1 mmol/L or 2-h glucose below or above 7 mmol/L. Due to an insufficient number of studies, we were unable to analyse the influence of fasting glucose and 2-h glucose separately. Although these measures are related, they reflect different aspects of metabolism and therefore it is questionable whether such measures should have been combined (Meyer et al., 2006). For example, both elevated fasting and 2-h glucose levels are associated with an increased risk of T2DM and cardiovascular disease (de Vegt et al., 2001). However, elevated fasting glucose levels are primarily due to hepatic insulin resistance and impaired basal insulin secretion and first-phase insulin release, whereas elevated 2-h glucose levels are primarily due to muscle insulin resistance and impaired first-and second-phase insulin release (Meyer et al., 2006).
It is yet to be established whether certain measures of GT exert a stronger moderating effect on the relationship between GL and cognition than others. A study by Owen et al. (2013) reported that fasting glucose levels moderated the glucose facilitation effect whereas 2-hour glucose levels did not. Future research would therefore benefit from comparing the moderating effect of different indices of GT. The influence of the susceptibility to postprandial hypoglycaemia remains relatively unexplored (Young and Benton, 2014), but provides an interesting avenue for future work. Given that the demands of a cognitive task influence postprandial blood glucose levels (Donohoe and Benton, 1999;Scholey et al., 2001), studies that utilise postprandial measures of GT may benefit from measuring glycaemia, via an OGTT, on a separate day to cognitive test sessions.
The findings from a recent study by Anderson et al. (2018) also suggest that research may benefit from analysing GT as a continuous, rather than dichotomous, variable. Anderson et al. (2018) compared the cognitive effects of a breakfast drink containing milk or apple juice in young adults. Using a linear mixed model, the authors identified specific fasting glucose levels where cognitive differences between breakfast conditions were observable. Importantly, these levels varied depending on the nature of the task (working memory, attention, or inhibitory control) and outcome measure (speed vs. accuracy), indicating that the domain specific response to GL might reflect a variability in the threshold at which specific domains are impacted.

Selection of tests
Apart from episodic memory, a range of cognitive tests have been used to measure the same cognitive subdomain. This observation suggests that tests may have been selected due to convenience rather than their sensitivity to previous nutritional interventions (Adolphus et al., 2021). As such, a null finding may be due to a lack of test sensitivity rather than a lack of effect of breakfast GL. For example, in order to measure attention, Ingwersen et al. (2007) created a composite score using reaction time and digit vigilance scores, whereas Ingwersen (2011) administered a Continuous Attention Task. Although the same breakfasts were administered in both studies, and children were of a similar age, only Ingwersen et al. (2007) reported that breakfast GL significantly influenced task performance. The authors suggested that the discrepant findings may be due to differences in cognitive demand and hence task sensitivity.
Future studies would benefit from using standardised, validated tests that are known to be sensitive to the subtle, but important, effects of nutritional interventions (Adolphus et al., 2017). Based on the results of this meta-analysis, word list recall tasks are sensitive to the effect of breakfast GL on episodic memory in adults. Both immediate and delayed episodic memory should be measured, and words matched for the number of syllables, the frequency with which they occur in English, the number of abstract and concrete words, and image-ability (Young and Benton, 2014). Due to the wide range of tests used to measure other cognitive domains and subdomains, it is difficult to state whether a test is sensitive to the effects of breakfast GL. Although performance on the Serial Sevens task was significantly influenced by variations in breakfast GL (Dye et al., 2010;Micha et al., 2011;Young and Benton, 2014), this task should be avoided due to its reliance on mathematical skill and the likelihood of significant practise effects (Karzmark, 2000).
The results of a recent systematic review by Peters et al. (2020) may inform task selection. The authors found that the medial temporal and frontal lobes and networks, which underpin episodic memory and attentional processes, may be preferentially affected by carbohydrate consumption. Despite the sensitivity of episodic memory to glycaemic manipulations, as evidenced in the present meta-analysis, many studies did not measure this subdomain. It would be useful if future work continued to assess episodic memory so that firm conclusions could be made regarding the conditions that elicit beneficial effects of LGL or HGL breakfasts.
Studies would also benefit from considering the influence of practice effects, which are a particular issue for tests involving memory and learning (Bartels et al., 2010). Practice effects tend to occur mostly between the first and second administration of a test (Bell et al., 2018). Using parallel versions of the same test, the influence of practice effects can therefore be reduced by incorporating a separate test familiarisation visit and a brief practice session immediately before testing begins (Bell et al., 2018). Practice sessions can also minimise the negative impact of stress and anxiety, due to a lack of task familiarity, on performance. However, it is important that the length of a practice session is appropriate so that fatigue effects do not impact performance (Süss and Schmiedek, 2000).
The final factor that needs consideration is the nature of outcome measures. Some studies only measured task speed (Deng et al., 2021;Micha et al., 2011). However, faster performance does not necessarily equate to better performance as speed may increase at the expense of accuracy, which is suggestive of an impulsive response style (Schmitt et al., 2005). To avoid misleading conclusions, measures of both speed and accuracy should be reported where possible, especially as studies reported that variations in breakfast GL influenced one outcome measure but not the other (Cooper et al., 2012(Cooper et al., , 2015Ingwersen et al., 2007;Nilsson et al., 2012;Wesnes et al., 2003).

Manipulation and measurement of GL
A major limitation of the literature to date, and hence the present meta-analysis, is that most breakfast interventions differed not only in terms of GL but also macronutrient and energy composition. This limits the extent to which findings can be attributed to differences in GL, as the provision of different amounts of energy, fat, protein, and carbohydrate can also impact cognitive performance (Fischer et al., 2002;Kaplan et al., 2001). The importance of matching the macronutrient and energy content of meals was highlighted in Section 3.2.4.6, whereby inclusion of studies that matched the macronutrient and energy content of meals revealed a trend towards a beneficial effect of a LGL breakfast on accuracy of attention scores during the late postprandial period. Similarly, the polyphenol, vitamin, and mineral content of breakfast interventions often differed, all of which can acutely modulate cognitive performance (Huskisson et al., 2007;Philip et al., 2019). To gain a better understanding of the impact of breakfast GL on cognitive performance, future research would benefit from matching the macronutrient and micronutrient content of meals or drinks. Studies have accomplished this by sweetening the same meal or beverage with different types of sugar (Deng et al., 2021;Dye et al., 2010;Benton, 2014, 2015) or by administering two types of rice varying in GL (Jansen et al., 2020). Manipulating GL using these methods would also allow studies to be blinded.
Another factor that needs consideration is the GL of breakfast interventions. As the GL of meals and drinks varied considerably between studies, standard GL thresholds could not be applied (i.e., LGL = below 10, MGL = 10 -20, or HGL = above 20). Instead, within each study, the two meals or drinks with the largest difference in GL were categorised as the HGL and LGL breakfast interventions. This resulted in large variability in the GL of LGL breakfast interventions, which ranged from 1.56 (Ginieis et al., 2018) to 50 (van der Zwaluw et al., 2014), and HGL breakfast interventions, which ranged from 11.3 (Anderson et al., 2018(Anderson et al., , 2021 to 71 (Lamport et al., 2014(Lamport et al., , 2013a. To facilitate more precise comparisons in future, research would benefit from administering breakfast interventions that fall within the thresholds stated above. Dose-response studies could also be conducted using a range of glycaemic loads to determine the optimal GL of breakfast. The difference in GL between LGL and HGL breakfast interventions also varied considerably. The smallest difference in GL was six (Benton et al., 2003) and the largest difference was 59 (Lamport et al., 2014(Lamport et al., , 2013a. It is unclear what impact this has on the relationship between breakfast GL and cognitive performance as significant effects were reported after consuming breakfasts with smaller and larger differences in GL. Due to a limited number of studies, we were unable to determine whether differences in GL influenced the results. However, there was some evidence to suggest that when the GL of breakfast was manipulated using different types of sugars, studies tended to demonstrate significant effects when there was a larger difference in GL. That is, when glucose was compared with isomaltulose or fructose (Ginieis et al., 2018;Taib et al., 2012;Benton, 2014, 2015), rather than when sucrose was compared with isomaltulose (Deng et al., 2021;Dye et al., 2010). However, this suggestion requires further investigation.
Many studies estimated the GL/GI of meals using published values. We also used this method when required. However, there are several issues with using published values. For example, depending on the type of food consumed, published values can overestimate the actual GI of a food by 22-55% (Dodd et al., 2011). The GI of the same two foods can vary depending on the degree of ripeness, country of origin, variety, or cooking/storage methods (Aston et al., 2008;Henry et al., 2005). Furthermore, it is questionable whether published values, calculated using adult samples, are applicable to children and adolescents. Although it is more costly, future studies would benefit from calculating GL values directly, on a separate day to cognitive test sessions. This would enable researchers to verify beforehand that HGL and LGL breakfast interventions produce significant differences in postprandial glycaemia, an issue that occurred in some studies (Smith and Foster, 2008;van der Zwaluw et al., 2014).
Lastly, the use of dairy products needs to be reconsidered. Dairy products are insulinotropic hence the addition of dairy to meals can shorten postprandial glucose profiles and produce lower GL values than anticipated (Blaak et al., 2012). This is problematic when the objective of a study is to compare the cognitive consequences of different glycaemic profiles. It is plausible that by using a dairy-based vehicle, the difference in postprandial blood glucose levels between two meals is reduced to the point where it is no longer cognitively relevant (Blaak et al., 2012). We intended to perform subgroup analyses to determine whether the amount of dairy used influenced the findings, however an insufficient number of studies were available. Nonetheless, dairy products should be avoided where possible.

Timescales
Cognitive performance was measured at various times throughout the postprandial period, a factor that may have played a key role in producing conflicting literature. The meta-analysis revealed multiple lines of evidence to suggest that the beneficial effect of a LGL breakfast in adults emerges during the mid-postprandial period (60 -119 min) and, in particular, the late postprandial period (120 min or later). Studies in children and adolescents have reported similar findings, whereby a significant beneficial effect of a LGL breakfast typically occurred between 120 and 210 min post-breakfast consumption (Cooper et al., 2012(Cooper et al., , 2015Ingwersen et al., 2007;Wesnes et al., 2003;Young and Benton, 2015). These findings suggest that a LGL breakfast may attenuate decrements in cognitive performance across the morning. As such, future studies would benefit from assessing cognitive performance at multiple time points, especially during the late postprandial period (120 min or later).
In many cases, participants did not undergo cognitive testing when the difference in blood glucose levels between breakfast conditions were greatest, times that are more likely to be cognitively relevant. For example, Deng et al. (2021) administered a cognitive test battery 60 min post-breakfast. Blood glucose levels after the HGL (GL = 32) and LGL (GL = 16) drink were almost identical at 60 min, possibly contributing to the lack of significant results. If GL values are calculated directly, on separate days to cognitive test sessions, then this information could be used to select the most appropriate time points to administer test batteries. This information would be particularly useful for researchers that choose to administer one test battery to reduce participant burden or school disruption. Lastly, the majority of studies included in this review did not measure baseline cognitive performance. However, one study reported that the effect of breakfast GL was influenced by baseline cognitive performance (Dye et al., 2010), suggesting that effects may have been masked in studies that did not statistically account for baseline performance. Measuring baseline performance would also assist future meta-analyses, as change scores can be calculated by review authors when necessary.

Condition of participants and controlling for confounding factors
In order to attribute differences in cognitive performance to difference in GL, it is critical that the influence of factors that might distort true effects are excluded or minimised (Schmitt et al., 2005). As shown in Tables 1 and 2, a phenomenon that was rarely considered is the second meal cognitive effect, whereby the composition of one meal influences glycaemic and cognitive responses to a subsequent meal. For example, Lamport et al. (2013b) reported that the consumption of a HGL evening meal, compared to a LGL evening meal, was associated with better episodic memory following the consumption of a HGL breakfast. An overnight fast may be insufficient to control for the potential confounding effect of an evening meal. This also applies to evening alcohol consumption, which has been shown to interact with the effect of breakfast GL on memory in adults (Benton and Nabb, 2004). Similarly, despite evidence showing that exercise can reduce mean 24-hour glucose levels (Munan et al., 2020), few studies standardised participants 24-hour physical activity levels. Given these findings, it is critical that studies instruct participants to fast overnight, consume standardised meals the day before testing, and avoid alcohol and vigorous exercise in the 24 h preceding testing.
Additional factors that can influence glucose metabolism and/or cognitive performance include mood, stress, illness, fatigue, hydration, hunger, quality of sleep, and motivation (Micha et al., 2010;Schmitt et al., 2005). This is especially the case for young children (Isaacs and Oates, 2008). Differences in GL may influence cognitive performance indirectly via some of these factors, hence it is important that studies consider their influence (Adolphus et al., 2016). In addition, it is likely that many participants included in the present meta-analysis underwent cognitive testing in a state of caffeine and/or nicotine withdrawal. This is less of a problem if a crossover design is used as the effect of withdrawal on cognitive performance is constant across conditions. Nonetheless, nicotine and excessive caffeine use should form part of the exclusion criteria.
The type of study design used should also be carefully considered. One of the main advantages of using a crossover design is that breakfast interventions are evaluated in the same group of participants thus reducing the confounding effect of between-person variability (Harris and Raynor, 2017). However, results can be complicated by order effects -for example, despite including a one-week washout period, three studies reported that cognitive effects only occurred when breakfast interventions were consumed in a specific order (Nilsson et al., 2009(Nilsson et al., , 2012Young and Benton, 2015). Furthermore, if breakfast conditions are not blinded, participants pre-existing knowledge or beliefs about the cognitive effects of a specific food (e.g., sugary cereal) may impact task performance (Adolphus et al., 2016).
In contrast, parallel designs reduce the risk of expectancy, fatigue, and order effects but increase the risk of between-subject variability distorting results. It is common practice for researchers to statistically test for baseline differences between groups. However, a covariate can be balanced between groups, according to a non-significant p value, but still exert a meaningful influence on the relationship between a treatment and outcome (Knol et al., 2012;Peterson et al., 2017). This is especially true for studies with small samples, often leading to the omission of important covariates (De Boer et al., 2015). Ideally, factors that are known to influence the cognitive effects of nutritional manipulations (e.g., baseline cognitive performance or socioeconomic status) should be identified a priori and incorporated into statistical models as covariates (Peterson et al., 2017). By adjusting for known covariates, whether significant or not, the effect estimate will be more precise and closer to the 'true' effect (De Boer et al., 2015).

Data availability and transparency
For this area of research to evolve, there needs to be more transparency and consistency when reporting results. Indeed, limited data availability and the use of different scores (i.e., post scores or change scores) prevented a meta-analysis of child and adolescent studies. It would therefore be beneficial if future studies reported means, effect sizes, and SD in a table and, ideally, raw datasets as supplementary material. This would facilitate a more accurate and robust synthesis of the literature, as well as a better understanding of the effect of time and other potential moderating factors (e.g., the difference in GL between meals). Future work may also benefit from providing more detailed information about, for example, the method of randomisation or the number of participant dropouts. This would ensure that a study is correctly classified as high or low quality.

Limitations
The findings are limited by a very low to low certainty of evidence. Studies were highly heterogeneous in terms of participant characteristics, the type of statistical methods employed, pre-test conditions, the composition of breakfast, sample size, and the type of cognitive test administered. The risk of bias assessment indicated that, overall, no study showed a high risk of bias, however 24 out of 33 studies showed some concerns of bias which were mainly related to the randomisation process, blinding of participants and researchers, and study protocol pre-registration.
There was some degree of overlap between the 'younger' and 'older' subgroups (below or above 35 years of age). For example, Sanchez-Aguadero et al. (2020) sample had a mean age of 28.1 years but an age range of 20-40 years. This should be considered when interpreting the results. Subgroup analyses according to GT status were performed by categorising participants as having 'poorer' GT if fasting glucose levels were above 6.1 mmol/L and/or two-hour glucose levels were above 7 mmol/L. This definition was chosen because it is clinically relevant (Petersen and McGuire, 2005) and did not markedly increase the number of studies excluded from the subgroup analyses. However, these measures reflect different aspects of metabolism (Meyer et al., 2006), therefore it is questionable whether they should have been combined. As the area evolves, subsequent meta-analysis should consider the moderating effect of postprandial glucose and fasting glucose separately. Lastly, we analysed measures of selective attention and sustained attention together. However, attention is not a unitary construct, therefore as the number of studies increase, future research might consider these aspects of attention separately.

Recommendations for future studies
To date, the considerable difference in experimental design between studies has limited the drawing of conclusions. It is suggested that factors listed above are all taken into account when designing studies. There are several other avenues for research. For example, few studies have investigated the cognitive effects of manipulating the GL of a lunchtime or evening meal, or the second meal cognitive effect. Future work could also place a greater emphasis on understanding the neurobiological mechanisms underlying the effect of breakfast GL on acute cognition. This could be achieved by using neuroimaging methods, which have not yet been applied to this area of research. Similarly, various biomarkers that are related to acute cognitive performance and/ or postprandial glycaemia could be measured, such as heart rate variability, insulin, glucagon, glucagon-like peptide-1, free fatty acids, or cortisol (Dybjer et al., 2020;Nilsson et al., 2008;Saito et al., 2018). Lastly, several studies have reported that the gut microbiome plays a key role in postprandial glycaemia (Berry et al., 2020;Mendes-Soares et al., 2019;Zeevi et al., 2015), and as such may influence the cognitive response to variations in breakfast GL. This could be investigated by administering different dietary fibres.

Conclusions
In conclusion, the meta-analysis revealed that the consumption of a LGL breakfast, rather than a HGL breakfast, was associated with better episodic memory during the late postprandial period in adults. Furthermore, the relationship between breakfast GL and episodic memory was influenced by individual differences in GT and age. A review of studies involving children and adolescents also suggested that a LGL breakfast may benefit episodic memory and attention during the late postprandial period.
Although there are many reports of a significant influence of the glycaemic nature of breakfast, these are not consistent. As such, it would be premature to suggest that public health guidelines recommend consuming a LGL breakfast to improve acute cognitive performance. Using the guiding principles discussed above, more comparable studies are needed in order to establish the critical variables that lead to a beneficial response. Such findings would have far reaching implications for public health policy and school breakfast programmes. The present review found that the nature of the task used, the timing of testing, population age, and individual differences in GT are relevant. No doubt there are other factors that should be considered, which will emerge as the field progresses.
Funding this research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability
the datasets generated and/or analysed during the present review are available from the corresponding author on reasonable request.

Declaration of interest
none.