Trial identification

A previous systematic review provided studies carried out before 2009 for the present review⁽ Reference Baldrick, Woodside and Elborn ^{14 )}. Updated searches were conducted (by L. J.) from April 2009 (last search date of previous systematic review⁽ Reference Baldrick, Woodside and Elborn ^{14 )}) to December 2013 in PubMed, EMBASE, CENTRAL, CINAHL and Web of Science using terms related to fruits and vegetables, dietary intervention studies and biomarkers (see online Supplementary information for detailed search strategy). Relevant systematic reviews were obtained and their reference lists were examined for additional references. Citations were screened by one reviewer (M. P. or L. J.) and hard copies of the relevant articles were obtained. These were screened by one reviewer (M. P.) and checked for inclusion by a second reviewer (L. J.).

Inclusion and exclusion criteria

RCT of different amounts of fruit and vegetable intake (where some food intake was observed or provided) with outcomes of plasma or serum vitamin C or carotenoids were included in the review. Interventions of any duration were considered for inclusion. Trials altering other aspects of the diet, in addition to fruit and vegetable intake – for example, low-fat diets – were excluded to avoid the possibility that changes in blood-based biomarkers may be a result of dietary changes other than fruit and vegetables. Intervention studies of a single fruit or vegetable were excluded. Findings from these types of interventions may underestimate the utility of biomarkers for measures of general fruit and vegetable intake as the single food source may contain a more limited range of nutrients. Trials where fruit and vegetable intake was encouraged through dietary advice were excluded as adherence to the advice is harder to estimate. Trials in healthy or unhealthy populations were included, including populations with high CVD risk factors or impaired glucose metabolism. However, trials in populations with abnormalities in micronutrient metabolism or vitamin-deficient populations were excluded. Trials were included if they reported biomarker measurements, either as changes from baseline or as baseline and post-intervention values, and if information was available on the amount of fruits and vegetables consumed in each intervention group.

Data extraction

Data on trial and population characteristics and outcomes were extracted into an Excel form that was piloted on a sample of trials before use (by M. P., M. S., L. J. and C. M.). Data extracted on trial characteristics included the type of trial (parallel or cross-over), duration of intervention, information on the duration of pre-intervention and within-intervention washout periods, the amount and types of fruits and vegetables consumed and the mode of administration (some meals eaten under supervision v. all meals at home), smoking status, fasting status at the time of biomarker measurement, the use of dietary supplements, inclusion and exclusion criteria and funding sources. Population characteristics included the sample size, country and type of sample – for example, clinical or general population – and participant demographics including age, sex and ethnicity were also extracted. Whenever available, data on baseline, post-treatment and change in biomarker concentrations were extracted for each trial arm. Where data on the amount of fruits and vegetables provided or biomarker levels were incomplete or lacked estimates of precision, authors were contacted. For four trials⁽ Reference Rantala, Silaste and Tuominen ^{15 –} Reference Berry, Mulla and Chowienczyk ^{18 )}, data were supplied by the authors and are included in the review.

Quality assessment

A risk of bias (ROB) assessment was conducted (by M. P.) using the Cochrane risk of bias tool⁽ Reference Higgins, Altman and Gotzsche ^{19 )}. Randomisation, allocation concealment, participant and assessor blinding, missing data and selective outcome reporting were assessed. Other items hypothesised to potentially introduce ROB were also included: the exclusion of participants taking supplements or smoking, participant fasting at the time of blood sampling, diet adherence monitoring and sufficient intervention washout periods (for cross-over trials) (≥4 weeks). The ROB for each trial was considered on the basis of whether any of the items, individually or in combination with others, were likely to have introduced bias, and trials were assigned as having no, possible or high ROB. The overall quality of the evidence for each outcome was assessed using the GRADE system⁽ Reference Guyatt, Oxman and Schunemann ^{20 )} that considers (1) the ROB across trials contributing to that outcome, (2) heterogeneity in the meta-analysis, (3) directness or the generalisability of the population in the trial, (4) precision of the effect size and (5) risk of publication bias.

Data analysis

Standardised mean change (smc) and sd of biomarker concentrations from the pre-intervention to the post-intervention period were computed using the baseline sd within each trial arm, owing to variation in the units reported across studies (μmol/l; mg/dl; μmol/μmol of cholesterol; μmol/mol of lipid). Effect sizes (standardised mean difference (smd) were the difference in the smc of biomarkers between arms with higher v. lowest fruit and vegetable intake. The se of the smd was computed from the variance of the smc and the sample size in each arm. For trials with more than two arms, the arm with the lowest fruit and vegetable intake was compared against all the other arms. To account for the use of the lowest intake arm in multiple comparisons, the sample size of that arm was divided by the number of comparison groups within that study⁽ Reference Abete, Romaguera and Vieira ^{21 )}. Fruit and vegetable intake was described in terms of number of portions using standard UK portion sizes – that is, one portion equates to 80 g of fruits or vegetables^{( 22 )}.

Mean differences in changes in biomarkers between groups allocated different doses of fruits and vegetables across the whole study in cross-over designs were assumed to be the same as mean differences between groups in parallel study designs. Where average biomarker concentrations before intervention and after intervention were described using medians or geometric means, they were assumed to approximate the mean; 95 % CI or interquartile ranges were transformed to approximate the sd assuming a normal distribution. Where data on change was not available, pre-intervention and post-intervention mean (sd) concentrations were extracted, and mean change was computed by subtracting the pre-intervention mean from the post-intervention mean in each arm. The sd of the smc was computed using standard equations⁽ Reference Abete, Romaguera and Vieira ^{21 )} based on the sd at baseline and the sd at follow-up within each arm and biomarker-specific correlations (r) based on published associations between baseline and follow-up concentrations of biomarkers⁽ Reference Gill, Haldar and Porter ^{23 ,} Reference Thompson, Heimendinger and Diker ^{24 )}. Post hoc sensitivity analyses were performed to check the influence of all assumptions on the results, and the pattern of findings was unaltered.

For each biomarker, smd (se) was pooled across all trials using random effects meta-analysis with inverse variance weights, and heterogeneity was estimated using I ^{2 (} Reference Higgins, Thompson and Deeks ^{25 )}. Heterogeneity was considered low or high if I ² was <25 or >75 %, respectively. For the primary analysis, data were combined for each biomarker for trials that included vitamin C and a common set of five carotenoids (α-carotene, β-carotene, β-cryptoxanthin, lutein and lycopene). Sub-group analyses planned a priori were conducted for each biomarker using meta-regression to investigate potential dose–response effect (difference in fruit and vegetable intake between arms in each trial in grams per d) and sources of heterogeneity, including differences by intervention duration (0–3 v. 4+ weeks, categories created based on data available); intervention compliance (meals observed v. eaten at home); trial design (cross-over v. parallel); health status (healthy v. unhealthy); location (Europe v. USA v. Asia-Pacific); type of food provided (fruits and vegetables v. vegetables only, categories created based on data available); baseline fruit and vegetable intake (<1 v. 2–3 v. 4–5 portions, categories created based on data available); fasting status (fasted v. not); blood sample fraction (plasma v. serum); ROB (low v. possible v. high); and sex (mixed v. male v. female). To check for a possible ceiling effect among participants with elevated biomarker concentrations, we also performed sub-group analyses using baseline biomarker concentrations (low v. high based on median split, categories created based on data available). For sub-group analyses, all trials with that biomarker measured were used, regardless of the simultaneous measurement of other biomarkers. As substantial (I ²>75 %) between-trial heterogeneity was observed, a post hoc sensitivity analysis was conducted to examine the effect of excluding trials with outlying results (>2 sd from the smd) from the analysis. Statistical evidence of association was considered important at P<0·05. Data were analysed in Stata, version 12 (StataCorp LP).

Trial selection

Of 3759 unique records, 144 full-text articles were assessed for inclusion, and nineteen trials were included in the review (Fig. 1). In all, nineteen trials were identified in this review, ten of which were also included in the previous systematic review⁽ Reference Baldrick, Woodside and Elborn ^{14 )}. Out of the nineteen trials, nine⁽ Reference Gill, Haldar and Porter ^{23 ,} Reference Baldrick, Elborn and Woodside ^{26 –} Reference Wallace, McEvoy and Hunter ^{33 )} assessed a common set of six biomarkers including five carotenoids and vitamin C (online Supplementary Table S2) and were included in the comparative (primary) analysis. Of the papers rejected on full-text screening, the majority were excluded on the basis of the intervention, often because trials involved only dietary advice or because the intervention targeted a single fruit or vegetable only. Other common reasons for exclusion were wrong study design (not RCT with food provision) or wrong outcomes (no biomarker concentrations).

Fig. 1 PRISMA diagram of search results.

Trial characteristics

Trial characteristics for all the included trials are shown in Table 1. Totally, twelve trials were conducted in healthy populations⁽ Reference Rantala, Silaste and Tuominen ^{15 ,} Reference Dragsted, Pedersen and Hermetter ^{17 ,} Reference Gill, Haldar and Porter ^{23 ,} Reference Briviba, Bub and Moseneder ^{27 ,} Reference Broekmans, Klopping-Ketelaars and Schuurman ^{28 ,} Reference Neville, Young and Gilchrist ^{31 ,} Reference van het Hof, Tijburg and Pietrzik ^{32 ,} Reference Brevik, Andersen and Karlsen ^{34 –} Reference Martini, Campbell and Gross ^{38 )}. Two trials were conducted in populations with increased CVD risk⁽ Reference Chong, George and Alimbetov ^{29 ,} Reference Wallace, McEvoy and Hunter ^{33 )}, and single trials were conducted in populations with obesity⁽ Reference Howe, Valentine and Hull ^{39 )}, overweight⁽ Reference Crane, Kubota and West ^{16 )}, hypertension⁽ Reference McCall, McGartland and McKinley ^{30 )}, elevated blood pressure⁽ Reference Berry, Mulla and Chowienczyk ^{18 )} or chronic obstructive pulmonary disease⁽ Reference Baldrick, Elborn and Woodside ^{26 )}. Within-trial differences in intake of fruits and vegetables ranged from 2 to 13 portions/d. The sample sizes ranged from twenty to 246 participants (median 64). For the nine trials included in the comparative analysis, the difference in the amount of fruits and vegetables between arms ranged from 2 to 7 portions/d.

Table 1 Characteristics of nineteen randomised controlled trials of fruit and vegetable (F&V) intake on biomarker concentrations

COPD, chronic obstructive pulmonary disease; GP, group; NR, not reported

Quality of the evidence

In the GRADE assessment of the quality of each outcome in the meta-analysis, no outcomes were downgraded for imprecision or indirectness. However, most trials were considered to have some ROB (Fig. 2). Trials did not state that there was allocation concealment and patient blinding was not possible. In a number of studies, there were inadequate pre-intervention and within-intervention washout periods and uncertainties around the true ingested amounts of fruits and vegetables (less adherence monitoring) (Fig. 2). In the absence of washout periods, there was considered to be risk of pre-intervention or cross-treatment contamination. In trials where consumption of fruit and vegetables was not directly observed, there was considered to be a likely overestimation of the true ingested amount. A concern in some trials was the inclusion of participants using nutritional supplements, a lack of fasting at the time of outcome measurement and the inclusion of patients who smoked. Funnel plots suggested the possibility of publication bias and heterogeneity for α-carotene, β-carotene, β-cryptoxanthin and vitamin C (based on the occurrence of studies outside of the triangular region indicating where 95 % of studies should be in the absence of bias or heterogeneity) (Fig. 3). All outcomes were downgraded for inconsistency as there was substantial heterogeneity in the meta-analysis. Overall, evidence for all outcomes was graded as having low quality.

Fig. 2 Summary of risk of bias among the nine studies with six biomarkers measured.

Fig. 3 Funnel plots of nine randomised controlled trials of different doses of fruit and vegetable intake on biomarker concentrations. smd, standardised mean difference.

Findings

The primary focus for this review was trials including measures of all six biomarkers so that their comparative utility could be assessed (Fig. 4). All biomarker concentrations, except lycopene, increased more from the pre-intervention to the post-intervention period in the arm providing higher amounts of fruits and vegetables compared with the arm providing lower amounts: α-carotene (smd 0·63; 95 % CI 0·25, 1·01), β-carotene (smd 0·27; 95 %CI 0·08, 0·45), β-cryptoxanthin (smd 0·52; 95 % CI 0·30, 0·74), lutein (smd 0·70; 95 % CI 0·37, 1·03) and vitamin C (smd 0·94; 95 % CI 0·66, 1·22). For lycopene, there was no evidence of greater change in plasma concentrations (smd –0·02; 95 % CI –0·27, 0·23) in response to higher fruit and vegetable intake. There was substantial between-trial heterogeneity in the pooled effects for all biomarkers (I ²=74–94 %). In the sensitivity analyses, where trials with extreme outlying results were excluded, seven out of nine trials remained in the analysis (online Supplementary Fig. S1). Effect sizes were smaller for all biomarkers but a similar pattern was observed, where there were significant effects for α-carotene, β-carotene, β-cryptoxanthin, lutein and vitamin C, but again no evidence of a difference for lycopene was observed. Heterogeneity was reduced for β-crytoxanthin, lutein, lycopene and vitamin C (I ²=46–66 %), but remained significant (online Supplementary Fig. S1). Further sensitivity analyses utilising information for each biomarker from all available studies (indirect comparisons) (online Supplementary Fig. S2) and excluding non-normally distributed data (online Supplementary Fig. S3) did not alter the pattern of the results.

Fig. 4 Summary of pooled difference between arms consuming higher v. lower amounts of fruit and vegetables for standardised mean change (smc) of biomarkers from the pre-intervention to the post-intervention period in trials with all six biomarkers measured. smc represents an sd of pre-intervention biomarker levels within each study. I ² is an indicator of between-trial heterogeneity. Random effects meta-analysis was used to pool mean differences. Includes the following studies for ALL biomarkers: Baldrick et al.⁽ Reference Baldrick, Elborn and Woodside ^{26 )}, Briviba et al.⁽ Reference Briviba, Bub and Moseneder ^{27 )}, Broekmans et al.⁽ Reference Broekmans, Klopping-Ketelaars and Schuurman ^{28 )}, Chong et al.⁽ Reference Chong, George and Alimbetov ^{29 )}, Gill et al.⁽ Reference Gill, Haldar and Porter ^{23 )}, McCall et al.⁽ Reference McCall, McGartland and McKinley ^{30 )}, Neville et al.⁽ Reference Neville, Young and Gilchrist ^{31 )}, van Het Hof et al.⁽ Reference van het Hof, Tijburg and Pietrzik ^{32 )}, Wallace et al.⁽ Reference Wallace, McEvoy and Hunter ^{33 )}. Total number of trials is nine; total number of arms being compared is twenty-two; total number of people included is 667.

Individual meta-analyses for each biomarker including up to nineteen trials are shown in the online Supplementary Fig. S4–S9. For these indirect comparisons, the same pattern was observed as for direct comparisons, with statistically significant effects for all biomarkers except lycopene. For these indirect analyses, we were able to additionally estimate the effects for zeaxanthin (online Supplementary Fig. S10) and total carotenoids (online Supplementary Fig. S11), which were available in a smaller number of studies. Both showed increases in response to high compared with low amounts of fruits and vegetables but were also highly heterogeneous (I ²=84 and 93 %, respectively).

All trials providing data on at least one biomarker were included in the investigation of dose–response and sub-group analyses. In meta-regressions of within-trial difference in the amount of fruits and vegetables (grams/d) against smd of biomarker level, there was no evidence of a dose–response effect (all P>0·05). When the difference in the amount of fruits and vegetables consumed in each arm was categorised into portions (2–3 v. 4–5 v. >5 portions), a trend towards higher biomarker concentrations among trials where the group difference in fruit and vegetable intake was greater emerged, but was only statistically significant for β-carotene (P=0·01, Fig. 5).

Fig. 5 Summary of pooled differences between arms consuming higher v. lower amounts of fruit and vegetables in standardised mean change (smc) of biomarkers from the pre-intervention to the post-intervention period in all trials with available data grouped by the amount of fruit and vegetables provided during the intervention. smc represents an sd of pre-intervention biomarker levels within each study. I ² is an indicator of between-trial heterogeneity. Random effects meta-analysis was used to pool mean differences. P value is from meta-regression test for trend across categories. Includes all studies up to n 19 based on availability of biomarker in each study.

Other notable findings from sub-group analyses included stronger effects for α-carotene, β-carotene, lutein and vitamin C in trials where participants ate meals under supervision compared with trials where food was eaten at home, accounting for 12–38 % of the heterogeneity (online Supplementary Fig. S12). Shorter interventions (0–3 weeks) were associated with significantly greater effect sizes compared with longer (≥4 weeks) interventions for α- and β-carotene. There were non-significant trends for a similar effect for lutein, lycopene and vitamin C, accounting for between 6 and 20 % of the heterogeneity (online Supplementary Fig. S13). Trials in healthy populations tended to show greater effect sizes compared with trials in unhealthy populations (online Supplementary Fig. S14), and this was significant for α- and β-carotene (accounting for 17–18 % of the heterogeneity). In the sensitivity analysis, excluding outlying results, there was still a significant effect of disease status for α- and β-carotene. In the sensitivity meta-regressions including intervention delivery, duration and participant health status together, associations were unaltered (data not shown).

Trials conducted in the USA had significantly greater effect sizes compared with those conducted in Europe for α- and β-carotene (online Supplementary Fig. S15), which was robust to adjustment for other factors for α-carotene. The effect size was greater for cross-over compared with parallel trials for β-carotene and lutein (online Supplementary Fig. S16), which was attenuated after adjustment for other factors (data not shown). For α-carotene and lutein, there was a greater effect size for trials where vegetables alone were given compared with trials where both fruit and vegetables were given (online Supplementary Fig. S17), but these findings were not robust to adjustment (data not shown). There was no evidence of differences across sub-groups defined by baseline fruit and vegetable intake, fasting status, blood fraction (plasma or serum) or ROB (data not shown).