figure a


Type 2 diabetes (hereafter referred to as diabetes) and its complications constitute a major threat to global health. Populations of South Asian origin have at least a two to four times higher risk than European-origin populations [1, 2]; they also develop diabetes and its complications at a younger age [3, 4]. Patterns of fat deposition, low lean mass [5] and low birthweight [6] are considered to contribute to early-onset insulin resistance and diabetes in South Asians, exacerbated by changing lifestyle, urbanisation and migration. Further, the proportion of deaths attributable to diabetes is almost 50% higher in South Asians than in populations of European origin, with no clear signs that the risk will level off over time [7]. Therefore, adequate actions for prevention of diabetes among South Asians are imperative.

From 2001, efficacy trials have documented a reduction in diabetes incidence in high-risk adults of up to 58% through lifestyle modification interventions incorporating a healthy diet and/or physical activity [8,9,10,11,12]. Recent meta-analyses of efficacy and pragmatic diabetes prevention trials found a 30–40% reduction in diabetes incidence [13, 14]. However, none of these meta-analyses reported effects specifically for South Asians [13, 14], although several lifestyle intervention trials, culturally adapted to suit the specific needs, cultural norms and values of the South Asian populations in different contexts, have been developed and tested [15,16,17,18]. Some trials in South Asian populations appeared moderately successful, but the effectiveness differed across studies. Although a few previous reviews have explored the effects of lifestyle modification interventions on glucose measures and anthropometric outcomes in South Asians [19], many were narrative and limited by inconsistent outcome reporting, and none reported estimates for diabetes incidence [20]. Owing to small numbers in individual studies and low power—the a priori estimation of effects was too optimistic, drop out was higher than anticipated or the incidence of outcomes was lower than expected—the strength of the evidence of individual studies is limited and further exploration of effects (e.g. subgroup analyses) are hampered.

Given the current knowledge gap about the effectiveness of lifestyle modification interventions in this high-risk population, we aimed to assess the overall effect from RCTs with lifestyle interventions involving dietary modification and/or physical activity to prevent diabetes in adult South Asians, using individual participant data meta-analysis [21]. In addition, we explored whether the effects differed by sex, BMI, age, study duration or study region.


The plan for this individual participant data meta-analysis was registered with the International Prospective Register of Systematic Reviews ([PROSPERO] registration CRD42017078003). The reporting is in line with the Preferred Reporting Items for Systematic Reviews and Meta-Analysis of Individual Participant Data (PRISM-P) [22].

Eligibility criteria

Individual and cluster RCTs that compared the effect of lifestyle modification interventions incorporating a healthy diet and/or physical activity with usual care/control groups not receiving the lifestyle intervention to prevent diabetes in the South Asian adult population (≥18 years) worldwide were eligible. To reduce the risk of bias in the estimates, studies were included only if judged to be at least of moderate quality by the Quality Assessment Tool for Quantitative Studies (see below) [23]. We decided a priori to exclude studies exclusively on drugs and surgical interventions, alone or in combination with lifestyle modification, and studies evaluating physical activity components not consisting of cardiovascular exercise (e.g. yoga, resistance training) and single dietary components (such as vitamin D supplementation).

Search strategy

We searched PubMed, EMBASE, Cochrane Library and Web of Science, first from the start to 30 September 2017 to allow for inclusion of studies for the individual participant data meta-analysis (search strategy shown in electronic supplementary material [ESM] Methods), supplemented with reference list tracing of key reviews [19, 20, 24, 25], included studies [12, 15,16,17,18, 26] and by searching trial registers [27]. The search was updated on 24 September 2018. The work builds on and complements our previous narrative systematic review of dietary and physical activity components recommended in experimental, quasi-experimental and before/after studies to prevent diabetes in South Asians [27].

Study selection and risk of bias assessment

Two reviewers independently screened titles and abstracts for eligibility, thereafter full texts. The risk of bias was assessed at the study level using the Quality Assessment Tool for Quantitative Studies [23] by three reviewers, as we perceived the quality assessment tool to be vulnerable to differences in interpretation. For all processes, a discussion was held to reach consensus in cases of discrepancy. Only studies judged to be at least of moderate quality were included [23].

Data extraction

Information on study design and population characteristics was extracted from published manuscripts, study protocols and personal communication with authors, using a piloted extraction form, by two reviewers working independently [27]. A discussion was held in case of discrepancies and a third independent reviewer arbitrated, if necessary. Principal investigators of eligible studies published prior to 30 September 2017 were contacted for permission to include individual participant data on sex, age, anthropometric measures and fasting and 2 h glucose results from oral glucose tolerance tests. To allow for transformation and analysis, the deadline for data inclusion was set to 1 March 2018. After each primary study gave ethical approval for the current study, data transfer agreements were signed before we received anonymised individual level data.

Data analysis and quality of the evidence

The primary outcome was diabetes incidence, defined as fasting glucose ≥7.0 mmol/l and/or 2 h glucose ≥11.1 mmol/l or registered as doctor diagnosed by self-report at follow-up visits [12, 16, 26]. Secondary intermediate outcomes were changes in fasting glucose, 2 h glucose, body weight (kg) and waist circumference (cm). Except for one study with age stratified into nine groups [16], we used age in years. Analyses were based on an intention-to-treat approach. All principal investigators confirmed preliminary analyses after variable standardisation and data cleaning.

Cox proportional hazards models were used to estimate age-adjusted HRs for diabetes incidence. The three studies with men and women were also adjusted for sex. The proportional hazard assumption was tested separately for each covariate and also overall (global test) and found to be met. Time to diabetes was used as the dependent variable in our survival analysis, and was calculated from the trial start date to either the diabetes diagnosis or the end of each trial. For secondary outcomes, we determined mean differences with 95% CIs, adjusted for baseline values for the outcome variable of interest, and age and sex when relevant, using the last available estimate [21]. Individual participant data meta-analyses were done in two stages. First, mixed regression models with random effects at individual level and group allocation as fixed effect were performed for each study. For secondary outcomes, we used generalised linear models with identity link function for studies with only one follow-up visit [17, 18]. Second, estimates from the multilevel models were used to estimate weighted averages across studies. We used the I2 statistic to express the amount of variance attributable to study heterogeneity. A priori-defined subgroup analyses were performed for sex, baseline BMI ≥27.5 kg/m2 or <27.5 kg/m2 [28], age >44 years or <44 years, short- (5–7 months) and long-term (2–3 years) study duration and region, testing for interactions between intervention and subgroups. We used Stata version 15.1 (StataCorp, College Station, TX, USA) for the regression analyses for each study and to generate weighted estimates across studies, and the metafor package in R ( for the forest plots. Because of loss to follow-up in some studies, we performed a sensitivity analysis, applying within-trial multiple imputations [29]. In trials with multiple follow-up visits [15, 16, 26] we used REALCOM-IMPUTE [30] to account for correlation between repeated measures. In trials with one follow-up [17, 18], we used multiple imputation by chained equations [31]. For each trial, we created 20 imputed datasets, thereafter combined for pooled estimates. We also performed funnel plots to visually explore possible publication bias/small study effect.

Finally, we used the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) approach to rate the quality of evidence and to generate absolute effect estimates for the outcomes [32]. In our previous systematic review, we identified RCTs that ended before or in 2015 and had not published results or updated their registrations [27]. Of the six trials that were identified and might have been eligible for the individual participant data meta-analysis, three have not yet reported, one has since published results [33] and two are ongoing. Although it is unclear if the underreporting is related to unsuccessful inclusion or negative results, it is indicative of a probable publication bias.


Included studies

We identified 4240 publications and, after removal of duplicates, 3009 titles and abstracts were screened, of which 2983 were excluded (ESM Fig. 1). Following full-text review of 17 articles, six RCTs fulfilled the inclusion criteria for the individual participant data meta-analysis [12, 15,16,17,18, 26]. The characteristics of excluded studies [33,34,35,36,37,38,39,40,41,42,43] are given in ESM Table 1. We obtained individual participant data for all six eligible studies for the meta-analysis (Table 1). When cleaning the data for analyses, we identified 29 individuals from two studies with glucose levels indicative of diabetes at inclusion [17, 18]. These individuals were excluded, leaving 1816 participants, of whom 604 (33%) were women. Three studies included both sexes, two were in men [18, 26] and one was in women only [17]. Four studies were of 2–3 years’ duration [12, 15, 16, 26], while two lasted 5–7 months [17, 18]. Four studies were from Europe [15,16,17,18] and two from India [12, 26]. Five studies combined a dietary and physical activity intervention [12, 15,16,17, 26], while one used physical activity only (primarily floorball/field hockey) [18]. Individual and/or group-based educational sessions were provided with varying intensity, one supported by mobile phone messages [26]. One study offered home visits from a dietitian and involvement of family members [16], three were performed in the community [15, 17, 18] and the two Indian studies were performed in a workplace setting [12, 26]. The quality of five studies was rated as strong (ESM Table 2).

Table 1 RCTs included in the individual participant data meta-analysis

Loss to follow-up differed from more than 20% in two studies of 7 months’ and 2 years’ duration [15, 17] to 0–2.3% in other studies that had 3 years’ duration [12, 16] (ESM Table 3). No clear patterns of differences in baseline characteristics between the intervention and control groups were observed (ESM Table 4). The age range at inclusion was 18–80 years, with mean age differing from 37 [18] to 52 years [16]. The mean BMI differed from 26 kg/m2 in the Indian studies [12, 26] to 30.6 kg/m2 in one European study [16]. At baseline, mean 2 h glucose levels were highest in the Indian studies, which only included participants with persistently impaired glucose tolerance [12, 26].

Diabetes incidence

Incident diabetes was observed in 118 of 936 participants (12.6%) in the intervention group and in 176 of 880 participants (20.0%) in the control group (adjusted HR 0.65 (95% CI 0.51, 0.81; p = 0.0002, I2 = 0%) (Table 2 and Fig. 1). This reflects an absolute reduction of 7.4% (95% CI 4.0, 10.2) and a number needed to treat (NNT) of 14 (95% CI 10.0, 28.0). In subgroup analyses, we did not find differences in relative estimates of effects by sex, age, BMI, study duration or region. For the long-term studies, the estimates were identical after exclusion of one study with high loss to follow-up [15] (0.65 [95% CI 0.51, 0.83], p = 0.0004). Overall, and for subgroups, we found no (I2 = 0%) or low between-study heterogeneity. According to GRADE, the quality of evidence can be rated as moderate; quality was rated down owing to some, but overall limited, concerns of risk of bias (e.g. lacking blinding of participants and providers) and possible publication bias, although not confirmed by visual inspection of funnel plots (data not shown).

Table 2 Diabetes risks overall and within pre-specified subgroups
Fig. 1
figure 1

Forest plot of HRs for diabetes in lifestyle intervention vs control groups, overall and for subgroups. IDPP-1, Indian Diabetes Prevention Programme; Innva-Diab-DE-PLAN, the InnvaDiab (Immigrant Diabetes) study, part of the European Union project DE-PLAN (Diabetes in Europe – Prevention using Lifestyle, Physical Activity and Nutritional Intervention); PAMH, Physical Activity and Minority Health; SMS, short message service

Other outcomes

For secondary outcomes, we observed a reduction in 2 h glucose (−0.34 mmol/l [95% CI −0.62 to −0.07], p = 0.02, I2 = 50%), weight (−0.75 kg [95% CI −1.34 to −0.17], p = 0.01, I2 = 71%) and waist circumference (−1.16 cm [95% CI −2.16 to −0.16], p = 0.02, I2 = 75%) in the intervention vs control group (Table 3 and ESM Fig. 2), with no effect for fasting glucose (−0.03 [95% CI −0.10 to 0.04], p = 0.37, I2 = 0%). No consistent subgroup differences in effects were found for secondary outcomes (ESM Table 5), although a difference in effect for weight (−1.10 kg vs −0.08 kg, p = 0.02 for interaction) was revealed for European vs Indian studies, with a similar but non-significant difference in the effect estimate for waist circumference (−1.59 cm vs −0.26 cm, p = 0.09 for interaction). The GRADE quality of evidence for 2 h glucose was rated as moderate, rated down owing to some concerns about risk of bias. For the other secondary outcomes, the quality was considered low because of imprecision and/or heterogeneity between studies (ESM Table 5). Sensitivity analyses after multiple imputations yielded similar effect estimates as the main analysis for secondary outcomes, although CIs were slightly wider and the result for weight was only borderline significant (p = 0.06; results not shown). Last, the distribution curves for fasting and 2 h glucose at the last visit for all studies merged showed a larger shift to the right in the control than in the intervention group (ESM Fig. 3). In the intervention group, 4.5% had developed diabetes at the last follow-up based on fasting glucose values ≥7.0 mmol/l and 8.2% based on 2 h glucose values ≥11.1 mmol/l, compared with 7.5% (p = 0.017) and 14.9% (p < 0.0001), respectively, in the control group.

Table 3 Overall effects of lifestyle interventions (mean difference) on continuous outcomes


In this individual participant data meta-analysis we found that the overall incidence of diabetes was reduced by 35% in the group receiving the lifestyle modification intervention compared with the control group, with an absolute risk reduction of 7.4% and an NNT of ~14 over a mean of 2 years. Results for diabetes were consistent for men and women and for other key subgroups. The studies included participants with normal and impaired glucose regulation and a range of age and BMI values at baseline, indicating high external validity. We also found evidence of a modest effect on 2 h glucose and simple measures of adiposity (weight and waist circumference). Of note, however, at the last follow-up we observed for both fasting and 2 h glucose a stronger shift on the right side of the distribution curve in the control groups, with significantly more diabetes cases compared with the intervention groups.

The main strength of this study is that it is the first to report summary effects of lifestyle interventions on diabetes incidence in the high-risk South Asian population. Starting with a systematic review of the literature, we used individual participant data meta-analysis from all the six eligible RCTs among South Asians published prior to the inclusion date to produce precise summary effects overall and across subgroups [21]. The data were also analysed for the secondary outcomes: fasting and 2 h glucose and adiposity measures. Compared with meta-analyses based on aggregate data on a study level from published papers, individual participant data meta-analysis facilitates standardisation of analyses and increases the precision of estimates and the quality of subgroup analyses [21]. Further, assessing the quality of evidence according to GRADE, we systematically and transparently assessed all factors that could impact on our certainty in the effects estimates, including risk of bias in each study, heterogeneity, indirectness, imprecision and publication bias [32].

However, our review is limited by a relatively small number of trials, some with high loss to follow-up, and relatively little variation in settings. The included studies counted approximately 300 incident cases of diabetes, the majority (83%) in the Indian studies. Thus, the power to detect subgroup differences and to further explore between-study heterogeneity for the adiposity measures was limited. Although we consider the evidence to be generally applicable to the target population, we cannot fully rule out potential indirectness due to differences between populations (country of origin or migrant status), interventions (type, content, intensity, mode of delivery, compliance) and settings (family- and community-based vs workplace).

Our meta-analysis included efficacy [12, 16] and more pragmatic trials [15, 17, 18, 26]. The 35% RR reduction in diabetes incidence resembles the RR reduction (39%) reported in a standard meta-analysis of 19 efficacy and pragmatic diabetes prevention trials in individuals with impaired glucose tolerance or impaired fasting glucose after a mean of 2.6 years of active lifestyle intervention [13]. Although ten studies in the latter meta-analysis were conducted in Asia, the results for South Asians were not reported. Of note, the absolute benefit on diabetes prevention was higher in our study than in this meta-analysis (7.4% vs 4.0%, respectively) and so the NNT was lower (14 vs 25) [13]. Our findings compare well with the first efficacy studies, reporting 6.2–12% absolute risk reductions (NNT 16–8, respectively) [8, 9], but are somewhat stronger than those of a meta-analysis of translational studies to prevent diabetes in high-risk populations other than South Asians (RR reduction, 29%; absolute risk reduction, 3%) [14].

Meta-analysis of diabetes prevention trials indicated that combined dietary and physical activity interventions were more effective than either strategy type alone, but the number of studies was limited for single strategy studies [13]. More studies to date seem to indicate that the dietary interventions are more effective than physical activity interventions to reduce diabetes incidence in high-risk individuals. There is, however, evidence that interventions focused on increasing physical activity in individuals with impaired glucose tolerance are effective at inducing sustained reductions in 2 h glucose concentrations [44]. Although one of the studies with a physical activity intervention that was included in our study found effects on weight and waist [18], more studies comparing dietary and activity interventions for diabetes prevention are needed in other ethnic groups than of European origin.

Although we could not assess this, the intensity of the interventions in the long-term studies in our review was lower [12, 26] or comparable [16] with those of the first efficacy studies, which had stronger effects [8, 9]. Even a low-cost community-based peer-support lifestyle intervention programme published too late to be included in our study found a 12% relative reduction in diabetes incidence in individuals with a high Indian diabetes risk score [33]. The smaller effect might be partly attributed to the selection of participants, with the majority having normal blood glucose levels or isolated impaired fasting glucose at baseline, as there are no interventions so far proven to reduce diabetes incidence in such individuals [33].

In contrast to our finding of an overall 35% reduction on diabetes incidence with a relative small mean 0.75 kg reduction in weight, studies in other populations found weight reduction to be the main driver of the effect [8, 9, 13], i.e. 16% reduction in diabetes incidence for each kg of weight loss [45]. Of note, despite small changes in mean weight and waist circumference values, both for fasting glucose (primarily reflecting hepatic insulin resistance) and 2 h glucose (reflecting muscle insulin resistance), the intervention had a more profound effect on the right side of the distribution curve, indicating reduced insulin resistance [46]. Further, achievements of dietary and physical activity goals, even without weight loss, may improve 2 h glucose and reduce diabetes incidence [8], as found in the Indian studies in our review [12, 26], where a reduction in portion size, consumption of carbohydrates and oil intake was related to a lower diabetes incidence even without weight loss [39]. Although trials among South Asians are few, there are indications that improvements in the quality of the diet (more complex carbohydrates and monounsaturated and polyunsaturated fatty acids) might improve blood glucose, serum insulin, lipids, inflammatory markers and hepatic fat [5].

Although only 33% of participants were women in our meta-analysis, as previous studies were underpowered to study potential sex differences, an important new finding was that their diabetes incidence was significantly reduced, despite a slightly smaller non-significant effect estimate for 2 h glucose, weight and waist circumference than in men. Furthermore, the larger effect on weight, with a concomitant non-significant effect estimate for waist circumference, in the European compared with the Indian studies in our meta-analysis is noteworthy. This may reflect different phenotypes, as it would be easier to achieve a larger reduction in weight measures among South Asians residing in Europe, who had higher BMIs, but differences in the diet and/or physical activity level at baseline or induced by the intervention may also be involved. Furthermore, perceptions of weight and health are changing across generations and differ by region [47]. South Asians living in Europe, particularly women, may be more sensitised to the importance of weight loss than those living in their country of origin [48]. Others report that sex did not influence the effects of lifestyle interventions [13]. Lifestyle trials aimed at weight loss which explored sex differences in anthropometric outcomes mostly report stronger effects in men, but actual sex differences were small [49]. Thus, there is little evidence yet to indicate that men and women should generally adopt different weight loss strategies.

The results of our meta-analysis, that lifestyle interventions had comparable and clinically important effects on diabetes incidence across different regions, is important for policymakers and clinicians. The studies outside India in our meta-analysis, conducted mostly among first-generation South Asian migrants in Europe, had suggested only a modest effect. The consistency of effects across subgroups of the South Asian population at risk of diabetes has made us rethink, contrary to our a priori beliefs based on the outcomes of single studies, that benefits may actually be achieved by lifestyle modification interventions, not only under ideal conditions, but also in real-life settings [15, 17, 18, 26]. Cultural adaptations to mode of delivery may be necessary across contexts as cultural adaptations likely promote the effectiveness of interventions among specific ethnic populations [50, 51], although evidence of the effects of cultural targeting on diabetes prevention outside India is still scarce [52]. Interestingly, one study using a culturally targeted physical activity intervention (floorball/field hockey) for men provided strong results for all secondary outcomes, although it was limited by its small size and short duration [18].

There are, however, several unanswered questions that should be addressed. First, lifestyle intervention studies have used generic recommendations (i.e. based on those for the local majority population) [27], while different targets for dietary and physical activity recommendations may be necessary [27, 53]. Further work is required on the mechanisms by which these interventions are having their effect, including the role and type of physical activity, diet quality, specific dietary components, cooking practices and timing of meals [54]. Further, we predict that larger effects on diabetes risk in South Asians might be achieved in studies aiming at larger weight reductions [55, 56] and with more intense dietary and physical activity changes [53]. Reach, retention and long-term sustainability may be enhanced through improvements in targeting and delivery, and benefits in the longer term (beyond 2–3 years) should be further investigated. It is hoped that ongoing studies will be complemented with new high-quality trials addressing the issues outlined above. Future systematic reviews should be extended to explore effects by mode of delivery, intensity and contextual factors. In addition to high-risk strategies, there are strong recommendations for population-based strategies as part of national public health policies [54].

In conclusion, pending deeper understanding of the causation of diabetes in South Asians and the development of new kinds of intervention, this individual participant data meta-analysis of lifestyle modification interventions in South Asian populations at high risk of diabetes provides evidence of a clinically important 35% relative reduction in diabetes incidence, with an NNT of 14 to prevent one case of diabetes over a mean of 2 years. Given the substantial and growing burden of diabetes, particularly in South Asian origin populations, this meta-analysis provides support for ongoing strategies underpinned by weight loss, dietary change and increased physical activity to prevent diabetes. Future work should aim to understand the mechanisms by which these effects occur, evaluate cost effectiveness and develop more effective interventions.