Search strategy

We searched Medline, Embase, Global Health, and Global Index Medicus from inception up to 14 November 2020 to identify systematic reviews of observational studies on the association between diabetes and CVD, CKD, dengue, depression, pneumonia, and TB. CVD encompassed coronary heart disease, myocardial infarction, angina, stroke and transient ischaemic attack, since diabetes has been shown to be associated with a markedly higher risk of these conditions [15 ]. The search included variations of the following as MeSH terms or keywords: diabetes, CVD, CKD, dengue, depression, pneumonia, and TB, as well as terms for LMICs, including individual country names (Appendix S1 in the Online Supplementary Document ) [16]. We also perused reference lists of identified relevant studies. Two authors (AAL and AL) independently screened titles, abstracts, and full-text articles. Any discrepancies were resolved through discussion.

Inclusion and exclusion criteria

We included systematic reviews, with or without meta-analyses, if they included cross-sectional, case-control or cohort primary studies. Primary studies had to have evaluated the prevalence or incidence of the relevant comorbidity in people with diabetes or the role of diabetes as a risk factor for the subsequent development of the condition. We included reviews of primary studies conducted solely in LMICs, or from a mixture of HICs and LMICs if at least two studies within a review were from an LMIC. Only LMIC studies from the latter were used in the analysis.

We excluded reviews which: focused on the association between diabetes and its comorbidities in the context of a third disease (for example, those which solely examined the association between diabetes and TB in an HIV-positive population); included diabetes as the outcome or focused solely on type 1 diabetes mellitus or gestational diabetes; examined studies of mortality outcomes; or were not conducted systematically (ie, narrative reviews, commentaries, editorials). Although our search strategy did not exclude non-English articles, we ultimately excluded articles written in languages other than English, Spanish, or Portuguese.

Study selection

If multiple reviews included overlapping primary studies, the review which was most comprehensive (eg, greater number of LMIC studies) was included. If reviews on the same outcome focused on different settings (for example, one country vs multiple), both were included. If there were prevalence and incidence or association reviews on the same comorbidity, provided they did not overlap in terms of primary studies or settings, all eligible studies were included.

Data extraction

Data were extracted independently by two authors (AAL and AL) using a standardised form adapted from the Joanna Briggs Institute [14]. Any discrepancies were resolved through discussion. If a review included primary studies from both HICs and LMICs, we only extracted data pertaining to LMIC studies. We categorised LMICs based on the World Bank’s historical income classification at the beginning of the primary study period [17]. Thus, the same country may be classified in different income levels depending on when the study began. We extracted the following data from reviews: first author; year; number of included LMIC primary studies; study designs; study settings; number of participants; country of primary study; population; exposure; exposure ascertainment; comparators (if any); outcome; outcome ascertainment; databases searched; and quality assessment method. For eligible meta-analyses, we extracted the following additional data items: study period; meta-analysis metric; number of outcomes (if more than one); type of model; factors adjusted for; effect estimate with 95% confidence interval; variables that were adjusted for; measures of heterogeneity (Q and/or I²) with P values; and measures of publication bias (Egger’s Test and/or Begg’s Test). If necessary data were not provided, or if there were any discrepancies identified in the review, we examined primary studies to extract the appropriate information.

Quality appraisal

The methodological quality of both systematic reviews and meta-analyses was assessed independently by two authors (AAL and AL) using an adapted version of the AMSTAR 2 checklist (Appendix S2 in the Online Supplementary Document ) [18]. AMSTAR 2 is an improved version of AMSTAR, a tool designed for the critical appraisal of systematic reviews of non-randomised and/or randomised studies of health care interventions. AMSTAR 2 is more inclusive to systematic reviews of observational studies and is comprised of the following domains: methods (ie, proper reporting of PICO, protocol, search strategy, eligibility criteria, data extraction and analysis), quality assessment (ie, risk of bias appraisal), and meta-analysis methods (ie, appropriate methods, investigation of heterogeneity and publication bias). Where the AMSTAR 2 checklist asked about randomised and non-randomised studies separately, we removed the parts related to randomised studies so the focus could be on observational studies. We also included an additional question on the type of quality appraisal tool used and whether quality descriptions were provided.

The checklist uses a rating system of high, moderate, low, and critically low confidence in the results, which is based on how well the identified review covers the domains deemed to be most important to the validity of umbrella reviews [18]. The critical domains included: adequate search strategy, sufficient reporting of study characteristics, risk of bias/quality appraisal, appropriate meta-analytical methods, accounting for risk of bias during results interpretation, and assessment of publication bias. If one critical domain was missing, the review was automatically dropped to a low confidence. If two or more critical domains were missing, the review was given a critically low confidence and thus excluded.

Statistical analysis

For systematic reviews which did not conduct a meta-analysis, or for which we did not conduct a secondary meta-analysis, we conducted narrative syntheses, which comprised a summary of primary study characteristics and main findings. For reviews which did include a meta-analysis, all included meta-analyses were re-run using random-effect models and summary estimates were pooled based on the reported effect estimates from each review [19]. Where reviews included both HIC and LMIC studies in their meta-analysis, we only included the LMIC studies in our secondary meta-analysis. We extracted the following information for each primary study from an LMIC to conduct the secondary meta-analysis: first author; year; country of study; study design; study size; income classification of country at first year of study period; study population; any information on diagnosis or type of exposure or outcome; effect estimate; and 95% confidence intervals (CI).

We assessed heterogeneity using the I² metric, with I² <50%, I² = 50%-75% and I² >75% suggesting low, moderate, and substantial heterogeneity, respectively [20]. If substantial heterogeneity was present and there were a sufficient number of studies, pre-defined subgroup analyses (study design and country income level) were conducted to explore potential sources of heterogeneity. If the identified meta-analysis conducted relevant subgroup analyses, we also re-ran those analyses if included primary studies differed or if the original analysis was not pooled using random effects models. If a review included at least ten primary studies, we explored publication bias using a funnel plot and Egger’s test to evaluate whether smaller studies contribute more to the effect estimate compared with larger studies [21]. For prevalence meta-analyses, we did not assess publication bias because these funnel plots were found to be inaccurate [22].

We used Stata version 16 (StataCorp LLC, College Station, Texas, USA) to conduct secondary meta-analyses [23].

Cardiovascular disease

The two reviews identified for CVD [24,25] included 22 primary studies (six cohort, six case-control, ten cross-sectional) from nine countries. There were 3 108 248 participants, of which 3 060 517 (98%) were from one primary study [39 ].

Einarson et al [24] did not perform a meta-analysis, but reported that among people with diabetes in upper middle-income countries, the weighted average prevalence of stroke, myocardial infarction, and coronary artery disease was 6.9%, 3.4%, and 17.7%, respectively. In lower middle-income countries the weighted average prevalence of stroke and coronary artery disease was 6.3% and 27.3%, respectively, with no study reporting on the prevalence of myocardial infarction.

Based on our secondary meta-analysis of the eight studies (six case-control, two cross-sectional) from Poorzand et al [25], we found an overall OR for premature coronary artery disease of 2.35 (95% CI = 1.71-3.21) in people with vs without diabetes in Iran, with low heterogeneity between studies (I²  = 25.5%, P = 0.225) ( Figure 2 ).

Figure 2.  Secondary meta-analysis of studies included in Poorzand et al. [25] showing the study-specific and summary estimates of the prevalence of premature chronic artery disease among people with diabetes in Iran.

Chronic kidney disease

We identified two reviews [26, 27] on CKD occurrence amongst people with diabetes, one of which only included two LMIC studies out of 71 included studies [26], while the other included 12 studies from Ethiopia [27]. The studies from the former were both cohort studies, including a combined total of 1114 participants with diabetes [ 26]. The annual cumulative incidence of CKD among people with diabetes in each primary study was 8.6% [40] and 8.1% [41].

We performed a secondary meta-analysis on the review by Shiferaw et al, which included two cohort, one case-control, and nine cross-sectional studies [27]. The authors provided prevalence estimates separately for CKD stages 1-5 and 3-5, so we did the same for our secondary meta-analysis. Seven cross-sectional studies were included for stages 1-5 and the pooled prevalence was 36% (95% CI = 26%-45%) ( Figure 3 , Panel A). For stages 3-5, 10 studies gave a pooled CKD prevalence of 15% (95% CI = 11%-19%) ( Figure 3 , Panel B). Heterogeneity for the two groups were both substantial, with I²>90%. Income level remained the same for Ethiopia throughout the study period, and all stage 1-5 studies were cross-sectional, which precluded subgroup analyses. Subgroup analysis by study design for stage 3-5 studies was not performed due to having too few studies in each subgroup.

Figure 3.  Secondary meta-analysis of studies included in Shiferaw et al. [27] showing the study-specific and summary estimates of the prevalence of chronic kidney disease among people with diabetes in Ethiopia for chronic kidney disease Panel A. Stages 1-5. Panel B. Stages 3-5.

Depression

Four reviews on depression [28,30-32] included 528 987 participants from 145 primary studies (142 cross-sectional and three case-control). An additional 41 cross-sectional studies, with an unknown number of participants, were from one review which did not provide specific study characteristics [29]. A total of 16 LMICs were represented across all five reviews, with the majority being from India and Iran. Based on the five reviews, we found that in LMICs the average prevalence of depression, including major depressive disorder, in people with diabetes ranged from 30%-61%. We conducted secondary meta-analyses on three of these reviews [30-32].

The systematic review by Mendenhall et al focused specifically on LMICs and found that the average prevalence of depression in people with diabetes was 36% [28 ]. While Uphoff et al. [29] did conduct a meta-analysis, this was done rapidly so the authors only provided aggregated study characteristics and a pooled depression prevalence. Due to this, we were not able to conduct our own secondary meta-analysis. Nonetheless, based on 43 estimates from 41 cross-sectional studies from Bangladesh, India, and Pakistan, they found a pooled depression prevalence of 40% (95% CI = 34%-45%) in people with diabetes. They found substantial heterogeneity (I²  = 97%) and conducted meta-regression analysis by publication year, country, disease (since they looked at various non-communicable diseases), sample size, diagnostic tool, and study quality. None of these variables fully explained the heterogeneity.

Hussain et al [30] included 43 cross-sectional studies from India and the pooled prevalence from our secondary meta-analysis was 38% (95% CI = 31%-45%). Heterogeneity was substantial (I²  = 99%, P < 0.001). Since India remained at the same income level when all the primary studies occurred, we could not do a subgroup analysis using income level.

Our secondary meta-analysis of Khalighi et al. [31] showed a much higher overall prevalence of depression in people with diabetes (61%, 95% CI = 55%-67%) compared to the other reviews, albeit with substantial heterogeneity (I² = 98%, P < 0.001). This review only included studies from Iran, which was considered a lower-middle income country before 2009 and an upper-middle income country from 2009 onwards [17]. Our subgroup analysis did not show any significant difference in the prevalence between the two income-level categorisations. Among the studies identified in this review, eight different depression assessment tools were used, with the Beck Depression Inventory (BDI) the most common, being used in 75% of studies. In a post-hoc subgroup analysis comparing depression prevalence in studies using the BDI vs all other tools, the pooled prevalence was higher in the former than the latter studies (65%, 95% CI = 59%-71% and 48%, 95% CI = 33%-63%, respectively). Whilst between-subgroup heterogeneity was statistically significant (P  = 0.028), there was still substantial heterogeneity within each subgroup ( I²>97%, P < 0.001) ( Figure 4 ).

Figure 4.  Subgroup analysis from secondary meta-analysis of studies included in Khalighi et al. [31] showing the study-specific and summary estimates of the prevalence of depression in people with diabetes in Iran by depression assessment tool.

In our secondary meta-analyses of LMIC studies included in the review by Wang et al [32], we computed both the pooled prevalence across all ten studies, and the pooled OR across the three case-control studies. The pooled prevalence of major depressive disorder in people with diabetes was 30.7% (95% CI = 16.6%-44.9%), with substantial heterogeneity between studies (I² = 99%, P < 0.001) (Figure S1 in the Online Supplementary Document ). Subgroup analyses were not possible given too few studies in each subgroup. The pooled OR for the case-control studies was 3.05 (95% CI = 1.11-8.37), and also had substantial heterogeneity (I² = 90.5%, P  < 0.001) (Figure S2 in the Online Supplementary Document ).

Tuberculosis

Of the five identified reviews on diabetes and TB [33-37], one focused on latent TB (LTB) [36] and one on multi-drug resistant TB (MDR-TB) [37]. We performed narrative synthesis on two reviews [33,34] and secondary meta-analysis on three [35-37]. In total, these reviews included 300 659 participants from 22 LMICs amongst 47 primary studies (16 cohort, 15 case-control, and 17 cross-sectional). One primary study from McMurry et al [34] reported on two separate cross-sectional studies from different countries, thus both were included in the study count. Generally, diabetes was shown to be associated with an increased risk of developing TB, although the association for LTB was not statistically significant and only included two primary studies.

Bailey and Ayles [33] examined three primary studies on the association between diabetes and TB in African settings, but did not perform a meta-analysis due to between-study heterogeneity. The study specific adjusted ORs from these primary studies suggest that diabetes is associated with increased odds of TB, although effect estimates ranged from 2.14 (95% CI = 1.32-3.46) to 19.3 (95% CI = 6.1-61.0). McMurry et al included 17 primary studies and found that prevalence of TB amongst people with diabetes ranged from 0.1% to 4.9% [34].

Al-Rifai et al identified nine of their primary studies as being from LMICs, but based on the World Bank income classification used in our review, ten were eligible and thus included in our secondary meta-analysis [35]. Since different effect measures were used across the primary studies, they could not all be combined. After pooling three primary studies which reported a hazard ratio, rate ratio, and relative risk, we found that diabetes was associated with a greater than 3-fold increased risk of TB (summary estimate = 3.40, 95% CI = 1.53-7.56), with substantial heterogeneity between studies (I²  = 87.4%, P < 0.001; Figure 5 , Panel A). Similarly, after pooling seven studies (five case-control, one cross-sectional, and one cohort) reporting odds ratios, we found that diabetes was associated with a 3-fold increased odds of TB (summary OR = 3.04, 95% CI = 2.29-4.03), with low heterogeneity between studies (I² = 23.5%, P  = 0.250; Figure 5 , Panel B). There was no evidence of publication bias (Figure S3 in the Online Supplementary Document ).

Figure 5.  Secondary meta-analysis of studies included in Al-Rifai et al. [35] showing the study-specific and summary estimates of tuberculosis among people with diabetes. Panel A. Hazard, risk, and rate. Panel B. Odds ratio.

Lee et al [36] only included two cross-sectional studies on the association between diabetes and LTB in LMICs. The pooled summary suggested no significant association between diabetes and LTB (OR = 1.16, 95% CI = 0.97–1.37; heterogeneity: I²  = 0.0%, P = 0.608; Figure S4 in the Online Supplementary Document ).

Finally, Tegegne et al. [37] included 17 estimates from 15 studies (six cohort, six case-control, three cross-sectional) from 10 LMICs across all WHO regions. Two studies [42,43] reported estimates separately for those with previously diagnosed TB and newly diagnosed TB. Our secondary meta-analysis of all 17 estimates found that people with vs without diabetes have 78% increased odds (OR = 1.78, 95% CI = 1.26-2.52) of developing MDR-TB, with moderate heterogeneity between studies (I² = 71.1%, P < 0.001). Our study design subgroup analysis found a higher OR and low heterogeneity for case-control studies (OR = 2.13, 95% CI = 1.37-3.29; I² = 26.6%, P  = 0.235) compared to cross-sectional and cohort studies ( Figure 6 , Panel A). Our income-level subgroup analysis found similar OR estimates between lower and upper-middle income countries, with substantial and moderate heterogeneity, respectively ( Figure 6 , Panel B). We conducted a post-hoc analysis on confounder adjustment and found that the pooled estimate was higher for those studies that adjusted for at least one confounder, of which age, sex, smoking, and HIV status were the most common, compared to studies that did not adjust for any other factors (OR = 2.57, 95% CI = 1.92-3.43 vs OR = 1.16, 95% CI = 0.70-1.92; Figure 6 , Panel C). We also conducted a post-hoc analysis by WHO region to see if there may be regional differences. Heterogeneity remained moderate to substantial across regions, with the exception of the European region which showed no heterogeneity (I²  = 0.0%, P = 0.715). Only the estimates for South-East Asia and Europe were statistically significant; these regions also had the strongest association between diabetes and MDR-TB ( Figure 6 , Panel D). There was no evidence of publication bias (Figure S5 in the Online Supplementary Document ).

Figure 6.  Subgroup analyses from secondary meta-analysis of studies included in Tegegne et al. [37] showing the study-specific and summary estimates of the odds of multi-drug resistant tuberculosis among people with diabetes. Panel A. Study design. Panel B. Country income level. Panel C. Confounder adjustment. Panel D. WHO region.