Telemedicine to deliver diabetes care in low- and middle-income countries: a systematic review and meta-analysis

Abstract Objective To determine the effectiveness of telemedicine in the delivery of diabetes care in low- and middle-income countries. Methods We searched seven databases up to July 2020 for randomized controlled trials investigating the effectiveness of telemedicine in the delivery of diabetes care in low- and middle-income countries. We extracted data on the study characteristics, primary end-points and effect sizes of outcomes. Using random effects analyses, we ran a series of meta-analyses for both biochemical outcomes and related patient properties. Findings We included 31 interventions in our meta-analysis. We observed significant standardized mean differences of −0.38 for glycated haemoglobin (95% confidence interval, CI: −0.52 to −0.23; I2 = 86.70%), −0.20 for fasting blood sugar (95% CI: −0.32 to −0.08; I2 = 64.28%), 0.81 for adherence to treatment (95% CI: 0.19 to 1.42; I2 = 93.75%), 0.55 for diabetes knowledge (95% CI: −0.10 to 1.20; I2 = 92.65%) and 1.68 for self-efficacy (95% CI: 1.06 to 2.30; I2 = 97.15%). We observed no significant treatment effects for other outcomes, with standardized mean differences of −0.04 for body mass index (95% CI: −0.13 to 0.05; I2 = 35.94%), −0.06 for total cholesterol (95% CI: −0.16 to 0.04; I2 = 59.93%) and −0.02 for triglycerides (95% CI: −0.12 to 0.09; I2 = 0%). Interventions via telephone and short message service yielded the highest treatment effects compared with services based on telemetry and smartphone applications. Conclusion Although we determined that telemedicine is effective in improving several diabetes-related outcomes, the certainty of evidence was very low due to substantial heterogeneity and risk of bias.

Although studies to evaluate the clinical efficacy or cost-effectiveness of telemedicine interventions for diabetes have previously been conducted, 9,[11][12][13][14][15][16][17][18][19] these reviews were performed over a global context and were dominated by evidence from high-income countries (mostly the USA). Such previous syntheses have little relevance in determining the effectiveness of telemedicine for diabetes treatment in low-and middle-income countries.
In this systematic review and metaanalysis, we aim to address the lack of evidence synthesis efforts in telemedicine for diabetes care in low-and middle-income countries. We designed our study to: (i) estimate the effectiveness of telemedicine in improving biochemical outcomes and patient characteristics such as adherence to treatment and selfefficacy; (ii) evaluate the implementation processes involved in telemedicine interventions; and (iii) determine the certainty of evidence for telemedicinebased interventions for diabetes care in low-and middle-income countries.

Database search
We conducted this systematic review and meta-analysis as per Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines (data repository). 20, 21 We registered the review protocol before at the PROS-PERO database for systematic reviews (CRD42019141271). 22 Using a pretested search strategy (data repository), 21 we searched Web of Science, PubMed®, MEDLINE®, Global Health Library and Cochrane Central Register of Controlled Trials from their inception to August 2019. We also searched New York Academy of Medicine and POPLINE databases for grey literature. We applied no restrictions in terms of participant age, year of publication or region of study at this stage.
We conducted an updated database search up to July 2020. We augmented the previously used search strategy with the keywords health informatics, wireless devices, text messaging, clinical decision support system, mobile app, blood glucose, diabetes mellitus type 2 and T2DM (Box 1; available at: http:// www .who .int/ bulletin/ volumes/ 99/ 3/ 19 -250068).

Inclusion and exclusion criteria
We included telemedicine interventions encompassing various modes of delivery including, but not limited to, short message service (SMS), smartphone applications (apps), telemetry (devices that allow remote monitoring of health data by automatic transmission from patients to clinicians), telephone and web-based systems. We only included randomized controlled trials (RCTs) and cluster RCTs that tested the effectiveness of telemedicine-based interventions in type 1, type 2 and gestational diabetes. We included studies conducted among participants aged ≥ 18 years resident in low-and middle-income countries. We considered all studies that reported serum glycated haemoglobin (HbA1c) levels as either their primary or secondary outcomes; however, to formulate a clinical recommendation for telemedicine interventions, we considered HbA1c levels as a primary outcome in our review.
We excluded studies employing semi-experimental designs, such as pre-post studies or those lacking control groups. In the case where two separate studies were based on an overlapping data set, we only included the study with the most complete information. We excluded short papers, brief reports, abstracts, conference papers, posters and letters to editors because these types of publications often lack important quantitative information. We also excluded studies published in languages other than English because of our lack of translational resources.

Outcome choice
We included biochemical parameters such as body mass index (BMI) and serum levels of fasting blood sugar, HbA1c, total cholesterol and triglycerides. We also included non-biochemical characteristics such as adherence to treatment, knowledge of diabetes and self-efficacy. However, our primary outcome was HbA1c levels reported post-intervention.

Screening process
Two independent reviewers performed the screening process to retrieve bibliographic records of eligible studies in two phases. First, the reviewers screened the titles and abstracts of all bibliographic records against the inclusion and ex-clusion criteria. Second, the reviewers thoroughly read the full texts of these eligible titles to ensure that all inclusion and exclusion criteria were met. Studies judged to be eligible at this stage were then included in the qualitative and quantitative synthesis where applicable. In the case of any uncertainty or difference in opinion between the reviewers, the reviewers and first author reached a final decision through discussion.

Data extraction
We extracted data related to the characteristics of the studies, primary endpoints and effect sizes of outcomes. We also compiled data on intervention implementation processes, such as mode of delivery, theoretical orientation, rationale, materials, and development and training procedures. We closely examined the individual elements of the interventions according to the World Health Organization guidelines on digital interventions. 10 We assessed the risk of bias in the RCTs using the Cochrane tool against randomization, allocation concealment, blinding of outcome assessors, attrition rate, selective reporting and any other bias matrices (e.g. a priori protocol registration and statistical power). 23 It was not possible to assess the rigour of the blinding procedures of participants and personnel in this review because of the nature of telemedicine-based interventions. We assessed certainty of evidence for telemedicine-based interventions using the grading of recommendations, assessment, development and evaluations (GRADE) guidelines. 24 We conducted the grading of evidence only for the HbA1c end-point as this was our primary outcome. We downgraded evidence by 1 or 2 points according to the presence and extent of flaws such as bias, inconsistency, publication bias, imprecision or indirectness related to patients, outcomes and interventions. 24 For quantitative synthesis, we assessed the mean and the standard deviation (SD) of outcomes for both the intervention and active control groups. 25 We used categorical data, such as frequency of events and sample size, for outcomes that lacked quantitative data in the form of mean and SD. 25 We then used these raw data to calculate standardized mean differences and their SDs for each outcome reported in the in-Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.
cluded studies. Because of the expected methodological and statistical heterogeneity, we calculated pooled effect sizes by performing random effects analyses. 25 We present these pooled effect sizes as forest plots depicting standardized mean differences and 95% confidence intervals (CIs). We ran a sensitivity analysis using the single-study knockout approach to assess the contribution of each study to the pooled effect size. We assessed publication bias by creating Begg funnel plots and performing Egger regression analyses (considered significant at P ≤ 0.1). 25 For outcomes demonstrating significant publication bias, we calculated adjusted standardized mean differences using the Duval and Tweedie trim-and-fill method. 25 For each outcome, we performed a series of subgroup analyses to quantify the specific difference in effect sizes for each mode of delivery. 26

Intervention characteristics
Out of 376 studies retrieved in our initial database search, we included a total of 22 studies describing 23 interventions in the qualitative analysis, and 21 studies describing 22 interventions in the quantitative synthesis (data repository). 21 In our updated database search, we retrieved 647 non-duplicate bibliographic records; we included 31 of these studies describing 32 interventions (an increase of nine compared with our initial search) in our synthesis ( Fig. 1). All included studies were published between 2010 and 2020. A breakdown of the studies by country revealed that the highest number of studies were conducted in China (eight studies), followed by India (five studies), the Islamic Republic of Iran (five studies), Malaysia (three studies), Turkey (two studies) and South Africa (two studies). A single RCT was conducted in each of Brazil, Egypt, Mexico, Mongolia and Pakistan. One of the studies described an intervention conducted at several sites in Cambodia, Congo and the Philippines. The minimum sample size of the RCTs ranged from 60 15,27,28 to 3324. 14 We list the properties of the 31 studies in Table 1, and the values for the outcomes considered that were extracted from the studies in the data repository. 21

Risk of bias
Our assessment of the risk of bias revealed that 19 studies had a high risk of bias, while 12 studies had a low risk of bias. Within the studies with a low risk of bias, the highest number of individual matrices were found to have a low risk of bias across matrices of random sequence generation (24 matrices), followed by attrition bias (20 matrices), other risk of biases (10 matrices), allocation concealment (eight matrices) and blinding of outcome assessors (six matrices; data repository). 21

Intervention strategies
Our included interventions varied in their strategies for the management of diabetes. We identified five modes of intervention delivery, through either smartphone apps (five studies), 14,37,47,50,51 SMS (nine studies), 13,30,36,39,42,46,49,53,54 telemetry (five studies), 32,38,40,48,52 telephone (10 studies) 13,15,18,27,29,33,40,41,44,48 and web-based systems including video conferencing (four studies). 31,35,43,45 Most studies focused on a single mode of telemedicine delivery; however, one study considered both telephone and SMS and another study investigated the use of both telephone and telemetry. 13,40 Major strategies included health record-keeping, follow-ups, reminders for follow-ups and logins, psychoeducation, glucose monitoring, monitoring prompts for serum glucose levels, pervasive alerts and online consultations ( Table 1 and data repository). 21 We did not observe any trend in the emergence of unique technologies; however, smartphone app-based interventions began to be tested from 2017. We provide details of individual interventions in the data repository. 21  Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.

Meta-analysis
We conducted our meta-analysis of the 30 studies, including 31 interventions, across eight outcomes. Our primary outcome, HbA1c levels, was reported in a total of 27 studies (28 interventions; Fig. 2). BMI was reported in 18 studies (19 interventions; Fig. 3 Adherence to treatment was reported in eight studies (Fig. 6), knowledge of diabetes in seven studies (Fig. 6) and self-efficacy in seven studies (eight interventions; Fig. 6).
We observed a significant treatment effect among several outcomes, with standardized mean differences of −0.38 for HbA1c (95% CI: −0.52 to −0.23; n = 7703; I 2 = 86.70%), −0.20 for fasting blood sugar (95% CI: −0.32 to −0.08; n = 5524; I 2 = 64.28%), 0.81 for adherence to treatment (95% CI: 0. 19   Funnel plots and Egger regression analyses revealed that publication bias was significant for the outcomes of HbA1c, knowledge of diabetes, fast-ing blood sugar and self-efficacy (data repository). 21 Publication bias was non-significant in outcomes including BMI, total cholesterol, triglycerides and adherence to treatment (data repository). 21 Our sensitivity analysis did not reveal any significant changes in the effect sizes of these outcomes.
Our subgroup analysis based on mode of intervention delivery revealed that the only outcomes that yielded statistical significance were BMI and self-efficacy. Telephone-and SMS-based telemedicine interventions yielded the highest treatment effects when compared with telemetry and smartphonebased services for a range of outcomes (Table 2; available at: http:// www .who .int/ bulletin/ volumes/ 99/ 3/ 19 -250068).
Our meta-regression analysis did not reveal any association between effect size and either quality assessment score or year of study for the HbA1c outcome (P > 0.05; data repository). 21 We could not run a meta-regression analysis based on sex, mean age and time at which the measurement of outcomes was reported because of missing data.

GRADE evidence
We assessed the certainty of evidence for the efficacy of telemedicine-based interventions for diabetes management in low-and middle-income countries, according to GRADE guidelines, for the critical outcome of HbA1c levels. We graded the certainty of evidence as very low because of substantial heterogeneity, publication bias and risk of bias (data repository). 21

Discussion
Our meta-analysis showing that telemedicine-based interventions are effective in improving serum levels of HbA1c and fasting blood sugar, adherence to treatment and self-efficacy is consistent with several individual RCTs conducted in both high-income countries and in low-and middle-income countries. 13,16,18,29,35,37,55 Our results are also in accordance with previously conducted meta-analyses; for example, in the 2014 global study of the clinical effectiveness of telemedicine in reducing serum levels of HbA1c, researchers reported a small but statistically significant decrease (standardized mean difference: −0.37). 19 Our results are also corroborated by the 2017 global systematic review, which showed moderate reductions in HbA1c post-intervention with reduced effect sizes at follow-up. 11 Telemedicine interventions were also found to be cost-effective for diabetes management; for example, for retinal screening alone, telemedicine interventions were reported to yield 113.48-3828.46 quality-adjusted life years. 12 Although an  Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.
outcome of interest, we could not find any cost-effectiveness data applicable to low-and middle-income countries.
Despite the encouraging effect sizes for a range of outcomes, we found no improvement in BMI or serum levels of total cholesterol or triglycerides. It was not possible to corroborate these results as our literature review did not yield any similar meta-analytic reviews exploring these indicators. The statistical non-significance or poor treatment effects of these interventions for these outcomes can be explained, however. First, most of these interventions were developed to target single outcomes, such as adherence to treatment, HbA1c serum levels or dietary behaviour. 13,55,56 None of the interventions focused on the measurement of lipid profile or BMI, and none assessed knowledge of diabetes (telemetric or otherwise). Second, the sample size calculations of these trials were based on improvement in HbA1c levels. We recommend that future interventions should be developed as comprehensive packages providing sessions on diet, physical exercise, monitoring of HbA1c levels and adherence to treatment.
For several outcomes, including BMI, adherence to treatment and self-efficacy, we found telephone-and SMS-based interventions (i.e. low-tech services) to be more effective than telemetry, smartphone apps or other webbased interventions. However, previous studies did not compare outcomes for different telemedicine delivery modes, meaning that we could not find any corroboratory or contradictory evidence. We suggest that our observed increased effectiveness of telephone-and SMSbased interventions may be attributable to the improved relationship between health-care professional and patient obtainable through these media. 57 Importantly, the fact that these delivery modes performed as well as or better than more advanced apps and software means that they can be adopted with confidence in low-resource settings.
Regarding other modes of delivery, smartphone apps developed in future studies should of course be user-friendly with a patient-centredness perspective, and considerate of the computer literacy levels of patients. Future researchers should consider conducting participatory approaches, including pilot surveys, design science, cost-effectiveness studies, and knowledge, attitude and practice surveys to explore levels of computer literacy among consumers.
Our review has several limitations. First, our search was limited to only five major academic databases and two minor grey-literature databases. Several additional databases, such as Embase and APA PsycINFO®, also warrant searching for potential articles. Although the databases Web of Science, PubMed® and Cochrane registry are highly inclusive, literature may have been missed. Our results are also limited by a high statistical and methodological heterogeneity across a few outcomes. For instance, outcomes such as self-efficacy and adherence to treatment were assessed using heterogeneous questionnaires. The heterogeneity in the outcomes could also be explained by different modes of intervention delivery, geographical regions and intervention contents.
Despite the above limitations, by synthesizing evidence from 31 studies conducted in low-and middle-income countries, we have provided more inclusive evidence in terms of number of articles than the previously published key reviews. 11,19 Our study also benefited from the exploration of additional outcomes, including a variety of biochemical indicators.
Our meta-analysis revealed a very low certainty of evidence that telemedicine interventions were effective in improving

Systematic reviews
Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.
diabetes management in low-and middleincome countries. Higher-quality RCTs are required before a solid recommendation for the use of telemedicine-based interventions can be made. We recommend that future interventions should be designed to address both the biological and socioeconomic determinants of diabetes. Studies that explore the evaluation, feasibility or acceptability of data are important in the scaling up of interventions.
To conclude, telemedicine-based services are frequently considered to be costly to the health system, but there should be more reviews addressing the cost-effectiveness of implementing and integrating such interventions within national health systems. Telemedicine would benefit from integration with the community-based health system with the support of community health workers. Despite barriers to this integration, telemedicine could improve the accessibility and quality of health-care services, improve personnel training and management processes, and optimize the use of epidemiological and clinical data. 58

Systematic reviews
Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.
Conclusión Aunque se determinó que la telemedicina es eficaz para mejorar varios resultados relacionados con la diabetes, la certeza de las pruebas fue muy baja debido a la considerable heterogeneidad y el riesgo de sesgo.

Systematic reviews
Telemedicine for diabetes in low-and middle-income countries Jorge César Correia et al.