The search, which was conducted in late August 2015, included studies published from January 1, 2000 to July 31, 2015 using MEDLINE, PsycINFO, EMBASE, CINAHL, and CENTRAL. Keywords and inclusion criteria from recent related reviews were examined [15,34,47,48] and a medical librarian was consulted to verify the search strategy (eg, keywords, choice of databases). Abstract and full-text screening were independently performed by two reviewers (LE, KB, or DW), with discrepancies resolved through discussion. Multiple outputs from the same dataset were linked for included studies and imported into Endnote X7.

Included studies were peer-reviewed English-language journal articles on telehealth RCTs recruiting adult (≥18 years) type 2 diabetes patients from Western countries where English is both an official and the majority language (ie, Australia, Canada, Ireland, New Zealand, United Kingdom, United States). All other research designs and publication types were excluded, as were RCTs that included type 1 or gestational diabetes patients, or that had explicitly targeted one or more ethnic minority groups in their recruitment strategy. Interventions could comprise any telehealth medium designed to treat or improve type 2 diabetes as the primary condition. Studies focusing on secondary diabetes-related complications (eg, retinopathy) or mental health were excluded as were telehealth interventions solely targeting health professionals rather than patients.

The data extraction form Multimedia Appendix 1 was developed using Cochrane guidelines [44] to describe study details, participant demographics, and intervention characteristics. Outcome data were extracted using Microsoft Excel and independently checked by two reviewers (DH, DW). The Cochrane Collaboration’s risk of bias (ROB) tool [49] was adapted to assess all 79 included articles using Cochrane’s Review Manager software (RevMan, The Cochrane Collaboration, Copenhagen, Denmark). Two of three authors (LE, DH, DW) independently evaluated each article for low, unclear, or high ROB, with discrepancies resolved through discussion. The assessment criteria were random sequence generation (selection bias), allocation concealment (selection bias), blinding of participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), selective reporting (reporting bias), and other sources of bias (other bias).

Due to the nature of telehealth interventions, patient and personnel blinding is generally unfeasible. Hence, for the purpose of this review, high risk of performance bias was interpreted as situations where unblinded research personnel interacted with participants across study groups, allowing for differential treatment. Assessment of detection bias focused on primary outcome detection. Studies that used an objective primary outcome measure (eg, laboratory-based blood test, administrative record of number of hospital visits) were assessed as low ROB because knowledge of participants’ allocation is unlikely to seriously affect the outcome. Conversely, studies where the primary outcome was subjectively assessed (eg, self-report measures) were deemed high ROB.

Data analysis of included studies addressed the following primary outcomes:

1. Proportion of studies that report on the ethnic composition of recruited participants and, where available, the overall prevalence of ethnic minority patients (between-study median and range);

2. Characteristics of studies that recruit a high proportion (≥30%) of ethnic minority participants (eg, telehealth medium, access to translation);

3. Proportion of studies that include English language proficiency or reading and writing literacy as a participant-screening criterion and, where available, the ways in which proficiency/print literacy is operationalized as a screening criterion; and

4. Language-related reasons for patient exclusion, if given (eg, informed consent, lack of resources).

In line with categorizations of race and ethnicity in the United States census [50] and conceptualizations of nonwhite or non-Caucasian respondents in other majority English-speaking countries (eg, United Kingdom [51]), ethnic minorities were defined as those of nonwhite ethnicity, including Hispanics, who may or may not be newcomers to the host country. In the case of studies reporting only the proportion of white patients recruited, all other participants were assumed to be ethnic minorities. In other studies that listed the proportion of participants belonging to an “other” group, “other” was interpreted as patients from an ethnic minority background not specified elsewhere in the study.

A narrative synthesis [52] was conducted to examine the characteristics of RCTs that reported a 30% or greater threshold of ethnic minority participants as a proportion of the total sample (considered relatively high), which is in line with the median prevalence of minority recruitment reported in earlier related reviews [3,33,34]. We would also note that this threshold is close to the proportion of ethnic minorities in the United States population. In the 2015 census, the “white alone, not Hispanic or Latino” category was reported at 61.6%, which implies that the remaining 38.4% are ethnic minorities [53]. After systematically extracting and tabulating the data, groupings and textual descriptions were used to explore heterogeneity within those studies and between studies with higher and lower proportions of ethnic minority participants. Exploration of relationships in the data was iteratively conducted to reveal factors that may promote or impede ethnic minority recruitment.

Finally, a post hoc subgroup analysis was conducted for studies that had explicitly targeted one or more ethnic minority groups as part of their recruitment strategy and that had been excluded from the systematic review for that reason only [54-65]. This additional retrospective analysis of ethnically targeted studies was undertaken to further examine recruitment strategies and features of trials that specifically recruited ethnic minority participants. This analysis involved the same methods as the synthesis described previously, but did not contribute to calculations reported in the main analysis.

The search yielded 2332 records after removing duplicates, which were submitted to abstract screening. After assessing 212 full-length articles for eligibility, 79 articles, consisting of 58 discrete RCTs, met the inclusion criteria (see Figure 1).

Multimedia Appendix 2 summarizes the characteristics of included studies in the review. The total number of participants across included RCTs was 12,916, with sample sizes ranging from 14 to 1665 (median 160). Nearly three-quarters (43/58) were recruited from the United States, whereas the rest were carried out in Australia (5/58), Canada (5/58), the United Kingdom (4/58), and one in both the United States and Canada. The studies involved a wide range of telehealth media and devices, the most popular of which was phone-based telehealth interventions. Other frequently used telehealth tools included Internet technologies (eg, static and interactive webpages, email, instant messaging), computerized self-management programs, and glucose meters integrated with mobile apps. Several studies combined different media and communication types as part of the intervention, including electronic medical records, educational websites, home monitoring, and videoconferencing (eg, [66]).

Parts A and B of Multimedia Appendix 3 show the ROB summary table and graph for included studies. Of the 79 articles, 34 described an appropriate randomization procedure, two [67,68] reported inadequate randomization, and the remainder provided insufficient detail for assessing risk of selection bias (judged unclear). Similarly, allocation concealment reporting was frequently insufficient to determine the ROB (eg, not clear whether allocation envelopes were opaque and sequentially numbered [69]). Most studies reported designs intended to reduce performance bias (within the parameters of unfeasible patient and interventionist blinding), with studies judged as having a high ROB in cases when nonblinded staff delivered both the telehealth intervention and usual care [69-83] and also collected follow-up measures [84,85]. Included studies employed both subjective and objective outcome measures, with objective glycated hemoglobin (HbA_1c) levels being the most common primary outcome. Approximately half of the studies included patient attrition information, with unclear ROB assigned when reasons for attrition were not given, attrition was not broken down by group allocation, or it was unclear how substantial missing data were dealt with in data analysis. Several studies were assessed as having a high ROB due to a large difference in attrition between study arms that was attributed to the intervention [86,87] or that had only presented outcomes for patients who had completed the intervention and/or all follow-up measures [88-91]. There was generally a low risk of reporting bias, with most studies reporting on prespecified outcomes. A high ROB was noted in studies where reported outcomes deviated from the protocol or specified methodologies [70,85,91-97] or insufficiently reported statistical or summary data [67,68,71,72,83,98-105]. Most studies were free of other sources of bias. However, some failed to report treatment dosage (eg, frequency of calls [74,90,106,107]) and its effect on outcomes even when adherence was low. Others were pilot studies or had small convenience samples, limiting their generalizability [71,85,95,101,108,109].

A post hoc subgroup analysis was conducted for 12 articles comprising 11 RCTs that had been excluded during full-text screening [54-65]. The findings echoed those of the narrative synthesis (see Table 2). All ethnically targeted studies took place in urban American settings, with nearly half (5/11) reportedly recruiting from economically deprived or medically underserved areas [56,57,60,61,65]. More than half (6/11) used phone calls as part of the intervention, although other telehealth media were also used. Although only one study offered patients intervention-related equipment (laptop and telehealth peripherals) [57], six offered financial remuneration ranging from US $40 to US $60 cash or vouchers [55,59,61-63,65], with one trial additionally offering some patients up to US $200 for reducing their HbA_1c levels by a prespecified amount [62]. The ethnically targeted studies frequently reported tailoring interventions to their target demographic, including offering written and spoken aspects of the intervention in a language other than English [56,61,64,65], designing interventions through consultation and feedback from minority groups [55-58,60,65], and using interventionists drawn from the target cultural communities [55,61,64,65]. This latter point included building input from minorities into the recruitment strategy and/or piloting materials and procedures with them.

This systematic review investigated research reporting practices and prevalence estimates of ethnic minority participation in telehealth diabetes RCTs, extending previous reviews by including a broader range of telehealth technologies targeting type 2 diabetes. Nearly 66% of included studies reported on the ethnic composition of their samples. Although this proportion is higher than in previous reviews [3,33,34], it confirms the underreporting of ethnicity in peer-reviewed journal articles. However, compared to the 2006 review that had also excluded studies recruiting participants from a single ethnic minority background [33], these results yielded lower median participation of ethnic minority patients as a proportion of the total sample (23.5% vs 39%).

All RCTs with 30% or greater ethnic minority recruitment were US-based, mostly recruited from urban areas, and frequently described recruited patients as low-income, socially deprived, or with little or no health care coverage. Participant remuneration or free telehealth monitoring devices incentivized participation in nearly 60% of these studies. Although the United States has a high proportion of ethnic minorities compared to other countries included in the review [147], its dominance as the setting for all studies meeting the 30% or greater cut-off for the narrative review and for all ethnically targeted studies in the post hoc subgroup analysis suggests that the American National Institutes of Health (NIH) Policy and Guidelines on the Inclusion of Women and Minorities as Subjects in Clinical Research [148] has been influential in promoting minority representation in trials. Overall, minority participation rates tend to be higher in the United States, which has legislation to mandate their inclusion, than in contexts where no such legislation exists (eg, United Kingdom) [26]. The availability of trial materials in multiple languages was also a recurrent feature of studies recruiting substantial numbers of ethnic minority participants and often occurred in conjunction with trial materials that accounted for cultural factors or presumed literacy levels. This suggests that having a language concordant interventionist and embedding input from members of target communities into trial materials could have positive effects on the recruitment of groups who may otherwise be underrepresented [149,150]. This supports existing research on strategies to optimize the recruitment of underserved patients, which reveals ways of working holistically to alleviate individual and external factors that impede participation [146,151]. The emphasis on patient and public involvement in research aligns with such practices [29,152].

Half of the included studies listed language proficiency among the patient eligibility criteria, although less than 5% provided a reason for including or excluding patients on this basis. The role of language in participant screening was described in different ways across studies, with emphasis either placed on different combinations of skills (speaking, listening, reading, writing) to reflect the nature of the intervention, or on patients’ status as primary speakers of English or another language where offered. There was no evidence of the use of objective measures or instruments or of a common procedure across studies to assess whether patients had the requisite language proficiency to participate and, in individual studies, this level of detail was not given. Thus, it was unclear how language-related determinations about inclusion or exclusion were made—that is, how the language-screening criterion as stated was operationalized in arriving at eligibility decisions.

This review has several limitations. First, only English-language articles that recruited patients from Western countries where English is both an official and the dominant language were considered. Findings may be different in reviews focusing on recruitment from other contexts or on other world languages. Second, ethnic minorities are not a monolith, and language barriers to engaging with the intervention are likely to be different for newcomers to a country (eg, migrants), who may have little knowledge of English, than for later generations of native English speakers who are visible minorities (eg, African Americans in the United States) [24]. It is often difficult to disentangle ethnicity from language in secondary data analysis because the language background of participants (ie, proportion of native English speakers and heritage language speakers) is often not reported as a separate category and, in some countries, is not captured in census data [40]. Third, and related to this, we used the ethnicity categories reported in the original studies in data extraction. However, race and ethnicity are complex constructs that frequently conflate social identity with other factors, such as genetic or biological characteristics (particularly as connoted in the former term), geographical origin, cultural practices, or religious persuasion [31]. Categorizations were inconsistent across studies included in this review, making comparisons between and within countries difficult. Fourth, the prevalence of minority patients in the area(s) from which they were recruited was not possible to collate, with several studies recruiting from different geographic sites and not reporting on local population or, in some cases, on the ethnic breakdown of their recruited sample. Finally, peer-reviewed articles were the sole publication type considered in this study, which was consistent with the goal of elucidating research reporting practices in academic journals. That is, the grey literature was not examined and study authors were not contacted to provide further information than was included in their published articles, taking into account linked publications. The brief descriptions of study setting, design, and recruitment in the included telehealth RCTs provided comparatively little qualitative data for the narrative synthesis. This has resulted in a relatively simple theoretical account of the trial features supporting ethnic minority participation, which would need to be corroborated through a more in-depth examination of key variables in subsequent research. Nonetheless, with 79 articles and 58 RCTs included, this review is comprehensive in its account of research reporting and includes more studies than the three previous telehealth diabetes reviews combined [3,33,34].

Despite the link between new technologies and improved outcomes, mixed evidence regarding reducing disparities from the research literature [122,129] suggests that telehealth has yet to fulfill its potential of being truly accessible to and effective for ethnic minority diabetes patients [15]. Cultural and linguistic tailoring to a diverse demographic and offering translation and interpretation services where possible could extend the benefits of telehealth type 2 diabetes interventions to a wider cross-section of patients, thereby promoting more equitable access to health care [43,153].

Findings from this and earlier systematic reviews suggest that between a third and half of telehealth diabetes trials provide no information on the ethnic composition of their samples. Further, other demographic characteristics, such as socioeconomic status, are not consistently reported across studies (see Multimedia Appendix 2). Providing data on participants’ gender and age, but no information on ethnicity for either the sample or the local target population is insufficient for assessing the external validity of the findings. Ethnicity information is necessary to evaluate claims about telehealth’s accessibility to all patients (and not just a subsection of the population), including its ability to foster social inclusion through the uptake of services.

The Consolidated Standards of Reporting Trials (CONSORT) statement [154], which articulates guidelines for best practice in research reporting and which has been widely adopted, advocates reporting baseline demographic and clinical characteristics for each group in the RCT. However, this is prefaced on the assumption that these will be collected, and the statement does not specify which baseline variables should be captured and described. In the case of clinical baseline variables, further specification of which variables to target would appear to be dependent on the nature of the condition being examined (eg, HbA_1c for diabetes). However, this is not the case for sociodemographic variables, which apply regardless of the disease type being investigated. In trials targeting chronic conditions in which prevalence is known to vary by ethnicity and other social factors, further specification of the baseline demographic characteristics that should be reported on could reduce the variability in between-study reporting. This would assist in examining crucial relationships between sociodemographic variables and health outcomes. Journal editors and editorial boards could converge on the most important sociodemographic characteristics that should be reported (eg, in a checklist) to achieve greater consistency and help streamline such information across studies.

Finally, the fact that telehealth trials seek to recruit patients with an adequate level of language ability is understandable— telehealth interventions rely on effective spoken or written communication as a key part of the “treatment.” However, the absence of an objective or standardized measure for assessing whether patients have the requisite language ability to successfully engage with the intervention suggests that such decisions are likely being made on the basis of trial recruiters’ subjective judgments. This creates a risk of selection bias, as trial recruiters might exclude ethnic minority participants based on the subjective view that they will not be able to adequately take part in the intervention, despite actually having sufficient language proficiency [43]. For example, recruiters may misjudge having a perceptible foreign accent as evidence of poor language ability in instances when this does not actually impede communication [155,156]. Conversely, including patients who do not understand the nature of the intervention due to language barriers is problematic, not least for ethical reasons [157]. Future research could focus on the development and validation of a tool to provide trial recruiters with a simple, practical means of assessing language proficiency for trial participation to minimize the possibility of patients being unfairly excluded based on arbitrary judgments.