A research protocol was developed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework [26].

One of the authors (W Ming) independently screened the titles and abstracts of identified citations for potential eligibility. Two authors (W Ming and J Hirst) independently examined the full-text articles of eligible papers and extracted information about the exposures and outcomes using a predefined data extraction table.

The primary outcome was maternal glycemic control. Owing to the challenges in quantifying glycemic control and lack of consensus in measuring and reporting this outcome in pregnancy, we chose to define glycemic control with respect to mean blood glucose during pregnancy monitoring (total, fasting, or 1-h or 2-h post-prandial blood, expressed in mmol/L), and final recorded HbA1c in pregnancy (reported as both % and mmol/mol).

Secondary outcomes included insulin usage (ie, the final dose of insulin in units), mode of delivery (vaginal delivery or cesarean section), and the proportion of cases of shoulder dystocia at birth. As poor glycemic control in pregnancy is associated with increased fetal size, we also compared differences in fetal size as defined by mean birth weight, rates of macrosomia (defined as birth weight >4000 g), and the proportion of babies that were large for gestational age (LGA; defined as birth weight for gestational age and gender >90th percentile using local references). Neonatal outcomes were also assessed including the need for any neonatal intensive care unit (NICU) admission, preterm birth <37 completed weeks, and neonatal hypoglycemia (defined as hypoglycemia requiring medical treatment; Multimedia Appendix 1).

Information on trial design and data on the primary and secondary outcomes were extracted by 2 reviewers, independently, using a predesigned Excel spreadsheet. Each trial was scored for the risk of bias using the Cochrane Collaboration Risk of Bias Tool. A third reviewer was available if there was a difference in opinion in interpreting the risk of bias.

The search and screening strategy is shown in Figure 1. Seven of the 54 trials selected for full-text review met the inclusion criteria, involving 579 women: 496 women with GDM (5 trials) [16,21,27-29] and 83 with type 1 diabetes (3 trials) [15,21,30]. The trial of Dalfra et al presented results separately for women with GDM and type 1 diabetes; thus, for analysis we present this trial stratified by diabetes type [21]. All trials were small in size, ranging from 19 to 203 women with a median of 57 (interquartile range 32-85). The 7 trials were all conducted in high-income countries (5 in Europe and 2 in North America). See Multimedia Appendix 1.

Technologies assessed were modem transmission of blood glucose readings to a central hospital computer [15], websites accessible to patients and health care professionals [17,18], a telephone system that translated blood glucose readings into audio tones to transmit them to a computer database [21], SMS transmissions of blood glucose readings to a central database [19], and a telemedicine hub located in the woman’s home, which transmitted data every week to a clinical team through the Internet [16]. All trials described the comparison groups as receiving “routine care.” However, this ranged from information given only about the method of blood glucose monitoring (ie, paper log books), to detailed descriptions of care pathways. The frequency of clinic visits differed between the trials, ranging from weekly to monthly visits.

Overall, all the trials displayed potential sources of methodological bias (Figures 2 and 3). Owing to the nature of the intervention, blinding of participants and health care providers was not possible and therefore we elected to not include this as part of the risk of bias assessment. Considering the method of randomization, 2 trials were found to be at low risk of bias, reporting the use of computerized stratified block randomization [15,16]. The remainder either used methods that were likely to be of high risk of bias, or did not report this component. Only 1 trial reported use of an adequate allocation concealment method [17]. Two trials gave a full description of participants and losses to follow-up during their trial [16,17]. Other trials reported losses to follow-up or postrandomization exclusions, which potentially may have affected the results. Reporting bias is the selective reporting of some outcomes but not others depending on the nature and direction of the results [31]. Only 1 included trial was judged to be at low risk of reporting bias [17], reporting a comprehensive range of glucose and clinical outcomes.

All the identified trials addressed clinical outcomes. Only 1 trial also reported maternal satisfaction; however, no comparative statistics were given between the intervention and the control groups. No trial presented any data on health economic outcomes.

HbA1c was the most commonly reported glycemic outcome in 5 trials [21,28-30]. Meta-analysis demonstrated a modest, but statistically significant, improvement in HbA1c associated with the use of a telemedicine. The mean HbA1c of women using telemedicine was 5.33% (SD 0.70) compared with 5.45% (SD 0.58) in the standard care group, representing a mean difference of −0.12% (95% CI −0.23% to −0.02%). When this comparison was limited to the 4 trials of women with GDM only, the difference was slightly greater [21,28,29]. The mean HbA1c of women with GDM using telemedicine was 5.23% (SD 0.70) compared with 5.37% (SD 0.61) in the standard care group, mean difference −0.14% (95% CI −0.25% to −0.04%). Three trials (175 women: 143 GDM and 32 type 1) compared the overall mean blood glucose levels between the intervention (telemedicine) and control (standard care) groups [27,28,30]. Meta-analysis of these trials demonstrated no evidence of difference in mean blood glucose levels; however, this was in keeping with the lack of difference in HbA1c also observed in these individual trials (Figure 4). Two of these trials reported differences between fasting and 2 h postprandial blood glucose, however, no significant difference was demonstrated between the groups [15,28]. One trial in women with type 1 diabetes reported the mean units of insulin used in each group [15]. For these 19 women, the telemedicine group used a greater total dose of insulin compared with standard care, 54 units (SD 7 units) and 36 units (SD 6 units), respectively.

Maternal outcomes were reported variously across the trials. A total of 4 trials (148 women using telemedicine and 145 controls) reported differences between rates of pregnancy-induced hypertension or preeclampsia [27-29]. In these trials, 7.5% of women overall had either of these conditions; however, there was no difference in the risk ratio between the telemedicine or control groups (Figure 5). When considering the mode of delivery, rates of Cesarean section were high in both the groups (50.0% in the telemedicine and 45.0% in the control) with no difference in the risk ratio. Only 2 trials (150 women) reported shoulder dystocia [29], however, with only 1 case of shoulder dystocia meta-analysis was not possible.

There was no significant difference between the groups with respect to mean birth weight. For the telemedicine group this was 3363 g (SD 115 g) and for the standard care group it was 3302 g (SD 121 g), with the mean gestational age at delivery of 37.9 weeks (SD 1.39 and 1.70) weeks in both groups (Figure 6). In the 2 trials that reported rates of macrosomia, there was no significant difference between the 2 groups, with an overall rate of 46% (129 cases and 159 controls, including 32 type 1 diabetic women) [21]. Three trials reported LGA as an outcome (124 women using telemedicine, and 119 with standard care) [27-29]. The overall prevalence of LGA in these 3 trials was 14.4%, with no difference demonstrated between the 2 groups.

There were 40 babies of 193 (20.7%) that were admitted to the NICU, however, this proportion was not significantly different between the 2 groups (Figure 7) [27,28]. Four trials reported the proportion of babies treated for neonatal hypoglycemia [27-29]. Overall, although 18.0% (18/100) of babies were treated for hypoglycemia, there was no evidence of differences between the intervention and control groups.

One trial reported mothers' satisfaction; however, this information was only presented for the intervention group. It reported that 90% (17/19) of women in the telemedicine group agreed or strongly agreed that they were satisfied with the system and would use it again.

No trials provided economic or health utilization analyses. One trial described differences in the duration of clinic visits, reporting that the telemedicine visits were 8 min less than those for standard care [19].

While telemedicine may offer a little advantage in terms of glycemic control in pregnant women with diabetes, there is insufficient evidence at this time to support that it has any effect on other clinical endpoints. However, the 7 trials included in our meta-analysis were all small, assessed different technologies and were deemed to contain moderate to high potential sources of methodological bias. Thus, while it is reassuring that there is no evidence of harm associated with telemedicine, it is not possible to conclude whether it offers genuine benefits.

The strengths of our review were the robust and rigorous search strategy used, identifying 3 additional trials those that had been considered in previous reviews of this topic [22,32]. We included pregnant women with all forms of diabetes, as the benefits of these technologies may not be limited to women with only GDM. There are some limitations of this review. With no agreement between the trials on the screening method and definition of GDM, or standard treatment protocols, patient groups across trials may not be precisely comparable. This is a problem for all research in GDM, and unifying clinical practice was part of the motivation behind the World Health Organization (WHO) or International Association of Diabetes in Pregnancy Study Group (IADPSG) guidelines for the diagnosis of GDM. With the rapid development of advances in communication technology, the same system has not been compared in different populations, and there has been no evidence of sustained scale-up of any of these technologies. This makes it difficult to recommend any 1 system over another. Despite these differences, as the underlying concept of remotely communicating blood glucose readings between outpatient visits was the same across these trials, therefore we deemed these trials as suitable for meta-analysis. A further limitation of this review is that some of the outcomes examined, such as Cesarean section rates, gestational age at delivery, and admission to the NICU, may be more influenced by local practice, rather than being directly influenced by the intervention itself. The recent initiative by the IADPSG to attempt to standardize reporting and outcomes in diabetic pregnancy research could be a valuable advance in the future to ensure results are more comparable in this area research [33].

As stated, the sample sizes in all these trials was small. In GDM research, trials powered to detect a difference in important adverse clinical outcomes generally need to recruit around 1000 women [34,35]. Even with meta-analysis therefore, this analysis is likely to be underpowered to detect any effect on severe less common perinatal outcomes, such as shoulder dystocia and death.

Two earlier reviews on telemedicine in the management of the pregnancy with GDM have previously been published [22,32]. Mastrogiannis et al presented a narrative synthesis of trials published on telemedicine for diabetic pregnancies published before 2012. The authors concluded that telemedicine solutions for pregnant women with diabetes could reduce patient visits and potentially improve quality of life, without increasing the risk of the maternal and neonatal outcomes. Rasekaba et al presented a meta-analysis limited to women with only GDM. They identified 4 publications from 1990 to 2013 and concluded that there was insufficient evidence to support clinical benefit. Other possible benefits, such as economic savings or patient satisfaction, were not assessed. Rasekaba concluded that there was a non-significant trend to better the HbA1c of the telemedicine group [22]. By identifying and including additional trials, we have been able to demonstrate that this difference is significant both for all women with any form of diabetes in pregnancy, and those with GDM only. However, this outcome should be interpreted with caution; iron deficiency and the increased turnover of red blood cells in pregnancy can make HbA1c a less sensitive indicator of glycemic control in pregnancy [36]. Similar to our findings, Rasekaba et al did not find any difference in other clinical outcomes [22].

Whereas there were no randomized trials that assessed maternal satisfaction, there is evidence from nonrandomized trials that telemedicine is associated with high levels of satisfaction. [24] Women report these systems to be convenient to use, particularly if they live far from the hospital, have other caring responsibilities, or need to take time off work to attend appointments [1,24,37]. These observations have only been assessed in women with GDM, and ideally should be confirmed for women with type 1 and 2 diabetes among whom a reduction in clinic visits may not be desirable, however greater supervision and support may be associated with benefits in itself.

There is limited evidence that fewer outpatient visits may be needed for women with GDM using telemedicine systems [38]. We did not identify any formal health economic evaluations of telemedicine systems for gestational diabetes. In nondiabetic pregnant women, an economic analysis was conducted for a telemonitoring system designed for high-risk pregnant women in the Netherlands [39]. The system evaluated involved self-measurement and transmission of blood pressure, temperature, cardiotocography (CTG), and weight and urine albumin to a clinical care provider. This system demonstrated a cost-benefit system when compared with in-patient care. However, as this system did not measure blood glucose, and as admission for blood glucose monitoring is rare in developed countries, results cannot be extrapolated to the diabetic pregnant population. In the nonpregnant population 1 meta-analysis has assessed the economic impact of telemedicine for adults with type 2 diabetes [39,40]. The authors identified 2 papers that assessed cost-effectiveness. However, owing to small numbers and lack of consistency in the reporting of costs and outcomes, no conclusion could be drawn. A comprehensive cost analysis of direct and indirect costs is ideally needed before widespread adoption of these systems into clinical care [41].

There is insufficient evidence to conclude that for women with diabetes in pregnancy, telemedicine systems produce superior clinical outcomes when compared with standard care. The reasons for this may be due to the existing studies being underpowered to detect small effect sizes and heterogeneity in the available technologies and methods by which they have been assessed. It may be however that the main benefits of these technologies are in improving maternal satisfaction and streamlining clinical care delivery. High-quality research is still needed to determine the efficacy, satisfaction, burden to pregnant women and to the health care system, and economic impact of telemedicine systems for this patient group.