This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mhealth and uhealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.
Mobile health apps for diabetes self-management have different functions. However, the efficacy and safety of each function are not well studied, and no classification is available for these functions.
The aims of this study were to (1) develop and validate a taxonomy of apps for diabetes self-management, (2) investigate the glycemic efficacy of mobile app-based interventions among adults with diabetes in a systematic review of randomized controlled trials (RCTs), and (3) explore the contribution of different function to the effectiveness of entire app-based interventions using the taxonomy.
We developed a 3-axis taxonomy with columns of clinical modules, rows of functional modules and cells of functions with risk assessments. This taxonomy was validated by reviewing and classifying commercially available diabetes apps. We searched MEDLINE, EMBASE, the Cochrane Central Register of Controlled Trials, the Chinese Biomedical Literature Database, and ClinicalTrials.gov from January 2007 to May 2016. We included RCTs of adult outpatients with diabetes that compared using mobile app-based interventions with usual care alone. The mean differences (MDs) in hemoglobin A1c (HbA1c) concentrations and risk ratios of adverse events were pooled using a random-effects meta-analysis. After taxonomic classification, we performed exploratory subgroup analyses of the presence or absence of each module across the included app-based interventions.
Across 12 included trials involving 974 participants, using app-based interventions was associated with a clinically significant reduction of HbA1c (MD 0.48%, 95% CI 0.19%-0.78%) without excess adverse events. Larger HbA1c reductions were noted among patients with type 2 diabetes than those with type 1 diabetes (MD 0.67%, 95% CI 0.30%-1.03% vs MD 0.37%, 95% CI –0.12%-0.86%). Having a complication prevention module in app-based interventions was associated with a greater HbA1c reduction (with complication prevention: MD 1.31%, 95% CI 0.66%-1.96% vs without: MD 0.38%, 95% CI 0.09%-0.67%; intersubgroup
The use of mobile app-based interventions yields a clinically significant HbA1c reduction among adult outpatients with diabetes, especially among those with type 2 diabetes. Our study suggests that the clinical decision-making function needs further improvement and evaluation before being added to apps.
Diabetes mellitus poses enormous challenges to China’s health care system due to its mortality, prevalence, and costs. Of 8.3 million deaths in China in 2010, 37.3% (3.1 million) were attributable to cardiovascular disease, which was also one of the leading causes of disability-adjusted life-years [
Once diabetes is diagnosed, lifetime diabetes self-management is critical to glycemic control and is associated with the long-term prognosis for patients with diabetes. Diabetes self-management includes self-monitoring blood glucose, making healthy lifestyle choices (healthy eating, physical activity, tobacco cessation, weight management, and coping with stress), taking and managing medications, preventing diabetes complications (self-monitoring of foot health; active participation in screening for eye, foot, and renal complications; and immunizations), and setting self-selected behavioral goals [
Mobile apps are the computer programs or software installed on smart mobile devices, with computing and connectivity capability built right into an operating system. With the rapid and ongoing growth in wireless connectivity, more than 500 million Chinese were smartphone and apps users in 2016 [
In the iTunes App Store for iOS and Google Play for Android, diabetes is one of the top-ranked categories [
Despite their variety and complexity, apps for diabetes self-management always share a limited number of basic functions, which can be classified into several simple categories (eg, self-monitoring, education, alerts and reminders, and communication) [
To address functional efficacy, a classification of app functions is required [
The aims of this systematic review of RCTs were to (1) develop and validate a taxonomy of apps for diabetes self-management, (2) perform a meta-analysis investigating the effects of mobile app-based interventions on glycemic control in adults with diabetes, and (3) explore the contribution of different functions to the glycemic efficacy of entire app-based interventions using the taxonomy and subgroup analyses.
We developed a preliminary taxonomy based on previous classifications, evidence-based guidelines, and authoritative recommendations, and validated it by reviewing commercially available apps for diabetes management. The contents of the taxonomy were confirmed if all functions of the available apps could be classified. After validation, we proposed a final taxonomy for diabetes management apps.
The preliminary taxonomy was validated by a review of commercially available diabetes apps, as shown in
We searched MEDLINE, EMBASE, the Cochrane Central Register of Controlled Trials (CENTRAL), and the Chinese Biomedical Literature Database using the terms “diabetes mellitus,” “blood glucose,” “blood glucose self-monitoring,” “mobile applications,” and “cell phones” from January 1, 2007, to May 30, 2016. We also searched for ongoing studies via ClinicalTrials.gov and checked the reference lists of relevant reviews and trials.
We selected RCTs that compared mobile app-based interventions with standard care (free of app-based interventions) in adult outpatients with diabetes. Mobile app-based interventions were those that could provide real-time interactions with users through apps running on smart mobile devices.
Our primary outcome was the change in hemoglobin A1c (HbA1c) concentration (%) from baseline. Our secondary outcomes were severe hypoglycemia (defined as the need for assistance from another person or very low glucose concentrations; this was study specific, eg, <2 mmol/L) and any other adverse events. We did further quantitative meta-analyses of primary and secondary outcomes if relevant data were available.
We excluded studies without any available data on HbA1c. We also excluded studies if their participants were children, adolescents, or pregnant women who required different therapeutic strategies for a more challenging or strict glycemic control [
Two reviewers (YW and YD) independently screened titles and abstracts and then full texts to select eligible studies. Reviewers resolved disagreements through discussion or, if necessary, through discussion with an arbitrator (SL).
For each trial, 2 reviewers (YW and YD) independently extracted data using a structured abstraction form and classified functions according to our taxonomy. Then, 2 reviewers (YW and YD) independently used the Cochrane Collaboration’s tool to assess the risk of bias of included studies [
We used a random-effects meta-analysis to pool the overall mean difference (MD) of the HbA1c changes and the risk ratios of adverse events due to the possible clinical heterogeneity of each included study. For trials with unreported change-from-baseline standard deviations, we imputed by standard deviations at the baseline and at the end of the intervention using the formula
We designed a preliminary taxonomy with a functional axis, a clinical axis, and a risk axis as shown in
We developed the risk axis based on the US Food and Drug Administration (FDA) risk-based recommendation [
During validation, we identified 1559 apps by searching the iTunes App Store and Google Play and excluded 1414 apps that were duplicated, were not for diabetes-management, were without English or Chinese versions, and had not been updated for at least 5 years. The remaining 145 eligible apps were downloaded onto smart mobile devices. After excluding those without real-time interactions and designed solely for health care providers, we included 96 apps and classified them by the preliminary taxonomy. As we could well classify all functions among the included apps by the taxonomy, and we identified all modules in the taxonomy in the included apps, we proposed the final taxonomy after this validation (
Taxonomy of apps for diabetes self-management.
Functional modules | Diabetes management modules | ||||
Monitoringb | Medication managementc | Lifestyle modification | Complication prevention | Psychosocial care | |
Logb | ⊕⊖⊖Recording self-monitoring parametersd; |
⊕⊕⊖Recording used medications and side effects | ⊕⊖⊖Recording activities, diets, and weightf | ⊕⊖⊖Recording complication-related statusg; |
⊕⊖⊖Recording mood |
Structured display | ⊕⊖⊖Displaying data in a structured way | ||||
General education | ⊕⊖⊖Instructions for monitoring; |
⊕⊕⊖Diabetes process and treatment options; |
⊕⊖⊖Incorporating nutritional management and physical activity into lifestyle | ⊕⊕⊖Preventing, detecting, and handling acute complications and chronic complicationsh | ⊕⊖⊖Addressing psychosocial issues and promoting behavior change |
Personalized feedback | ⊕⊖⊖Reminding to monitor; |
⊕⊕⊖Reminding to take medications; |
⊕⊖⊖Reminding to eat healthily and be active; |
⊕⊖⊖Reminding to quit smoking, visit doctors, and prevent acute complications | N/Ak |
Communication | ⊕⊖⊖General communication, connecting users with their peers and families through social networking, chat forums, or websites; |
aRisk assessment of a function: low risk (⊕⊖⊖), potential risk (⊕⊕⊖), and high risk (⊕⊕⊕). The overall risk assessment of an app was determined by the highest risk of included functions.
bMonitoring and log are basic modules.
cMedications for diabetes include insulin, oral antidiabetic agents, aspirin, antihypertensives, lipid-lowering medications, and vaccines.
dSelf-monitoring parameters include blood glucose, blood pressure, heart rate, and pulse.
eOther medical parameters include cholesterol levels, hemoglobin A1c, urine test, and ketones.
fActivities include steps, duration, heart rate, and consumed calories; diets include food, water, nutritional values, carbohydrate counting, and calorie calculator; weight includes body mass index, body fat, and circumference.
gComplication-related status includes smoking, drinking, snoring, feet, eyes, teeth, and sensory status.
hAcute complications include hypoglycemia and hyperglycemia; chronic complications include cardiovascular disease and microvascular complications (ie, nephropathy, retinopathy, neuropathy).
iClinical decision making is recommending treatment (eg, oral agents and insulin) by algorithms alone without the participation of health care providers.
jSelf-management decision making is decision making on lifestyle modification by algorithms.
kN/A: not applicable.
We identified 3131 references using our search strategies and identified 544 references by checking the reference lists of relevant articles, 68 of which underwent a full-text review. This process excluded 55 studies, with the reasons listed in
Across the 13 included references, the HbA1c was obtained from 12 trials with 974 participants after a median follow-up period of 6 (range 3-12) months, and severe hypoglycemia was extracted from 4 trials of 346 participants after a median follow-up of 6 months. There were 5 trials that enrolled patients with type 1 diabetes mellitus (T1DM), 5 with type 2 diabetes mellitus (T2DM), and 2 with both types of diabetes.
Of the 12 included mobile app-based interventions, 1 is available in the iTunes Store and Google Play at the time of our study [
Various technologies were applied for data transmission between users and mobile devices. Across the 12 included trials, 6 (50%) used wireless transmission through Wi-Fi, Bluetooth, near-field communication, or public switched telephone network, 5 (42%) used manual entry, and 1 (8%) used wire transmission through a data port connection.
Of the 12 included app-based interventions, we determined 3 (25%) to be of high risk due to having a clinical decision-making function. The definition of the clinical decision-making function was recommending treatment (eg, oral agents and insulin) by algorithms alone without the participation of health care providers. We determined that the other 9 interventions (75%) carried potential risk.
Characteristics, modules, risk assessments, and technologies of the included mobile app-based interventions.
Study | Country | No. |
Diabetes type | Follow-up (months) | Mean (SD) HbA1ca, %: baseline; end; change | Intervention | FMb | DMMc | Risk assessmentd | Technology |
Hsu, 2016 [ |
US | Ie: 20/15; |
2 | 3 | I: 10.8 (1.0); 7.7 (1.6); –3.2 (1.5) |
Cloud-based diabetes management program | L, StD, GE, Co | M, MM, LM, CP | Potential | Wireless |
Baron, 2017 [ |
UK | I: 45/40; |
Both | 9 | I: 9.1 (1.8); 8.6 (1.6); |
Mobile telehealth | L, StD, GE, PF, Co | M, MM, LM | Potential | Wireless |
Drion, 2015 [ |
Netherlands | I: 31/30; |
1 | 3 | I: 7.73 (NRg); 7.91 (NR); |
Diabetes Under Control (DBEES) | L, StD | M, MM, LM | Potential | Manual entry |
Holmen, 2014 [ |
Norway | I: 51/39; |
2 | 12 | I: 8.1 (1.1); 7.8 (0.9); |
Few Touch Application (FTA) | L, StD, GE, PF, Co | M, LM | Potential | Wireless |
Waki, 2014 [ |
Japan | I: 27/24; |
2 | 3 | I: 7.1 (1.0); 6.7 (0.7); |
DialBetics | L, StD, GE, PF, Co | M, LM | Potential | Wireless |
Kirwan, 2013 [ |
Australia | I: 36/28; |
1 | 9 | I: 9.1 (1.2); 8.0 (0.7); |
Glucose Buddy | L, StD | M, MM, LM | Potential | Manual entry |
Rossi, 2013 [ |
Italy | I: 63/55; |
1 | 6 | I: 8.4 (NR); 7.9 (NR); –0.5 (NR); |
Diabetes Interactive Diary | L, PF, Co | M, MM, LM | High | Manual entry |
Charpentier, 2011 [ |
France | I: 60/56; |
1 | 6 | I: 9.2 (1.1); 8.6 (1.1); |
Diabeo system | L, StD, PF, Co | M, MM, LM | High | Manual entry |
Rossi, 2010 [ |
Italy | I: 67/58; |
1 | 6 | I: 8.2 (0.8); 7.8 (0.8); –0.4 (0.9); |
Diabetes Interactive Diary | L, PF, Co | M, MM, LM | High | Manual entry |
Yoo, 2009 [ |
Korea | I: 62/57; |
2 | 3 | I: 7.6 (0.9); 7.1 (0.8); |
Ubiquitous Chronic Disease Care (UCDC) system | L, GE, PF | M, LM | Potential | Wire |
Istepanian, 2009 [ |
UK | I: 72/NR; |
Both | 9 | I: 7.9 (1.5); 7.8 (NR); |
Mobile phone telemonitoring system | L, Co | M | Potential | Wireless |
Quinn, 2008 [ |
US | I: 15/13; |
2 | 3 | I: 9.5 (NR); 7.5 (NR); |
WellDoc Communications | L, StD, GE, PF, Co | M, MM, LM, CP | Potential | Wireless |
aHbA1c: hemoglobin A1c.
bFM: functional modules are communication (Co), general education (GE), log (L), personalized feedback (PF), and structured display (StD).
cDMM: diabetes management modules are complication prevention (CP), lifestyle modification (LM), monitoring (M), and medication management (MM).
dThe overall risk assessment of an intervention was determined by the highest risk of its functions.
eI: intervention group.
fC: control group.
gNR: not reported.
Study selection. CBM: Chinese Biomedical Literature Database; CENTRAL: Cochrane Central Register of Controlled Trials; CGM: continuous glucose monitoring; CSII: continuous subcutaneous insulin infusion; HbA1c: hemoglobin A1c; HCP: health care provider; PHR: personal health record.
Only 67% (8/12) of the trials adequately reported allocation sequence generation, and 58% (7/12) adequately reported concealing the allocation sequence. As an objective outcome, all trials adequately blinded the assessment of the primary outcome (HbA1c changes). The corresponding proportion for incomplete outcome data was 25% (3/12), for selective reporting was 25% (3/12), and for other sources of bias was 50% (6/12).
Risk of bias for the primary outcome (hemoglobin A1c changes): review authors’ judgments about each risk-of-bias item presented as percentages across all included studies.
Risk-of-bias summary for the primary outcome (hemoglobin A1c changes): review authors’ judgments about each risk-of-bias item for each included study.
The use of mobile app-based interventions was associated with a clinically significant HbA1c reduction of 0.48% (95% CI 0.19%-0.78%,
We performed a post hoc exploratory analysis for 5 trials enrolling patients with T1DM and 5 trials enrolling patients with T2DM. The use of app-based interventions did not achieve statistical significance among patients with T1DM (MD 0.37%, 95% CI –0.12%-0.86%,
Effects of app-based mobile health interventions on hemoglobin A1c (HbA1c). MD: mean difference.
Effects of app-based mobile health interventions on hemoglobin A1c (HbA1c) for patients with type 1 diabetes (T1DM) and type 2 diabetes (T2DM). MD: mean difference.
We noted a greater HbA1c reduction when interventions included a complication prevention module (with complication prevention: MD 1.31%, 95% CI 0.66%-1.96%,
For high-risk interventions with a clinical decision-making function, the reduction of HbA1c was 0.19% (95% CI –0.24%-0.63%,
Interventions using manual entry showed an associated lower HbA1c reduction without statistical significance (wire connection: MD 0.70%, 95% CI 0.33%-1.07% vs wireless connection: MD 0.53% CI 0.15%-0.92%,
Effects of modules, risks, and technologies of app-based mobile health interventions on hemoglobin A1c (HbA1c). MD: mean difference.
Adverse events were reported variably among the 5 included studies [
For severe hypoglycemia, 1 study reported significantly fewer episodes in the intervention group (0.33 vs 2.29 events/patient-year) [
Overall, we rated the quality of the evidence for severe hypoglycemia as low due to imprecision (wide confidence intervals including null effect) and study limitations (risk of bias in 4 trials), and as very low for adverse events owing to inconsistency (substantial diversity in the definitions of outcome measures), imprecision (small sample sizes and low event rates), and study limitations (risk of bias in 5 trials) (
As most commercially available apps for diabetes self-management were not tested by RCTs, both the patients and the clinicians needed indirect evidence to guide their assessment while choosing apps. The purposes of this review were to investigate the glycemic efficacy of mobile app-based interventions, and to explore the differential effectiveness of their functions. We could not use existing classifications for the functions of the app-based interventions because of inconsistency and incompleteness. As a result, we developed and validated a comprehensive taxonomy for the functions of diabetes self-management apps. To our knowledge, this is the first comprehensive taxonomy with clinical, functional, and risk axes, and this is the first review exploring the contribution of each function to the effectiveness of entire apps.
The meta-analysis of 12 RCTs demonstrated that app-based interventions were associated with a statistically and clinically significant HbA1c reduction of 0.48% (95% CI 0.19%-0.78%). We noted larger HbA1c reductions for patients with T2DM (MD 0.67%, 95% CI 0.30%-1.03%) than those with T1DM (MD 0.37%, 95% CI –0.12%-0.86%). The exploratory subgroup analyses showed that having a clinical decision-making function in app-based interventions was not associated with a greater HbA1c reduction (with clinical decision making: MD 0.19%, 95% CI –0.24%-0.63% vs without: MD 0.61%, 95% CI 0.27%-0.95%; intersubgroup
Consistent with previous reviews involving mobile app-based interventions [
Our study developed a 3-axis taxonomy for diabetes apps, with rows of functional modules, columns of diabetes management modules, and cells of functions with risk assessments. The functional, clinical, and risk axes were developed based on previous classifications, the ADA’s guidelines, and the FDA’s risk recommendation, respectively. The 3-axis design of the taxonomy is comprehensive and decreases the possibility of misclassification. Additionally, this 3-axis design is applicable for diseases other than diabetes by adjusting the modules in the clinical axis. The validation process guarantees that our taxonomy can be used to classify commercial diabetes apps. Differences in the detected effect sizes in subsequent subgroup analyses indicated the utility of our taxonomy.
Our taxonomy has some advantages. First, it is a comprehensive taxonomy with functional, clinical, and risk axes. The taxonomy permits subsequent exploratory subgroup analyses of multifunction apps, which give insights into the efficacy and risk of each module in diabetes apps. Comparatively, existing classifications appear to be incomplete or inconsistent. Previous classifications have mainly focused on the functions of apps, which, as a result, have made them applicable only for functional evaluation [
Second, our taxonomy can be of some help in the development and evolution of diabetes apps. App developers are usually technicians without a clinical background. As a result, the evidence-based guidelines for diabetes management are easily ignored during app development. For example, we found that complication prevention and psychosocial care were uncommon in the app-based interventions we examined. However, complication prevention behaviors and emotional well-being are associated with positive diabetes outcomes according to the guidelines [
Third, our taxonomy permits subsequent exploratory subgroup analyses of multifunction apps, which give insights into the efficacy and risk of each module in diabetes apps.
Our exploratory subgroup analyses suggested a limited efficacy of clinical decision making, which was defined as recommending treatment (eg, oral agents and insulin) by algorithms alone without the participation of health care providers and was determined to be high risk according to our taxonomy. Traditionally, clinical decisions are made during a face-to-face interview after a complete assessment. Built-in clinical decision support systems, however, are less likely to collect data and assess status as thoroughly as face-to-face consultations do. Without adequate data and well-designed algorithms, clinical decision-making functions can make inappropriate decisions and pose risks to patients [
Our subgroup analyses indicated that having a complication prevention module in the apps was associated with a greater HbA1c reduction. Complication prevention behaviors such as smoking cessation and hypoglycemia prevention are critical components of diabetes management according to current guidelines [
Having a lifestyle modification in app-based interventions was associated with a trend toward reduced HbA1c, as was having a general education module. The modules of lifestyle modification and general education may raise awareness of lifestyle change and self-management. Since these 2 modules pose limited risks to patients with diabetes, it might be reasonable to add lifestyle modification and general education to diabetes apps.
The data suggested limited glycemic efficacy of having a personalized feedback module. However, considerable uncertainty and limitations exist regarding its efficacy. Given that the personalized feedback module has a relatively high risk, further evaluation is required before adding a personalized feedback module to diabetes apps. Consistent with a previous review [
Our study also has some limitations. First, the exploratory and observational nature of our subgroup analyses and the possibility of misclassification prevented us from drawing a solid conclusion about the modular efficacies and risks. Second, we examined only 12 trials in our study, which may limit the strength of this systematic review. Third, we noted the asymmetry of the funnel plot, which indicated a potential risk of publication bias in our systematic review.
In our study, we developed a 3-axis taxonomy for diabetes self-management apps. Mobile app-based interventions improve glycemic control in adult outpatients with diabetes, especially in those with T2DM. Our analyses suggest that clinical decision making requires further improvement and evaluation before being added to apps. Safety issues such as hypoglycemia and other adverse events are being overlooked and need attention in future investigations.
Taxonomy development and validation.
MEDLINE search strategy.
List of excluded references with reasons for exclusions.
The detailed information, taxonomic classification, and risk of bias of included trials.
Funnel plot.
GRADE for primary and secondary outcomes.
Meta-analyses for adverse events.
Classifications of the patient-clinician communication function.
American Diabetes Association
Cochrane Central Register of Controlled Trials
Food and Drug Administration
Grading of Recommendations Assessment, Development and Evaluation
hemoglobin A1c
mean difference
randomized controlled trial
type 1 diabetes mellitus
type 2 diabetes mellitus
The authors thank Dr Xiaodan Li, Department of Gastroenterology, West China Hospital, Sichuan University, for her comprehensive revision and advice for this manuscript. The corresponding authors were supported by grants from the National Natural Science Foundation of China (grant nos. 81400811 and 21534008), National Basic Research Program of China (2015CB942800), and the Scientific Research Project of Health and Family Planning Commission of Sichuan Province (grant nos. 130029 and 150149).
YW, XY, HT, and SL conceived this study. YW and YD contributed to data extraction and quality assessment. YW and XY contributed to the Web-based searches of the literature and apps. YW and SL wrote the report. YW, JK, and LL conducted the statistical analysis. GV, AN, and XS guaranteed the study methodology. All authors discussed and interpreted the results and reviewed the manuscript before submission.
None declared.