The GlucoNote app was built with Apple’s ResearchKit [7]. Users enter body weight, blood sugar levels (fasting blood sugar or postprandial plasma glucose), and blood pressure levels. Changes in these parameters are displayed as graphs (Figure 1). Steps are counted by each iPhone’s built-in pedometer, and after the user sends the physical activity data to the server, step counts are displayed as a graph (Figure 1). Each time the user sends the physical activity data to the server, the step counts recorded in the past 30 days are also sent to the server, including the step counts of the 30 days recorded before study enrollment. If the user does not send the physical activity data, the step counts will not be sent to the server. GlucoNote facilitates easy input of dietary information by meal photos; this function was developed for DialBetics and is described in detail elsewhere [13]. To record dietary information, users can enter photos of a meal or choose from a menu list of 2913 items based on Eat Smart, a database provided by Eat Smart, Inc. When dietary information is entered from the menu list, the app automatically calculates each meal’s intake of calories, protein, fat, carbohydrate, dietary fiber, cholesterol, and salt, which are displayed as a graph. The recommended intake for 1 meal—calculated as one-third of the recommended daily intake—is also displayed for comparison (Figure 1). As it has been reported that self-monitoring is crucial in weight loss [14] and improved glycemic control [15], the GlucoNote app was designed to provide users with visual feedback of the parameters, which helps in self-monitoring of body weight, blood sugar level, physical activity represented by step counts, and diet. Physicians, including a diabetologist, nurses, and dietitians were involved in the development of GlucoNote.

Before release, the study was approved by the Institutional Review Board of the University of Tokyo. All iPhone users aged ≥20 years in Japan with type 2 diabetes or prediabetes were eligible for the study after they provided electronic consent. They could download GlucoNote for free.

The GlucoNote app was made available on March 14, 2016, (but only in Japan) through the Apple App Store [16]. The GlucoNote release was announced in the homepage of the department and a press release was issued [17] (Multimedia Appendix 1). After downloading and opening GlucoNote, participants were questioned for eligibility, that is, “Are you 20 years old or older?”; “Have you been diagnosed as type 2 diabetes or prediabetes?”; “Are you able to understand and follow the consent forms?”; and “Are you living in Japan?”. Those who met the eligibility criteria proceeded to the informed consent screens. Before providing consent, participants had to read about the risks and benefits of participating in the study and their right to withdraw from it; they could withdraw at any time without giving reasons. Informed consent was by digital signature. Participants who enrolled between March 14, 2016, and March 13, 2017, were analyzed (Figure 2).

Participants were encouraged to fill in the following profile information: sex, body height, body weight, wake-up time, bedtime, smoking habits, age at diagnosis, presence of retinopathy, presence of neuropathy, and regular visit to a dentist. Target daily calorie intake could be calculated based on height and activity level.

Participants also filled in the results of medical examinations: date of examination, height, weight, waist circumference, blood pressure, blood sugar, hemoglobin A_1c (HbA_1c), total cholesterol, high-density lipoprotein cholesterol, triglycerides, serum creatinine, aspartate transaminase, alanine transaminase, gamma-glutamyl transferase, urine protein, urine sugar, and urine albumin-creatinine ratio.

As shown in Figure 2, the observation duration was defined as the time from the day of study enrollment to May 13, 2017, 2 months after the end of enrollment. Retention duration was defined as the time from the day of study enrollment to the last day during the study period on which the user recorded at least 1 of 4 items: weight, blood sugar, blood pressure, or diet. Steps are different from other data—body weight, blood sugar, blood pressure, or dietary information—that require user input each time; they were automatically counted by the iPhones, and these data of steps recorded in the past 30 days were sent to the server at one time. Therefore, steps were discounted from analyses of retention rate.

The number of first-time app downloads was obtained from App Store Connect (Apple Inc), excluding app updates, downloads from the same Apple identification onto other devices, and redownloads [18].

For the participants’ demographics in Table 1, body mass index (BMI) was calculated from their first recorded body weight.

To compare the data between weeks, each participant’s mean values per week were calculated for weight, blood pressure, blood sugar, energy intake, or number of steps when the participant recorded those at least once during that week. When the user did not record that parameter for the entire week, the mean value of the corresponding week was treated as a missing value. When the user recorded the same parameter twice or more times on the same day, the data recorded first that day was used to calculate the mean value per week. The mean value recorded in the first 7 days starting from the day of study enrollment was defined as the value at week 0 (Figure 2). For the number of steps, the mean value recorded in the 7 days starting from 4 weeks before the day of enrollment was defined as the value at week −4 (Figure 2).

Daily energy intake was calculated by summing the calories of each meal when at least 3 meals were recorded per day. When only 1 or 2 meals were recorded, the daily energy intake for that day was treated as a missing value because one cannot tell whether participants did not record the meals or did not have the meals.

Retention rates were calculated by the Kaplan-Meier method [19], and the difference in retention rates between the 2 groups was evaluated by the log-rank test. The participants who continued to use the app to the end of the observation period were treated as censored cases and shown as vertical tick marks on the Kaplan-Meier curve. To compare characteristics between robust users and nonrobust users, the Fisher exact test was used for categorical variables and the Mann-Whitney U test was performed for continuous variables. A paired t test compared the data between week 0 and week 4 for parameters with a normal distribution, and a Wilcoxon signed-rank test was used for those with a non-normal distribution. Statistical analyses used R and Easy R [20].

After GlucoNote was released on March 14, 2016, it was downloaded 1703 times during the first year, with 581 users consenting to take part in the study. The daily number of downloads during the year is shown in Multimedia Appendix 1. Of these, 59 users withdrew and the remaining 522 were analyzed (Figure 3). Their demographics are shown in Table 1. Males accounted for 79.9% (417/522) of the users, their median body weight was 70.6 (interquartile range [IQR] 62.0-81.5) kg (n=274), and BMI was 24.53 (IQR 21.99-27.82) kg/m² (n=270); 200 users answered the question about their smoking habit, 32 (16.0%; 32/200) of them being current smokers; 178 users answered the question about age at diagnosis, 111 (62.4%; 111/178) were diagnosed between the ages of 40 and 59 years; and 41 users recorded their HbA_1c level, the median value was 6.3% (IQR 5.9-7.1; Table 1).

A total of 467 participants recorded at least 1 of 5 items: body weight, blood sugar, blood pressure, diet, or steps (Figure 3). The number of participants who recorded body weight, blood sugar, blood pressure, dietary information, or step counts was 274 (58.7%; 274/467), 172 (36.8%; 172/467), 169 (36.2%; 169/467), 275 (58.9%; 275/467), or 428 (91.6%; 428/467), respectively (Figure 4). The measured data of participants at week 0 are shown in Table 2.

Retention rates were analyzed for 357 participants who recorded at least 1 of 4 items: weight, blood sugar, blood pressure, or diet (baseline users; Figure 3). With the median observation duration of 382 days (range 63-426 days, IQR 275-423 days), the median retention duration in baseline users was 8 days (IQR 1-63 days; Table 3).

Retention rates for 2 days and 4, 8, and 12 weeks were 0.627 (95% CI 0.575-0.675), 0.353 (0.304-0.403), 0.272 (0.227-0.319), and 0.220 (0.179-0.265), respectively (Figure 5). Male participants had a higher retention rate than female participants (P=.01; Figure 5). No significant differences were found between the groups with age at diagnosis <40 years and ≥40 years (P=.13) or the groups with BMI <25 kg/m² and BMI ≥25 kg/m² (P=.62; Figure 5).

Of the 357 participants, 126 used GlucoNote longer than 4 weeks (29 days or longer) and were analyzed as robust users. The median observation duration and retention duration for robust users were 375 (IQR 256-422) and 124 (IQR 60-275) days, respectively (Table 3).

Characteristics of robust users (n=126) and nonrobust users (n=231) were compared. Consistent with the result above, men were more likely to be robust users than women (P=.02); no difference was observed between the groups when age at diagnosis was <40 years and ≥40 years (P=.11; Table 3). Moreover, no differences in smoking status (P=.45), HbA_1c (P=.34), or body weight (P=.59) were observed. Those who answered the questions about smoking status and age at diagnosis were more likely to be robust users than were those who did not answer those questions (P<.001).

Data at week 0 were compared between robust users and nonrobust users (Table 4). The daily number of steps at week 0 was higher among robust users than nonrobust users (median 6108 vs 5171; P=.001), whereas step counts at week −4 were comparable between the 2 groups (median 5876 vs 5594; P=.37). In addition, daily energy intake at week 0 was higher among robust users than nonrobust users (median 1595 kcal vs 1451 kcal; P=.04; Table 4). No significant differences in body weight, blood sugar, or blood pressure were observed.

The data for robust users were compared between weeks 0 and 4 (Figure 2). Strikingly, body weight significantly decreased from week 0 to 4 (mean 71.3 [SD 14.1] kg to mean 70.8 [SD 13.9] kg; P=.002; Table 5). The mean difference in body weight between week 0 and week 4 was −0.5 kg (SD 1.2), corresponding to a mean decrease rate of 0.6% (SD 1.6). No changes in blood sugar level, blood pressure level, daily energy intake, or daily number of steps were observed between weeks 0 and 4 (Table 5).

After the observation period, a questionnaire was sent to the participants asking whether they found GlucoNote useful for health management. Only 22 out of 522 participants (4.2%; 22/522) replied: 4 (18%; 4/22) people found the app very useful, 6 (27%; 6/22) useful, and 12 (55%; 12/22) not very useful; none of the participants pronounced it not at all useful.

In our study, nearly 80% (79.9%; 417/522) of the participants were men (Table 1). This is much more biased than expected, even taking into consideration that men, in Japan, are more likely to have diabetes or prediabetes than women (28.5% vs 21.4%, respectively) [21]. There have also been reports that men use information and communication technology (ICT) more frequently, have less ICT anxiety, and have a more positive attitude toward ICT self-efficacy than women [22-25]. Consistent with those reports, our recent survey of patients with chronic conditions suggested that male patients were more likely to express willingness to use a personal health record than female patients (paper in preparation). It has further been reported that women are more likely to perceive diabetes as a stigma than men [26], and it may be that women were reluctant to use an app that was plainly meant for diabetes patients. Efforts to remove such barriers for women may contribute to broadening usage. The bias was not because of the difference in ownership of smartphones: a 2018 survey in the Tokyo area reported that 77.9% of men and 80.9% of women aged between 15 and 69 years owned a smartphone [27].

To get enough participants, we had to forego gathering information such as age because the more the input required before setting up GlucoNote, the fewer comply—meaning fewer study participants. That is why complete user demographics are unavailable. Moreover, many values are missing from the users’ profiles, a problem shared by similar studies [10]. For example, only 38.3% (200/522) of the participants answered the question about smoking (Table 1). Not surprisingly, those who answered questions such as smoking status or age at diagnosis were likely to use GlucoNote longer (Table 3). Improvement in usability and security may be helpful to entice participants to enter more information. In our recent survey of patients with lifestyle-related diseases, time and effort needed and concern over security were identified as the main barriers to using such personal health records as health care apps (paper in preparation). For example, instead of relying on users to manually input the data, developing functions such as automatic recognition from a photo image of medical examination results may be useful—as may be reassuring users by improved security.

We do not have information about the use of other apps by GlucoNote users or people who visited the product page of GlucoNote, although that would have been helpful in identifying potential users.

A rapid decline in retention rates has been reported in other studies using ResearchKit [12,28]. Although the ResearchKit platform offers the advantage of completely remote recruitment and enrollment, lack of human communication may mean less motivation for participants to continue compared with studies conducted face to face. In this study, the 2-day retention rate was only 0.627 (95% CI 0.575-0.675; Figure 5), meaning that more than 1 in 3 participants used the app only for 1 or 2 days. The median retention duration was as short as 8 days (Table 3). However, long-term use is essential for the app to affect users. Clearly, additional efforts to improve retention rates are necessary. Our previous study using DialBetics—conducted at a university hospital—showed a much higher retention rate of over 70% at 3 months, partly because the research team nurse contacted the participants and encouraged them when they missed measurements for 2 weeks [5]. Such intervention by medical staff certainly helps improve retention, thought it obviously means a higher cost. Alternatively, some incentives built into the app—such as reward points for long-term users—may be helpful.

The characteristics associated with the retention time of mHealth have not been well studied. In a study of asthma patients who used ResearchKit, being female and older correlated with longer retention [10]. By contrast, in our study, men had a longer retention rate than women (Figure 5 and Table 3). Age at study enrollment was not collected, but age at diagnosis did not correlate with retention rate (Figure 5 and Table 3). As mHealth is relatively new, factors predicting willingness to use it—or good adherence to mHealth—have not been well understood. To deliver mHealth effectively, it is important to identify the characteristics that mark suitable candidates for the app.

Interestingly, robust users had significantly higher step counts at week 0 than nonrobust users (median 6108 vs 5171 steps; P=.001), whereas step counts were comparable before using the app (median 5876 vs 5594 steps; P=.37; Table 4), suggesting that using the app may have prompted an increase in step count by robust users but not by nonrobust users. In contrast to steps that were counted automatically, daily energy intake data may be incomplete because users may not have recorded all their food and drinks. The higher daily energy intake among robust users at week 0 (Table 4) might reflect a likelihood that these users input meal information more completely than nonrobust users, in turn, a reflection of robust user attitudes toward using the app in contrast with those of nonrobust users.

We observed a significant decrease in body weight (P=.002) after 4 weeks, with a mean decrease rate of 0.6% (SD 1.6; Table 5). Of 67 users with weight data for weeks 0 and 4, 47 (70%; 47/67) showed a decrease at week 4, suggesting that using the app might have prompted behavior modification leading to weight decrease.

It was previously reported that weight loss was the dominant predictor of reduced diabetes risk in patients with impaired glucose tolerance, with every kilogram of weight loss resulting in 16% reduction in risk [29]. A nationwide Japanese intervention program showed that for people with metabolic syndrome, a 1% to 3% weight loss after a 6-month lifestyle modification program resulted in a significant decrease in HbA_1c [30]. Although body weight loss in this study (mean −0.6% in 4 weeks) is only marginal and may be of little clinical significance, the long-term use of GlucoNote by improving the retention rate may lead to a reduced diabetes risk in patients with prediabetes. More importantly, it would be of interest to follow up to determine if the weight loss is maintained over time.

It is difficult to explain the exact reasons for weight loss only from the data recorded; no change in steps per day or energy intake was observed between weeks 0 and 4. However, one can speculate that some behavior changes undetectable by an app were prompted—such as increased physical activity other than walking. In addition, it is possible that the app failed to detect changes in diet because users may not have recorded all their food and drinks.

There are several limitations of this study. First, as discussed above, its major limitation is the low retention rate. Second, as noted above, the participants’ complete demographics were unavailable, and data collection was incomplete. Consequently, many analyses were based on different sample sizes, a common defect shared with similar studies [10-12]. Correcting this will require effort to improve response rates and collect more complete data. Moreover, only robust users were available for the analysis, increasing the risk of selecting bias. Third, the study design was free of controls and eligibility, including diagnosis of diabetes and prediabetes, and the measured data were entirely based on self-reports and had not been validated by medical professionals—a common problem shared by similar studies. Fourth, the overall number of downloads was lower than expected, with only 1703 in the first year after the release. A more effective way to inform people of the app is desirable. In addition, to widen its targets to include people with a metabolic syndrome, including overweight or even healthy people may be considered. Finally, to evaluate the effects of the app, a randomized controlled trial must be performed.

We developed and released GlucoNote, a novel app that uses ResearchKit to support self-management in patients with type 2 diabetes and prediabetes. This afforded a valuable opportunity to evaluate usage patterns. Analyses of the participants who enrolled in the study within 1 year of the release revealed the potential advantages and challenges of GlucoNote. Future tasks include improving retention rates and evaluating its effects.