The eligibility requirements for participating in the FGIs included having diabetes, being an adult, and having a smartphone capable of running apps. Participants were recruited through an advertisement posted on the website of a self-help online cafe for diabetes patients from May 11 to 17, 2016. The 19 recruited diabetes patients were divided into two groups. Two nurses interviewed each group for about 1 hour with the aid of the semi-structured questions used in the study of Choi et al. [15].

All of the interviews were recorded and then analyzed using the following qualitative content analysis procedure [20]. First, the audio-recorded FGIs were transcribed verbatim by the first author. A PhD student in nursing who is an expert in qualitative research reviewed the transcripts to check for any missing information. Second, three members of the research team independently developed an initial codebook of themes that arose after an FGI had been conducted and all of the FGI transcripts had been read. The rule applied for the extraction of themes was that the narratives were expressed by at least one FGI participant and could be implemented in the app. Third, the same members reviewed and had periodic discussions to achieve a consensus about the codes to be used and to identify convergent, salient, and/or unique themes. Fourth, the same members then classified all the themes according to the four-category schema (information, personal motivation, social motivation, and behavioral skills) developed by Choi et al. [15]. The final codebook of themes and categories was then reviewed by the PhD student.

Functions were extracted based on the functional requirements of users and literature about the implementation of DSM apps. The functional requirements of users were determined by reviewing previous DSM intervention studies, and additional functions were extracted by three nursing informatics experts based on a healthcare report from IMS Health and the systematic review of Chomutare et al. [17].

The knowledge provided in the DSM app was extracted from clinical practice guidelines. Two diabetes nurse educators (with 2 years and >10 years of experience) advised the first author about selecting appropriate guidelines and extracting DSM knowledge.

The DSM knowledge was extracted based on the DSM key areas of the AADE excluding psychosocial areas: healthy eating, being active, self-monitoring of blood glucose (SMBG), taking medication, problem-solving, and reducing risks. The extracted knowledge was organized as an assessment of the DSM knowledge and previous behaviors of a patient based on goal setting (setting a realistic goal), feedback (commenting on goal achievements), and evaluation (evaluating the DSM behaviors of the patient).

To implement the knowledge and functional requirements extracted from the analysis stage, we designed the overall system diagram. The DSM app consists of a client part running on the user's smartphone and a server part running on the system. We depicted the flow of data between the client and the server. We have drawn the system diagram using Microsoft Office Power Point program.

A data model is an abstract model that organizes elements of data and defines the relationships between entities and the relationships to properties [21]. In this study, we defined all the data entities necessary for the DSM app. These data entities each had a data type (such as Boolean, Integer, Date time, Coded text), possible values, and a unit. For example, for a body mass index (BMI) data entity, the data type is float, unit kg/m², and possible values from 10 to 100. A database was designed to collect data entered by the user in the DSM app, such as blood glucose, diet, exercise, and diary. We used the Erwin software to draw an entity-relationship diagram.

An algorithm is an ordered sequence of steps that describe solutions of problems. In this study, algorithms for collecting patient information at the first visit, setting goals, providing DSM education, and evaluating DSM behavior on return visits were developed. We used the free online diagram software (draw.io) to design a flow chart of the algorithms.

The display screens were designed to maximize the convenience and concentration of users during their interactions with the DSM app. The menu was designed to minimize the need to move between screens of the DSM app. The display screens and menus were drafted by the authors and finalized after several rounds of meetings with the designer.

The recommendations selected by two diabetes nurse educators and provided by the DSM app were compared using 13 scenarios that were developed by the first author to evaluate the algorithms. The scenarios used in the evaluation included at least one decision-making node from the decisionmaking process for the DSM app. The professional diabetes nurses read each scenario and selected appropriate recommendations based on clinical practice guidelines. The data included the same scenarios that the nurses used and were entered into the DSM app, and then the recommendations from the app were recorded.

A score of +1 was assigned when both the category and details of the recommendations from the DSM app and the diabetes nurses matched. A score of 0 was assigned when the two recommendations were in the same category but with different details, whereas a score of −1 was assigned when the recommendation categories differed. The proficiency and efficiency scores were calculated as follows [22]:

The above equations indicate that proficiency was the ratio of the sum of consistent, partially consistent, and inconsistent recommendations from experts and the app to the total number of recommendations, while the efficiency was the ratio of consistent recommendations from experts and the app to the total number of recommendations.

The heuristics of the DSM app was evaluated using a mobile heuristics tool by two experts with majors in nursing informatics and one expert with a major in medical informatics, each of whom had more than 5 years of experience in his or her field. The following eight principles of Bertini's mobile heuristics tool was measured on a 5-point Likert scale (0 ‘no heuristic problem’ to 4 ‘heuristic problem exists and must be fixed’): visibility of app status, match between the app and the real world, consistency and mapping of the app, ergonomic and minimalist design, ease of input and readability, flexibility and efficiency of personalization, aesthetics and privacy, and realistic error management [23].

Three experts were provided with Android smartphones running the DSM app for 1 day to use and evaluate the heuristics problem and severity scores. When at least two experts raised the same heuristics problem or only one expert did this but the severity score was greater than 4, the app was modified appropriately [23].

The usability of the DSM app was evaluated by diabetes patients who agreed to participate. We recruited the participants from an online diabetes self-help group. These patients used the diverse functions of the DSM app, such as blood glucose input, diet and exercise input, bulletin board, and interfacing glucometer, for 2 weeks. They received a text message once a week to encourage use of the DSM app. After 2 weeks, we surveyed the users about their experience and satisfaction through an open questionnaire about the blood glucose input function, diet and exercise input function, and feedback function as well as their overall satisfaction with the DSM app.

The comments from the heuristics evaluation and usability evaluation were used to modify the app after it had been reviewed by the three nursing informatics experts.

Table 1 presents the demographic characteristics of the FGI participants. These participants were aged between 33 and 55 years (mean, 41.61 years), and more than half of them (68.4%) were male. Seven participants had type 1 diabetes (36.8%), 11 had type 2 diabetes (57.9%), and 1 had impaired glucose tolerance (5.3%). The mean duration of diabetes was 22.8 years in those with type 1 diabetes and 5.1 years in those with type 2 diabetes.

Table 2 presents the categories, themes, and some representative quotes from the focus groups. The 11 themes extracted from the FGIs were organized into the following 4 categories developed from the IMB model:

Sixteen DSM app functions were extracted (Figure 2).

Ten functions (education, recommendations, feedback, setting a goal, recording, checking, diary, sharing, communication, and viewing) were extracted based on the requirements of users from the FGIs. The gamification requirement was not translated into a function because its implementation would require excessive economic and human resources. Two functions for reminding and alerting were extracted from the report of IMS Health [5], one function for the glucometer interface was extracted from Chomutare et al. [17], and three functions for managing the app (registering, login, and notices) were added.

The knowledge sources selected included four clinical practice guidelines published by the AADE [4], the American Diabetes Association (ADA) [24], the Korean Diabetes Association (KDA) [25], and Korean Association Diabetes Nurse Education (KADNE) [26]; a relevant paper on DSM published by the Diabetes Education Center of Seoul National University Hospital [27]; the Diabetes Knowledge Test of the Michigan Diabetes Research Center [28]; the DSM assessment report tool (D-SMART) [29], and the food and exercise calorie database of the Korea Food and Drug Administration (KFDA) [30].

The following 49 knowledge rules for DSM were extracted based on the DSM key areas of the AADE: 10 rules for diet, comprising 6 on assessment, 1 on goal setting, 2 on feedback, and 1 on evaluation; 7 rules for exercise, comprising 2 on assessment, 1 on goal setting, 2 on feedback, and 2 on evaluation; 9 rules for SMBG, comprising 4 on assessment, 2 on goal setting, 2 on feedback, and 1 on evaluation; 3 rules for medication adherence, comprising 2 on assessment and 1 on evaluation; 9 rules for reducing risks, comprising 2 on assessment, 3 on goal setting, and 4 on feedback; and 11 rules for problem-solving, comprising 8 on assessment and 3 on evaluation.

Knowledge in the SMBG domain may be illustrated by the following examples. If a patient measures his or her blood glucose less than once a week, the DSM app provides a recommendation about the importance of measuring blood glucose (assessment). If a patient sets a pre-meal blood glucose target below 70 mg/dL, the DSM app provides a recommendation about the appropriate blood glucose target range (goal setting). If a patient achieves the goal, the DSM app provides positive feedback, such as “You have a good blood sugar. Keep your status steady” (feedback). If a patient measures his or her blood glucose less than four times a week or if fewer than half of the recorded blood glucose values are within the normal range, the DSM app provides recommendations about the importance of measuring blood glucose, its normal range, and how to measure it.

The DSM app consists of two parts (Figure 3), a client part that runs on the user's smartphone and a server part for sharing data with other users. The client part provides personalized education and recommendations based on the user's personal information, health-related information, medication history, and DSM knowledge and performance level. It receives blood glucose levels wirelessly from a glucometer and provides feedback based on individual goals, records of DSM behaviors, and blood glucose levels.

The server part has a bulletin board for sharing records with other users and expressing empathy using a ‘like’ button. The record of a user's DSM behaviors, alarm setting, and app access history are collected to the server from the client part after 4 weeks of app usage.

The data used in the DSM app were 78 grouped into the following 15 categories: personal information, health-related information, medication history, goal setting, dietary record, food list, exercise record, exercise list, diary, blood glucose, medication compliance, blood pressure and weight, foot care, evaluation rules, and overall. Data models were created for the 78 data entities with possible values, data types, and units. One of the data models is presented in Table 3.

The database consists of 10 tables without any identifiable data elements in order to minimize security problems. A primary key was created by combining the registration time, email address, and a nickname, rather than using identifiable data elements. The 10 tables comprise personal information, dietary records, food list, exercise record, exercise list, blood glucose, medication adherence, blood pressure and weight, foot care, and an alarm.

The DSM app algorithm shown in Figure 4 was developed based on the functions and knowledge extracted in the analysis step.

The patients who used the DSM app for the first time were asked to complete consent forms. Their personal information, health-related information, medication history, and DSM performance level were assessed (box A in Figure 4). Next, the patients set their goal, that is, amount and periods of DSM behaviors. If a goal was assessed by the system as unrealistic, more-appropriate values for the amount and period of DSM behaviors based on clinical practice guidelines were recommended in a pop-up window (box B in Figure 4).

After setting a goal, the patient's knowledge level of DSM behaviors was assessed, and the app provided personalized education according to this level of knowledge (box C in Figure 4). For example, a patient who did not select the correct answer (“one glass of diet cola”) in response to the question “what is incorrect hypoglycemic therapy” was considered to be lacking knowledge about how to cope with problems. In such a case, the app would provide personalized education about a “hypoglycemia coping method.”

After providing education, the app assessed the adequacy of the DSM behaviors of the patient and provided appropriate individualized recommendations (box D in Figure 4). For example, patients who had never cared for their feet were considered to have inappropriate foot-care habits, so the app provided individualized information on the “significance of foot care” and “foot-care method.”

In the case of patients who reused the DSM app, the app identified their goals set during their first use and provided them with feedback according to goal achievement (box E in Figure 4). The app then provided individualized information according to the evaluation of record fidelity, exercise, diet, SMBG, percentage of normal blood glucose, and adherence with medication based on the record of DSM behaviors in the app, which was written by the user (box F in Figure 4). For example, if the recorded dietary calorie intake over 1 week was less than 60% of the target intake, the app determined that the dietary habits were incorrect and provided individualized information on “normal weight and daily required calories,” “three general principles of healthy eating,” and lists of replacement foods.

The menu structure of the DSM app comprises 6 main menus: initial input, evaluation, bulletin board, record, information, and option. The initial input menu appears as a pop-up screen when the user accesses the DSM app for the first time, and the evaluation menu appears when the user accesses the DSM app after the set target period has elapsed; thus, it was designed without a submenu. The bulletin board, record, information, and option main menus were designed as fixed menus of the DSM app, and the main menu of the bulletin board was designed as a shared diary submenu. The record main menu consists of profile, diary, blood glucose, diet, exercise, medication, blood pressure and weight, and foot-care submenus. The information main menu consists of SMBG, healthy diet, exercise, medication, problem-solving, daily record view, graph view, and statistical analysis submenus. The option main menu consists of goal setting, reminder, notices, communication, and glucometer-interface submenus.

Forty user screens were designed to implement 6 main menus and 22 submenus in the app. The screen for the bulletin board, SMBG information, medication adherence checking, and blood glucose level recording are shown in Figure 5.

The proficiency of the DSM algorithm was evaluated by two evaluators as 91.43% and 90.48% (mean, 90.96%). The efficiency of the DSM algorithm was evaluated as 94.29% and 90.48% (mean, 92.39%).

The recommendations of the experts and the app did not match for SMBG. The DSM app did not provide SMBG information to users who made measurements at least twice daily, in accordance with the AADE evaluation criterion of at least once daily. However, evaluator A suggested that people who faithfully self-monitor their blood glucose should be consistently provided with SMBG information. SMBG requires more-specialized knowledge than diet and exercise management, so the algorithm was modified to provide SMBG information to users who exhibited good behavior in checking blood glucose.

The heuristics evaluation of the DSM app revealed 15 issues (Table 4). The following three issues were considered by at least two evaluators to be serious, so the app was modified accordingly (italicized in Table 4): “since every record is shared, there may be privacy concerns,” “it is difficult to grasp the DSM behaviors in the record view at a glance,” and “it is inconvenient to be required to input the blood glucose level when entering a diet and exercise record.”

Usability evaluation was performed with the participation of 14 diabetes patients (8 men, 6 women) with a mean age of 43 years. Five major modifications were proposed (Table 5). They are blood glucose input screen; number of blood glucose inputs; diet, exercise input screen; blood glucose trend graph; and change of food and exercise name.

We modified the DSM app based on the heuristics and usability evaluation. First, we reduced the number of questionnaire items from 100 to 20 and mandatory input items. Second, we removed the limit for the number of blood glucose inputs and modified the blood glucose test time to be more detailed (‘morning’ - ‘before and after’ - ‘10 minutes’ combination). Third, we divided the input window into two for recording and viewing. Fourth, we added a function for users to select the level of record disclosure as “only the author can see” or “share”. Fifth, we upgraded the data visualization of the DSM app, query period, and viewing of blood glucose values. Finally, we modified the food database to provide representative food names and calorie information, including recipes, ingredients, and local food names.