Following analysis of the literature on the available apps for diabetes management [18], main design functions for diabetes apps [19,20], app evaluations [21], and user reviews of the top diabetes apps available in the Google Play and Apple Store markets, we defined the most important points to consider within our project. For a mobile monitoring system, we derived the following core features: (1) systematic collection and persistence of patient data locally on the user device and centrally on a server, (2) data export into commonly used formats for assessment by patient and doctor, and (3) personalized recommendations based on evaluation of PPG response. Further, for the system to be sufficient for long-term monitoring, the following requirements must be satisfied: the system must be based on a commonly used architecture (in terms of devices, software, etc), the system must be sufficiently simple to use by patients without prior computer literacy, and the process of data collection and exchange should be as simple as possible.

The conceptual scheme of the GDM recommender system is shown in Figure 1. The core of the recommender system comprises dynamically changing BG predicting models, which are used for personalized recommendations. The architecture utilizes the data collected from different sources for all patients to constantly improve its BG predictions. The patient state is a vector in multidimensional space and contains data on preceding events (eg, meals and insulin injections) and information on BG levels collected within the continuous glucose monitoring (CGM) system signal. The patient state is combined with data prototyping patients (with the data relevant to BG regulation, such as glycated hemoglobin A_1c [HbA_1c] and oral glucose tolerance test [OGTT] parameters) and the data for all patients in a centralized database that is used to train the BG prediction models. The models are dynamically retrained as new data are uploaded making their predictions more sustainable when they are used by new patients for whom only a limited amount of data has so far been recorded.

In the first step of the system design, we formulated a list of parameters for the BG predicting algorithm (Textbox 1). Because it was not feasible to track these records in a traditional paper diary and there were no solutions matching our requirements, such as apps for recording meals in a simple and hassle-free (mobile) manner and corresponding software to acquire the records of these electronic diaries alongside CGM system signals, we developed our own app to obtain the necessary records for the BG prediction algorithm.

From a technical viewpoint, the developed GDM advisory system contained the following elements: (1) a mobile app for data collection and presentation on the side of the patient, (2) a CGM system for continuous BG monitoring, and (3) a centralized server with appropriate software for data aggregation, processing, and training of BG prediction models.

The mobile app for data collection and exchange was developed for the Android OS. A desktop app with the same functionality was also developed for users not in possession of an Android device (both are referred to simply as “app”).

The app contains 18 different screens, including the main menu, user input forms, record management and information, user settings, data export form, and help. Some of these screens are presented in the animation included in Multimedia Appendix 2.

The SQLite database in the app consisted of 13 tables (tables for records on BG, insulin, physical activity, sleep, ketones, meals, and meal items; tables with built-in and user food databases and user data) containing the data for different types of user records as well as the built-in food database. The built-in food database, collected from open sources (including the Scientific Research Institute of Nutrition of the Russian Academy of Medical Sciences and the USDA food databases), made it possible to track 27 food parameters (macronutrients and micronutrients) without patient input, because involving the patients could lead to mistakes and additional burden from the patients’ perspective.

The app allowed the users to export data in Excel spreadsheets and store them on their devices as well as to send them remotely to physicians (Figure 3).

Using the preceding methods, an algorithm transformed the data for the amount and kind of consumed food, the start time of food intake, physical activity, duration of sleep, and current BG level (received from the CGM system) into a BG prediction parameter used to establish a recommender system. An example of the results of data matching between the CGM system signal and the electronic diary is shown in Figure 4 and more examples are shown in Multimedia Appendix 3.

The main characteristics of the participants are presented in Table 1. The women with GDM had higher BMI and higher levels of HbA_1c and plasma glucose (PG) during OGTT than the controls.

The average characteristics of meals and glycemic responses are presented in Table 2. Patients with GDM consumed significantly lower amounts of carbohydrates and fats in their meals; therefore, the energy content of their meals was considerably lower than that of the control group. Fasting and postprandial BG levels were significantly higher in patients with GDM than in those in the control group, whereas the actual rise in BG level after meals did not vary significantly in these groups owing to lower average carbohydrate consumption in the GDM group. The area under the curve (AUC) for BG level 1 and 2 hours after the beginning of the meal was also larger for patients in the GDM group, even considering the lower average carbohydrate consumption in this group.

The significant model coefficients are presented in Table 3. The BG level 2 hours after OGTT was a heightening factor for all the variables describing glycemic response, corresponding to worse BG regulation in participants with GDM. Some of the parameters describing lifestyle during pregnancy were significant in predicting PPG response: reported physical activity was a lowering factor for all the variables describing glycemic response, where they were found to be significant, as well as reported high consumption of legumes, although high consumption of coffee appeared to be a heightening factor.

Despite their simplicity, the developed linear regression models proved to be highly efficient in their prediction of the PPG response feature. Model performance was estimated using standardized metrics. The correlation between real and predicted values (R), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) were estimated for each of the proposed models (Table 4). The resulting table shows only marginally worse results for the test set, which may be due to overfitting. When comparing the MAPE of the presented models, it might be considered that the models predicting AUC on a postprandial curve perform better.

Our infrastructure, including the mobile app for patients’ electronic record management and data processing algorithms for matching of the CGM system signal and electronic diary, enabled the collection and analysis of data on 909 food intakes and corresponding postprandial BG curves from 62 pregnant women. Combining these data with patients’ characteristics (eg, HbA_1c, BMI, age, and lifestyle parameters) facilitated the development of models that accurately predict PPG response for real-life meals and can be implemented in our app for personalized prediction of PPG response and subsequent decision-making support.

To the best of our knowledge, no previous study has developed a tool for personalized prediction of PPG response in pregnant women. As the physiology of pregnancy differs significantly from that of the nonpregnant state, models created on nonpregnant participants have no potential to be applied to the management of pregnant women.

A unique feature of our app is the ability to integrate food choices based on our PPG response models in a decision-making algorithm. This will enable the app to give personalized advice concerning each upcoming meal in order to achieve desired BG levels. To adapt the personalized model to her specific requirements, a woman simply needs to fill in a short questionnaire when the app is first started (the questionnaire covers parameters 9-16 listed in Table 3). In order to receive on-site personalized nutrition advice, participants have to enter the desired food components with exact weights before meals. The app will calculate the predicted PPG response and, in case the recommended levels are exceeded, will suggest how to reduce the carbohydrate content of the desired meal by reducing the amount of carbohydrate-rich products or by suggesting variants of products for replacement.

The significant potential of our app lies in the ability to capture data, provide decision support, and share data with health care providers, thus promoting communication. Although the efficacy in terms of use of our app is yet to be tested in an RCT, we believe that it already has potential for supplementing traditional care, especially between visits to the clinic, when patients can be provided with on-site personalized recommendations and education. Our app may also be helpful for the collection of large datasets for future statistical analyses.

For short-term use (eg, pregnancy), mHealth solutions might be more widely accepted and there will potentially be less app attrition compared to patients with chronic diseases.

To test and train PPG response prediction models, we used data of women with normal glucose tolerance in line with the data from GDM patients. We consider this approach appropriate because the postprandial BG levels of women with normal glucose tolerance and GDM diagnosed under the new IADPSG criteria often overlap, as shown in Table 2. Because women with normal glucose tolerance are not required to adhere to a diet, they consume much more carbohydrates. Consequently, they often have BG levels similar to GDM patients, thus enabling us to assess the impact on PPG response of products with a more variable carbohydrate content.

Our approach of making patients select foodstuff from a prebuilt database is a good way to standardize and ease the procedure of food tracking as well as to increase data reliability. On the other hand, this created an additional problem for patients who had to search for their desired food items.

We provided patients with the possibility to create their own food items by combining existing items, but only a few people found this feature useful. There are two main reasons why we refrained from allowing users to add food items to the app database on their own. Firstly, there is virtually no way to find a comprehensive list of all the nutritional characteristics we would like to track (previous 25 features) for potential food items. Secondly, there is a concern with the quality of data available on publicly available Internet sources. Allowing users to add new items into the database might lead to mistakes and incomplete data, which might in turn cause bias in the statistical analyses.

The most important predictors of PPG response that remained significant for all models were BG level before meal, quantity of carbohydrates in the meal, type of food (breakfast was a heightening factor), 2-hour BG level at OGTT, frequency of coffee consumption as a heightening factor, and the level of physical activity expressed as climbing stairs during pregnancy, which lowered PPG response. All these parameters are expected with regard to physiology and are in line with the results of previous studies addressing factors determining glucose metabolism [33-35].

Table 5 compares the prediction quality of the developed predicting models to that of BG predicting models of different types presented in the literature. All models exhibit adequate accuracy that allows them to be used for patient assistance. The predictive power of the developed models and those presented in recent scientific papers is of equal quality in terms of R, RMSE, MAE, and MAPE. Considering that the developed model is simple, interpretable, and requires little time to tune its parameters, in which it fulfills the requirement for dynamical coefficient updates, the model must be stated as adequate for the pursued goal.

Our study has several limitations. Because the data on food intake were self-reported by participants through the app, there is the potential for biases due to inaccurate reports. For example, women who are overweight or gain excessive weight during pregnancy often underestimate their actual consumption of foods that are considered harmful. This is a typical drawback of any epidemiological study assessing nutrition. However, we used hints found in the data to reduce this kind of bias, as described previously (see Control of the Accuracy of Self-Reports section).

In addition, the PPG response prediction models are based on data from a relatively small sample size. However, the resulting PPG response prediction models have predictive power of equal quality to those presented in recent scientific papers [10,36-39]. Overall, we believe that the developed PPG response prediction models are sufficiently accurate to form the basis for a subsequent self-management intervention.

There are several concerns limiting the current use of the developed app with the built-in decision-making algorithm. Our app is so far only available for Android devices and PCs; unfortunately, this prevents iPhone users from accessing the app using their mobile phone (they can access it only using their desktop computer). We plan to increase the number of potential users by implementing an app for iPhone users as well. We are also considering implementing a responsive Web app to allow usage of a wide range of platforms and form factors.

This study did not include insulin treatment. Thus, the current models are not designed to predict PPG response after insulin injections. This will need to be evaluated in the future.

Our app currently lacks the ability to wirelessly upload BG readings. We also plan to overcome this inconvenience in the future.

The algorithm used for data processing is currently only semiautomatic and requires the intervention of the researcher in the final data processing steps. In the future, all the system software should conceivably be rewritten in a high-level programming language.

The developed PPG response prediction models are based and validated on the data of women in the third trimester of pregnancy. Therefore, it should not be used in the first and second trimesters. However, GDM is usually diagnosed in the third trimester and, according to the recommendations of some diabetes associations, can be diagnosed only in the third trimester [8]. Because Russian guidelines suggest diagnosing GDM at any time during pregnancy [24], we are planning an additional study to validate the PPG response prediction models for the first and second trimesters.

To improve BG control, as well as maternal and fetal outcomes, the efficacy of this app for the management of women with GDM needs confirmation in an RCT and we plan to perform such an RCT.

This infrastructure comprising the data processing algorithms, the BG prediction models, and the mobile app for patients’ electronic record management can be useful for guiding BG prediction-based personalized recommendations for GDM patients. The accuracy of the prediction models was validated on training sets of patients with 20-fold cross-validation for parameters tuning. The efficacy of the implementation in terms of providing health care to women with GDM to reduce BG levels and pregnancy complications will be evaluated in a future RCT.