Development of a Smartphone Application to Capture Carbohydrate, Lipid, and Protein Contents of Daily Food

Abstract

Background:

Meal lipids (LIP) and proteins (PRO) may influence the effect of insulin doses based on carbohydrate (CHO) counting in patients with type 1 diabetes (T1D). We developed a smartphone application for CHO, LIP, and PRO counting in daily food and assessed its usability in real-life conditions and potential usefulness.

Methods:

Ten T1D patients used the android application for 1 week to collect their food intakes. Data included meal composition, premeal and 2-hour postmeal blood glucose, corrections for hypo- or hyperglycemia after meals, and time for entering meals in the application. Meal insulin doses were based on patients’ CHO counting (application in blinded mode). Linear mixed models were used to assess the statistical differences.

Results:

In all, 187 meals were analyzed. Average computed CHO amount was 74.37 ± 31.78 grams; LIP amount: 20.26 ± 14.28 grams and PRO amount: 25.68 ± 16.68 grams. Average CHO, LIP, and PRO contents were significantly different between breakfast and lunch/dinner. The average time for meal entry in the application moved from 3-4 minutes to 2.5 minutes during the week. No significant impact of LIP and PRO was found on available blood glucose values.

Conclusion:

Our study shows CHO, LIP, and PRO intakes can be easily captured by an application on smartphone for meal entry used by T1D patients. Although LIP and PRO meal contents did not influence glucose levels when insulin doses were based on CHO in this pilot study, this application could be used for further investigation of this topic, including in closed-loop conditions.

Glucose control following meals is an important component of the management of type 1 diabetes (T1D) to reach optimal diabetes control aiming at an effective prevention of long term diabetic complications. Carbohydrate (CHO) counting of meal intakes is recommended in the guidelines of T1D management in clinical practice.¹ The benefits of CHO counting have been well documented, including reduced postmeal glucose excursions, improved HbA1c levels, increased flexibility in the choice of meal content and better quality of life.² These positive outcomes are mediated by a more accurate computing of insulin doses to be delivered to reduce glucose excursions related to CHO intakes. The availability of bolus calculators in glucose meters or insulin pumps further help the appropriate implementation of insulin delivery at mealtimes according to meal CHO content by taking into account premeal glucose level in meter and pump calculators, and ‘active insulin on board’ in pump calculators.^3,4 An identified limit of the usefulness of CHO counting is the frequent lack of ability of reliable CHO counting by the patients.⁵

However, it has been reported that non-CHO components of meals can also contribute to postmeal blood glucose control. Lipids (LIP) and proteins (PRO) have been the most investigated components of mixed meals which can influence postmeal glucose levels.^6,7 A recent review underscored the variability of impact of LIP and PRO on postmeal glucose levels according to their quantities and combinations but also the considered individuals.⁸ The most common effect of high-fat or high-protein meals is to delay postmeal glucose increment later than 3 hours; hence higher meal insulin doses are needed but the timing of insulin action must be taken into account. Of note, the studies considered in this review were characterized by an investigational background where the amounts of LIP and PRO were predefined and precisely computed to assess their impact on postmeal blood glucose levels.

We developed an Android application following endocrinologist and dietician guidelines and patient advices to have it easy to use in daily life. The main functionalities of this application include CHO, LIP, and PRO counting, based on the CIQUAL French food database,⁹ insulin dose computing according to counted CHO and annex functionalities to personalize the application with favorite meals, new foods, meal history, container pictures (eg, plates) to help in evaluating quantities. During our development, we aimed at reducing to the minimum the time spent and the number of steps to enter a meal, knowing that it is one of the most important critiques from users about this type of applications.

A preliminary assessment of the application use compared to patients’ own CHO counting during 1 week in ambulatory conditions shows that patients mainly fail in evaluating the amount of CHO in food intakes at lunch and dinner. Hence, use of the application leads to a more accurate estimation of CHO amounts and could allow a reduction in the occurrence of postmeal hypoglycemia and hyperglycemia. In this pilot study, we assessed as primary objective whether the amounts of meal LIP and PRO could be captured by using the application. We also looked whether computed LIP and PRO contents of daily food impacted glucose levels. However, because it was a pilot feasibility study limited to a small number of volunteers, the study was not powered to assess the specific effects of LIP and PRO meal contents on blood glucose levels. Moreover, no attempt was made to promote high-fat and/or high-protein meals since we primarily aimed at assessing the application usability for daily food in real life with no medical recommendation.

Methods

Study Design

Eligible patients were T1D adults more than 18 years old, trained on CHO counting and considered as mastering it from specific tests at inclusion, with a good to fair diabetes control (HbA1c between 6.5% and 8.5%). Participants were volunteers among the patients usually followed by the Department of Endocrinology, Diabetes and Nutrition at Montpellier University Hospital and living in the Montpellier region. Each patient was included after a medical outpatient visit to validate his or her ability to follow the study procedures. Then, each patient was given an Android smartphone with the developed application, a user manual, and a logbook. The study was performed in agreement with our Institutional Ethical Board.

For 7 days, the patients had to enter each meal composition in the application (food name and quantity in grams) and the premeal blood glucose level, while they noted in the logbook their own estimation of the CHO content of each meal as they usually do. They were asked to refrain from taking snacks between the main meals except if needed. The application was used in blinded mode, which means that the application saved computed CHO, LIP, and PRO components of each meal without showing the result to the user. Hence, meal insulin doses were based on patient estimation of CHO content of each meal. The patients were also asked to note in the logbook complementary information, including time taken to enter each meal composition in terms of food category and amount, 2-hour postmeal glucose levels, and any events that occurred before or after the meal (eg, physical activity). Whether they had to treat them for hypoglycemia or hyperglycemia within the 3 hours following meals was also recorded in the notebook. The data entered in the smartphone were also collected remotely by 3G Internet connection in a distant FTP. At the end of the 1-week trial, a patient visit was scheduled to collect the smartphone and the logbook, to answer a questionnaire about the ease of application use, and to record their proposals to improve the application.

Statistical Analysis

The preliminary descriptive analysis included frequencies for categorical variables and means ± standard deviation for continuous ones. Since the observations are not independent because they are repeated by the same patients and for different types of meals, mixed effect models were used to account for the patient effect as a random effect and the effect of the type of meal as a fixed effect.

To assess whether the LIP and the PRO contents of the meal had an impact on blood glucose levels, we used a linear mixed effect model with LIP and PRO estimations as fixed effects, after accounting for the CHO estimation, the type of meal (fixed effects), and the patient (random effect). The impact of LIP and PRO estimations on the glucose status of the patient after the meals (normal, hypoglycemia, hyperglycemia) was modeled by a mixed effect multinomial logistic regression model with patient as a random effect, type of meal, LIP, PRO and CHO as fixed effects.

Finally, the evolution of the time for meal entry in the application was also analyzed using a linear mixed model with time as an additional fixed effect and then with age as an additional fixed effect. Statistical significance was set at 5%. All analyses were carried out using the SAS/UNIX statistical software (SAS version 9, SAS Institute, Cary, NC; proc freq, proc univariate, proc mixed, proc glimmix).

Results

The patients (4 males, 6 females) who participated in the study were 36 ± 13 years old, their HbA1c level was 7.6 ± 0.6%, their last learning session for CHO counting had occurred 35 ± 13 months ago, and each patient took an average of 19 ± 4 meals during the study week (Table 1).

Table 1.

Characteristics of Included Patients.

ID number	1	2	3	4	5	6	7	8	9	10
Age (years)	32	34	38	61	58	30	29	21	33	25
Gender (male or female)	F	F	M	M	F	F	F	F	M	M
Time since last CHO counting training (months)	48	48	18	48	36	>48	48	33	16	9
HbA1c at inclusion (%)	8.2	8.0	7.3	8.5	7.5	7.1	7.3	7.7	7.8	6.5
Number of meals per week	20	19	13	24	23	17	19	24	12	16

Overall, we collected data from 187 meals taken by the 10 patients. The meals included 57 breakfasts, 64 lunches, 63 dinners and 3 snacks. The snacks were taken by a single patient and were removed from analysis to obtain sufficient power for the mixed effect model. Glucose levels were within the normal range after 118 meals, in the hypoglycemic range after 37 meals, and above glucose target after 32 meals.

Table 2 presents meal component statistics, depending on the meal type. The average meal CHO content was 74.37 ± 31.78 grams, with a significant difference between breakfast and lunch/dinner. The average LIP content per meal was 20.26 ± 14.28 grams and the average PRO content per meal was 25.68 ± 16.68 grams. There was a significant difference between breakfast and lunch/dinner for LIP and PRO contents. The relative CHO meal content was 63.8 ± 13.8%, the relative LIP meal content was 16 ± 8% and the relative PRO meal content was 20 ± 9%.

Table 2.

Meal Nutritional Components as Reported by Study Patients.

	All meals		Breakfast		Lunch		Dinner		P value
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
CHO (g)	74.37	31.78	57.28	27.36	79.86	23.46	86.14	35.68	<.001
LIP (g)	20.26	14.28	13.32	9.80	23.97	16.46	22.76	13.23	<.001
PRO (g)	25.68	16.68	11.49	5.73	35.86	16.60	28.18	14.60	<.001
CHO-rel (%)	63.80	13.80	70.50	14.3	58.20	12.20	63.50	12.50
LIP-rel (%)	15.90	8.20	15.40	9.80	16.20	7.70	15.90	7.20
PRO-rel (%)	20.30	9.40	14.10	5.80	25.60	9.50	20.60	8.60

CHO, carbohydrates; LIP, lipids; PRO, proteins. X-rel values correspond to the calculation of X / (CHO + LIP + PRO); they show the distribution of the 3 main nutritional components of meals. P value indicates the statistical difference between breakfast versus lunch and dinner.

We then assessed the impact of LIP and PRO recorded amounts per meal on 2-hour postmeal glucose levels and we found no influence of LIP and PRO contents. The same parameters had also no impact on postmeal glucose status, that is, glucose in normal range, hypo- or hyperglycemia. We also assessed whether LIP and PRO meal contents impacted blood glucose levels before the next meal when no correction (insulin injection, food intake) or physical exercise had been made in the interval and found no significant relationship (Table 3).

Table 3.

Impact of Meal Lipids (LIP) and Proteins (PRO) on Early and Late Postmeal Glucose Control: Mixed Model Results of the LIP and PRO After Accounting for the CHO Estimation and Type of Meal.

		Postmeal blood glucose level
		2-hour measured	Before next meal
Number of assessable meals		127	100
Fixed effects
CHO		0.15 ± 0.29 (P = .59)	0.31 ± 0.31 (P = .32)
LIP		−0.15 ± 0.65 (P = .8)	0.77 ± 0.63 (P = .22)
PRO		−0.07 ± 0.64 (P = .9)	−0.39 ± 0.67 (P = .57)
Meal	Dinner (ref)
	Breakfast	−1.35 ± 20.14	−27.06 ± 20.41
	Lunch	−4.86 ± 18.42	−11.26 ± 19.40
Random effects
Patient		249.68 ± 424.69	425.65 ± 461.84
Residual		6510.32 ± 873.76**	4524.71 ± 693.40**

Estimate ± standard error are provided for fixed and random effects.

^**P < .01.

Finally, we evaluated the time needed to enter a meal in the application, and we found that it took on average 175 ± 122 seconds. This average time was reduced significantly (P = .0003) from the beginning of the week when it was around 220 seconds to the end of the week when it stabilizes around 150 seconds. We also observed that younger patients needed less time to enter a meal in the application than older patients (P = .017). Patient-collected feedback on the application use showed that patients were overall satisfied with the application, found it easy to use, and adapted themselves easily to the application interface (Table 4). Suggestions were made on how to improve the user interface of the application and on new functionalities to implement in the next version.

Table 4.

Satisfaction Questionnaire: Number of Patients Who Gave the Different Scores to Each Question.

Question	Score
	1	2	3	4	5
Have you been satisfied with the application use? (1 = not satisfied, 5 = very satisfied)	0	0	3	7	0
Do you think the app would be helpful daily? (1 = useless, 5 = very helpful)	0	1	5	2	2
Is the application easy to use? (1 = very hard, 5 = very easy)	0	0	1	6	3
How do you estimate your adaptation to the app? (1 = very hard, 5 = very easy)	0	0	1	4	5

Discussion

The primary objective of our study was to assess whether the Android application on smartphone that we developed was able to capture CHO, LIP, and PRO contents of meals from declarations by T1D patients of their daily food intakes. It appeared that patients reported no significant issue in entering meal components in the application and the application could compute CHO, LIP, and PRO contents of the food intake as recorded by the patients. Lunches and dinners included higher CHO, LIP, and PRO contents than breakfasts. The relative content of CHO and PRO was also higher in lunches and dinners while LIP showed the same proportion in all meal types. Nevertheless, we observe that reported CHO intakes were almost the same as in the overall French population but PRO intakes were 10% lower and LIP intakes almost 30% lower.¹⁰ The patient interviews performed at the last study visit disclosed that patients frequently failed to enter in the application the dressing and cooking components of their food intakes. This may be coherent with the lower LIP part of their food as reported in the application when compared to the average French population. Hence, specific education would be needed to prevent this omission. Of note, patients also reported that including all meal components in the application instead of considering CHO-rich food only as they usually do was perceived as burdensome and time-consuming. The time needed to enter data in the application could be significantly reduced, and then be more convenient, if CHO-rich food entry would only be requested.

An ancillary objective of the study was to look whether LIP and PRO contents of the daily meals had some impact on blood glucose levels while insulin doses were computed according to CHO component. Although the view on postmeal glucose was limited to 2-hour postmeal values which was available in only 127 meals out of 187, reported status of postmeal glucose in or out of target range and blood glucose levels before the next meal when no correction or physical exercise had been made in the interval (applicable in 100 cases), no significant impact of LIP and PRO meal contents could be identified. This observation is in contrast with previous studies showing the hyperglycemic effect in late postmeal period (3-5 hours after meal intake) of predefined LIP and PRO meal contents.⁸ Even if recordings of LIP and PRO appeared to be lower than true intakes, it is likely that in real-life patients with T1D are not prone to high-fat and/or high-protein meals. Hence, data from studies based on predefined high-fat and/or high-protein contents may not be frequently applicable in daily life. Such types of meals may be restricted to unusual situations which seldom occur. Besides, some controversy still remain about the real impact of LIP and PRO on postmeal glucose levels as shown by some studies reporting rather reduced insulin needs when CHO were combined with LIP and PRO.¹¹

We are aware of the limitations of our study. The quite small number of patients and the short duration of the study are evident limitations of the study and extensions to more numerous patients and longer duration of the experiment could provide stronger data. Concerning the accuracy in capturing meal contents by the application, we lack references for the true food intakes in comparison to those reported by the patients. Pictures of the meals taken by the patients could provide such information and should be considered in an extended assessment of the application usefulness. Since the patient interviews argue in favor of underreporting of LIP and perhaps PRO meal contents, the absence of identified significant effect of LIP and PRO components on available blood glucose values while insulin doses were computed according to CHO intakes only does not preclude true effects of LIP and PRO meal contents on glucose control. A clear limitation to our observation is the relatively scarce data on postmeal glucose levels. Only 2-hour postmeal levels could be obtained in only 127 meals out of 187 which were taken during the study period. Most studies which reported impacts of LIP and PRO components on postmeal glucose control included 3-hour or later postmeal glucose values.⁸ The reported need for postmeal corrections for out-of-range glucose levels and the assessment of glucose values before the next meal can hardly compensate for the lack of predefined late (3 to 5 hours) postmeal glucose measurements.

Conclusion

Our study shows that our developed application to capture food intakes in patients with T1D is easily usable in daily real-life practice. It provides valuable information on the CHO, LIP, and PRO components of meals taken by these patients; it would benefit from a more extensive study in more patients and on a longer duration. Besides, a specific learning of patients would be needed to avoid omissions in the reported intakes. More accurate inputs from the patients may however increase the time for data capture. Including specific predefined items for dressings and cooking adjuncts in the application could prevent this risk of unacceptable burden. The usefulness of this additional information on meal components for a better management of meal insulin doses cannot be reliably assessed from our observation that LIP and PRO components looked as having few impacts on glucose control in common life. Further investigations will be needed to assess more precisely whether capturing LIP and PRO meal contents could be helpful for each meal or only in occasional cases of high-fat and/or high-protein meals. Testing the application in closed-loop conditions could answer whether it should be incorporated in artificial pancreas models to modulate insulin delivery for allowing better postmeal glucose control.