A Randomized Controlled Study of an Insulin Dosing Application That Uses Recognition and Meal Bolus Estimations

Abstract

Background:

Throughout the insulin pump therapy, decisions of prandial boluses programming are taken by patients individually a few times every day, and, moreover, this complex process requires numerical skills and knowledge in nutrition components estimation. The aim of the study was to determine the impact of the expert system, supporting the patient’s decision on meal bolus programming, on the time in range of diurnal glucose excursion in patients treated with continuous subcutaneous insulin infusion (CSII).

Methods:

The crossover, randomized study included 12 adults, aged 19 to 53, with type 1 diabetes mellitus, duration ranging from 7 to 30 years. Patients were educated in complex food counting, including carbohydrate units (CU) and fat-protein units (FPU). Subsequently, they were randomly allocated to the experimental group (A), which used the expert software named VoiceDiab, and the control group (B), using a manual method of meal-bolus estimation.

Results:

It was found that 66.7% of patients within the A group statistically reported a relevant increase in the percentage (%) of sensor glucose (SG) in range (TIR 70-180 mg/dl), compared to the B group. TIR (median) reached 53.9% in the experimental group (A) versus 44% within the control group (B), P < .05. The average difference in the number of hypoglycemia episodes was not statistically significant (–0.2%, SD 11.6%, P = .93). The daily insulin requirement in both groups was comparable—the average difference in total daily insulin dose between two groups was 0.26 (SD 7.06 IU, P = .9).

Conclusion:

The expert system in meal insulin dosing allows improvement in glucose control without increasing the rates of hypoglycemia or the insulin requirement.

Keywords

type 1 diabetes bolus calculator randomized control trial insulin pumps

Since insulin pumps are widely used for diabetes treatment, the supporting systems commonly called the bolus calculators, have also been developed to support patients’ decisions on insulin dosing.¹ Furthermore, it is a well-known fact that the metabolic control in diabetes measured in HbA1c and blood glucose values strongly correlate with the appropriate insulin dose, related to the size and composition of the meal that patient is planning to eat.² On the other hand, the patients—insulin pump users or their caregivers, are making decisions on the type of bolus and amount of insulin doses programmed for a meal on the daily basis. In this complex process, a number of variables are to be taken into account in a mathematical formula.³ We also observe, how this very decision is influencing individual patients’ emotions and their personal experiences affect the final results of insulin applying, for example, people who have declared severe hypoglycemia may be eager to reduce the insulin dose.⁴

For the available bolus calculators, the prandial insulin is counted on the basis of carbohydrates contained in the meal—it is either entered by the patient or calculated automatically after typing the name and size of the food product in the food database.⁵ However, not all calculators include a database of food products. In the bolus calculators, one might set individual parameters such as insulin to carb ratio (ICR). Currently two models of ICR are used: the one that is popular in the United States, with a designated amount of carbohydrate per unit of insulin, and another one, used in many European countries, where the stable variable of 10 grams of carbohydrates is provided for a particular meal and a dose of insulin is fixed for a 1 carb unit (CU; carbohydrate units; 10 grams of carbs). In addition, the insulin sensitivity factor, applied to correct hyperglycemia conditions, is also established in calculator parameters.⁶

Another issue is the concept of the principle rules playing an important role in the meal insulin calculation.

Today, according to the standard of functional intensive insulin therapy, only the carbohydrate content matters. The authors of a new VoiceDiab bolus calculator suggest a new formula, where the insulin dosage is based on the total calorie value of a meal.

Since the moment continuous glucose monitoring systems have become commonly available and used, many individual observations of blood glucose show an increase of glycemia (particularly) after meals rich in fats, even up to 8 hours following a meal. There are also clinical trials proving that the rise of one’s blood glucose level lasts longer after consuming fats rather than carbohydrates.^7-9 Currently, there are studies confirming that thesis, indicating an increase for insulin demand following the meals that are rich in fats and proteins.^10,11

The new insulin dosing algorithm was set up in relation to the use of three types of auxiliary boluses. The algorithm allows to adapt a precise insulin dose and its release time, corresponding to a specific meal. This algorithm is also available as Diabetics, software that has been described in a scientific publication.¹² Its safety has been proved in prospective clinical trials and observations on a large group of patients with type 1 diabetes mellitus.¹³ This program has also been used as a foundation for the presented application.

A fundamental assumption of the concept is to calculate the dose of prandial insulin considering the total calories of the entered meal, not only for amount of carbohydrates. It means that the regular bolus dose should be enlarged so to reach an adequate bolus, calculated for the proteins and fats, and delivered over a certain period.

The primary purpose of the authors of the VoiceDiab’s bolus calculator was to help patients with the calculation of the full dose of insulin, to obtain satisfactory results of a daily glycemic profile. In addition, its goal was to facilitate the process and decision-making process in meal-insulin programming.

Methods

The randomized, two-arm, crossover trial conducted to evaluate the impact of the bolus calculator on the blood glucose profile in adults patients with type 1 diabetes treated with insulin pumps.

The primary end point was the impact of the expert system on the changes occurring over the time in range of diurnal glucose value. It measures the time spent within the recommended range, from 70 mg/dl to 180 mg/dl in the experimental group, compared to the control group.

The study has been conducted in outpatient clinics, in accordance with Good Clinical Practice, and the protocol was approved by the local ethics committee (KB11/06/2014). The eligibility criteria included having type 1 diabetes, utilizing continuous subcutaneous insulin infusion (CSII), and being above 18 years of age. Patients with poor metabolic control (hemoglobin A1c [HbA1c] > 10%), viral or bacterial infection, physical or mental limitations, or lack of continuous computer or web access were excluded.

Study Group

After the screening visit, 14 patients 18 years or older with type 1 diabetes from a recognized base using ADA criteria have been enrolled. One patient missed the study before a randomization and another one quit the process due to unwillingness to follow the study procedures.

Intervention

The VoiceDiab expert system supporting the patients’ decision on meal insulin dosing was the intervention in the experimental (A) group. It is an application dedicated to smart phones, featuring a new system that includes the software for automatic speech recognition and the software for morphological language analysis combined with a nutrient database (887 products) and a meal-insulin dose calculation algorithm. During testing the application, the patient describes vocally the meal that is planned to be consumed. The patient provides the details, such as a dish size, for example, a beef roast (200 grams), boiled potatoes (120 grams), lettuce with the vinaigrette sauce, and vanilla ice cream (a small serving of 100 grams). After recording, the user receives the information about carbohydrates, proteins, and fats described by the units of measure (such as grams, carbohydrates units, fat protein units) and total calories of the meal, displayed on the app screen. Then, after confirmation, the user receives information about the recommended type of bolus—normal wave, dual wave, or extended—and a dose of insulin needed to be given and a recommended time for the extended bolus (Figure 1).

Figure 1.

The diagram of the VoiceDiab expert system. (1) Picture presenting the meal. (2) Application screen in the stage of describing by voice the meal and below information about the value of total calories, carbohydrate, fat, and protein contents in this meal. (3) Screen presenting the type of bolus with dose of insulin for each part and time for extended bolus.

The Meal-Insulin Algorithm

A prandial insulin dose is calculated according to the rules of the Diabetics algorithm software.

It is calculated for total calories value of carbohydrate, fat and protein products in meal. Carbohydrate units—CHO units (CU)—were defined as 10g of CHO product, and a new measure—a fat-protein unit (FPU), being an equivalent to 100 kcal of fat or/and protein food, was added up for calculation purposes. The individual patient’s insulin to carb ratio (ICR) is included in the calculation of the insulin dose. When ICR is introduced into the equation, the insulin dose is multiplied by the number of CHU and FPU. This algorithm does not consider the current blood glucose concentration, the target blood glucose, and the duration of insulin action.

Procedure

During the second visit, the patients were randomly allocated into an experimental group (A) that used the mobile phone application and a control group (B), applying traditional manual calculations. The scheme of study procedure is shown in Figure 2.

Figure 2.

The schema of the study procedures.

For all patients from both groups, the insulin titration, including the basal rate and the ICR, was carried out on an individual basis. For each subject allocated to the A group, individual ICR has been set up in the application. A blood sample to assess HbA1c was taken at V2. Metabolic outcomes: the interstitial glucose (sensor glucose SG) value has been assessed by the continuous glucose monitoring system DexCom (San Diego, CA) over the study time, lasting four days for each subgroup (experimental and control). Patients wore blinded CGM. Both insulin pumps and CGM were read by the Diasend system (Diasend AB, Sweden), while Medtronic insulin pumps were supported by Carelink software (Northridge, CA). After the CGM personal data collection, insulin titration was reconciled in both study groups. In the whole group, only short-acting insulin analogues (aspart, lyspro) were used.

Education

At V2, all patients from each surveyed group were trained to calculate the number of CU and FPU and, finally, to adapt the type of bolus for a specific meal and to program the time of the square-wave bolus (S-W). Furthermore, the patients allocated to the experimental group were educated how to use the application, through the manual (version Pl 1.0 01.09.2015) they received, to learn how information about meals should be recorded. All the subjects had been trained on recording the meals in the food diary and on the procedures applied to the CGM system. Throughout the study the patients recorded information about the meals they consumed: the time of the meal, the meal’s products, and the size of each meal’s components. In addition, meal descriptions in the experimental group have been noticed by the ASR server.

Statistical Analysis

All numerical data are presented as mean ± SD. Shapiro-Wilk’s W test was applied to check the regularity of distribution related to analyzed variables. The results indicated that the distribution of all assessed parameters, except the percentage of the continuous glucose monitoring data in the hypoglycemic range, might be considered regular. Thus, in the case of all the normally distributed variables, the parametric Student’s t-test for independent variables was used to assess significance of differences related to means applied over particular periods with and without application. Student’s t-tests for dependent variables were applied to assess whether the mean differences of the analyzed variables in these two periods reached the values significantly varying from zero. In the case of the percentage of the continuous glucose monitoring results in the hypoglycemic range, the corresponding nonparametric tests were applied, that is, the Mann-Whitney U test and Wilcoxon’s test, respectively, instead of Student’s t-tests. The results were considered to be statistically significant for P < .05.

Results

The data gathered from 104-day records (12 patients) were considered the grounds for a statistic evaluation.

In the study group, 12 patients with type 1 diabetes (five males and seven females) finished two arms of the study; the average age was 27.9 years (range 19-53), the average diabetes duration was 14 years (range 7-30), with av. HbA1c 7.25% (range 5.7-9.4).

The difference in blood glucose range was assessed individually for each patient. We have selected three ranges of blood glucose levels: (I) recommended range—TIR 70-180 mg/dl, (II) hyperglycemia range—above 180 mg/dl; and (III) hypoglycemia range—below 70 mg/dl. It was found that 66.7% of patients within the experimental (A) group got statistically significant increases in the percentage of SG events in the recommended range, compared to the control group. TIR (median) was 53.9% in the experimental (A) group versus 44% in the control group (B) (P < .05). The statistically significant improvement in the daily glucose excursion was observed among 10 of 12 patients (Table 1). In terms of hypoglycemia, average episodes of SG below 70 mg/dl were higher in the experimental group. Six out of 12 subjects had experienced more SG episodes. However, an average difference in hypoglycemia episodes was not statistically significant (–0.2%, SD 11.6%, P = .93; Figure 3).

Table 1.

Mean Values of Sensor Glucose in Range (70-180 mg/dl) in the VoiceDiab and Control Groups, With the Partial Differences Between Two Groups.

Number	Sensor glucose—time in range (70-180 mg/dl) (%)
Number	Control group B, non-VoiceDiab	Experimental group A, VoiceDiab	Mean difference	P value
1	41.7	40.7	1.1	.69
2	47.8	77.6	29.8	.000
3	26.6	66	39.4	.000
4	44	53.6	9.6	.000
5	19.9	40	20.1	.000
6	58.1	68.8	10.7	.000
7	62.8	68.7	5.9	.005
8	38.9	43.2	4.4	.041
9	62.7	39.7	−23.1	.000
10	62.5	60.9	−1.6	.45
11	38	54.2	16.2	.000
12	70.1	48.6	−22.1	.000

Figure 3.

The partial differences in hypoglycemia episodes measured as a percentage of episodes bellow 70 mg/dl recorded by CGM between the VoiceDiab and control (non-VoiceDiab) groups. The average difference was 0.2% (SD 11.6%, P = .953).

The daily insulin requirement in both groups was comparable, an average change in daily total insulin dose was 0.26 (SD 7.06 IU, P = .9). In seven subjects, the basal insulin dose decreased; in two patients, it did not change, and an average difference was 0.14 (SD 1.65 IU, P = .77).

The boluses dose of insulin in the group A increased in 7 of 12 patients; an average difference was 0.40 (SD 6.91 IU, P = .84).

Discussion

The VoiceDiab application is a unique expert bolus advisor aimed at insulin-treated people. For the first time, speech recognition software, a food database, and an insulin algorithm have been integrated into one system, supporting decisions in meal insulin programming. The main idea was to facilitate the process of meal bolus adjusting in a daily routine, through matching the insulin dose and the type of bolus so to fit the meal. The patient was supposed to not count carbohydrates and calories contained in protein and fat within the meal he or she selected. They provided information only about the components of meals, using a voice system built into the mobile phone.

An assessment of the system’s voice recognition accuracy was performed in an additional study, which was carried out before the clinical research presented in this article. It is noteworthy that the presented new system successfully passed a series of technical tests in patients in different age groups demonstrating that the accuracy of the ASR server in transforming verbal description of meals into written text was equal to 97% (without repeating the verbal description by the user) and that only in 1.5% cases the system failed to recognize the content of the meal even though the verbal description was repeated for three or more times.¹⁴

In our study we found out that the majority of patients using the apps and treated with an insulin pump reported 54% of diurnal glucose level in a recommended range. Nearly 70% of patients improved their diurnal glucose profile, when supported by the VoiceDiab. We need to emphasize that we present a preliminary study with a short observation time.

The literature often claims that the Diabetics algorithm (considering a larger dose of insulin calculated also for fat and protein contained in food) causes a high risk of hypoglycemia.^7,11 We noticed that in both arms (A and B), the rate of hypoglycemia (<70 mg/dl) was very low and at the level of around 6%, although the percentage of hypoglycemia increased by 1.0%. Only in one subject participating in our study the hypoglycemia events increased up to 30%, while in others it decreased to 0.9%. Despite this one case, we have found the insulin algorithm implemented in the expert system to be safe.

In the Kordonouri et al study,¹⁵ focused on comparing two systems of meal boluses, the authors highlighted higher rates of hypoglycemia episodes among patients using a complex option. But in other research, Blazik and Pankowska,¹⁶ in an RCT where the same algorithm was used, the hypoglycemia index ranged between 4.9 and 4.7 (hypoglycemia <70mg/dl) within the whole surveyed group. The rational for the different risk of hypoglycemia can be the higher basal insulin dose. In the Blazik and Pankowska study, the basal insulin was close to 30-25% of a total daily dose, usually lower then it is practiced in other diabetes schools.¹⁷

It seems to be interesting to compare hypoglycemia results obtained in our study on the system linking insulin pumps with CGM, where an insulin inflow—considered the base in the event the glucose drop—is automatically ceased, based on the model predicting hypoglycemia. This research proved that the rate of hypoglycemia reached 15%¹⁸ and it was triple the results obtained in the VoiceDiab project with complex meal insulin dosing. Therefore, when referring to our result, we can say that neither this algorithm nor the expert system VoiceDiab increase the risk of hypoglycemia. Although it doesn’t include the algorithm of the insulin on board (IOB), the rate of hypoglycemia has been very low. IOB is incorporated in all bolus calculators integrated with insulin pumps. On the other hand, in the analysis conducted by Schmidt and Norgaard, while reviewing the outcomes from bolus calculators in terms of the hypoglycemia risk, no differences related to the algorithm using or not using IOB, have been found.³

According to our survey, more accurate food component calculations carried out by the software have an impact on the increase of a bolus dose of insulin. This observation leads into the discussion on an insulin adjustment adopted for the bolus calculator. It would be advisable to carry out both research studies and debate on the establishment of new insulin parameters: insulin ratio or share of a base insulin in a 24-hour profile, and the calculation should cover the insulin content versus total meal calories and not only refer to the carbohydrates. It seems to be adequate to decrease the ICR when dual-wave bolus with an additional dose of insulin for fat and protein is programming. This study shows how important it is to open a discussion on the bolus advisor protocol and the insulin algorithm. Moreover, it needs to cover the issue of the insulin algorithm for carbohydrate products or for whole components (CU and FPU).

Currently, there are numerous kinds of software on the market, addressed to people with diabetes. Moreover, there are some applications available on the market, that include insulin bolus calculators built in smartwatches.¹⁹ It is important to emphasize that most of the software has not been validated by clinical studies.^20,21 We are aware that only one of them, “Diabeo,”²² combined with the tele-consultation system, has been validated by an in-depth legitimate study.

Conclusion

The VoiceDiab algorithm, focused on the adaptation of insulin dosing for the caloric content of a meal, using the voice speech system, has the potential to support the diet personalization process. The patients’ opinions about the application functionality in everyday life have been presented in an additional report. Finally, this preliminary analysis enlarges the base of evidence necessary for verifying the software supporting patients in insulin dosing.

Footnotes

Abbreviations

ASR, automatic speech recognition; CGMS, continuous glucose monitoring system; CSII, continuous subcutaneous insulin infusion; CU, carbohydrate units; DAFNE, Dose Adjustment for Normal Eating; FPU, fat and protein units; ICR, insulin to carbohydrate ratio (insulin dose for 10 grams of carbohydrate); IOB, insulin on board; N, normal wave bolus; SG, sensor glucose; SMBG, self-monitoring of blood glucose; S-W, square-wave bolus; TIR, time in target range.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by the Polish National Center for Research and Development grant PBS1/B9/13/2012. We would like to thank the following people for their involvement: Maria Wierzchowska, the study nurse, and Katarzyna Ruszkowska, the study coordinator from the Institute of Diabetology in Warsaw.

References

Lepore

Dodesini

Nosari

. Bolus calculator improves long-term metabolic control and reduces glucose variability in pump-treated patients with type 1 diabetes. Nutr Metab Cardiovasc Dis. 2012;22(8):e15-e16.

DAFNE Study Group. Training in flexible, intensive insulin management to enable dietary freedom in people with type 1 diabetes: Dose Adjustment for Normal Eating (DAFNE) randomised controlled trial. BMJ. 2002;325(7367):746.

Schmidt

Norgaard

Bolus calculators. J Diabetes Sci Technol. 2014;8(5):1035-1041.

Delamater

Kurtz

Bubb

. Stress and coping in relation to metabolic control of adolescents with type 1 diabetes. J Dev Behav Pediatr. 1987;8(3):136-140.

Shapira

Yodfat

Fau - HaCohen AravaHaCohen

Fau – Feigin . Bolus guide: a novel insulin bolus dosing decision support tool based on selection of carbohydrate ranges. J Diabetes Sci Technol. 2010;4(4):893-902.

Alcantara-Aragon

Gonzalez

Corcoy

. Carbohydrate-to-insulin ratio in a Mediterranean population of type 1 diabetic patients on continuous subcutaneous insulin infusion therapy. J Diabetes Sci Technol. 2015;9(3):588-92.

Smart

Evans

M.O’Connell

S. M

. Both dietary protein and fat increase postprandial glucose excursions in children with type 1 diabetes, and the effect is additive. Diabetes Care. 2013;36(12):3897-3902.

Neu

Behret

Braun

. Higher glucose concentrations following protein- and fat-rich meals—the Tuebingen Grill Study: a pilot study in adolescents with type 1 diabetes. Pediatr Diabetes. 2015;16(8):587-591.

Lodefalk

Aman

Bang

Effects of fat supplementation on glycaemic response and gastric emptying in adolescents with type 1 diabetes. Diabet Med. 2008;25(9):1030-1035.

10.

Bell

Smart

Steil

. Impact of fat, protein, and glycemic index on postprandial glucose control in type 1 diabetes: implications for intensive diabetes management in the continuous glucose monitoring era. Diabetes Care. 2015;38(6):1008-1015.

11.

Wolpert

Atakov-Castillo

Smith

. Dietary fat acutely increases glucose concentrations and insulin requirements in patients with type 1 diabetes: implications for carbohydrate-based bolus dose calculation and intensive diabetes management. Diabetes Care. 2013;36(4):810-816.

12.

Pankowska

Blazik

Bolus calculator with nutrition database software, a new concept of prandial insulin programming for pump users. J Diabetes Sci Technol. 2010;4(3):571-576.

13.

Pankowska

Szypowska

Lipka

. Application of novel dual wave meal bolus and its impact on glycated hemoglobin A1c level in children with type 1 diabetes. Pediatr Diabetes. 2009;10(5):298-303.

14.

Foltynski

PŁ

Pańkowska

Mazurczak

Migalska-Musiał

An algorithm based on voice description of meal insulin dose calculation compensate food intake. IFMB Proc. 2015;51:1441-1444.

15.

Kordonouri

Hartmann

Remus

. Benefit of supplementary fat plus protein counting as compared with conventional carbohydrate counting for insulin bolus calculation in children with pump therapy. Pediatr Diabetes. 2012;13(7):540-544.

16.

Blazik

Pankowska

The effect of bolus and food calculator Diabetics on glucose variability in children with type 1 diabetes treated with insulin pump: the results of RCT. Pediatr Diabetes. 2012;13(7):534-539.

17.

Holterhus

Bokelmann

Riepe

. Predicting the optimal basal insulin infusion pattern in children and adolescents on insulin pumps. Diabetes Care. 2013;36(6):1507-1511.

18.

Choudhary

Olsen

Conget

. Hypoglycemia prevention and user acceptance of an insulin pump system with predictive low glucose management. Diabetes Technol Ther. 2016;18(5):288-291.

19.

Arsand

Muzny

Bradway

. Performance of the first combined smartwatch and smartphone diabetes diary application study. J Diabetes Sci Technol. 2015;9(3):556-63.

20.

Hirsch

Parkin

CG.

Unknown safety and efficacy of smartphone bolus calculator apps puts patients at risk for severe adverse outcomes. J Diabetes Sci Technol. 2016;10(4):977-80.

21.

Klonoff

DC.

The current status of bolus calculator decision-support software. J Diabetes Sci Technol. 2012;6(5):990-4.

22.

Charpentier

Benhamou

Dardari

. The Diabeo software enabling individualized insulin dose adjustments combined with telemedicine support improves HbA1c in poorly controlled type 1 diabetic patients: a 6-month, randomized, open-label, parallel-group, multicenter trial (TeleDiab 1 Study). Diabetes Care. 2011;34(3):533-539.