Qualitative meal-size estimation has been proposed instead of quantitative carbohydrate (CHO) counting with automated insulin delivery. We aimed to assess the noninferiority of qualitative meal-size estimation strategy.
We conducted a two-center, randomized, crossover, noninferiority trial to compare 3 weeks of automated insulin delivery with 1) CHO counting and 2) qualitative meal-size estimation in adults with type 1 diabetes. Qualitative meal-size estimation categories were low, medium, high, or very high CHO and were defined as <30 g, 30–60 g, 60–90 g, and >90 g CHO, respectively. Prandial insulin boluses were calculated as the individualized insulin to CHO ratios multiplied by 15, 35, 65, and 95, respectively. Closed-loop algorithms were otherwise identical in the two arms. The primary outcome was time in range 3.9–10.0 mmol/L, with a predefined noninferiority margin of 4%.
A total of 30 participants completed the study (n = 20 women; age 44 (SD 17) years; A1C 7.4% [0.7%]). The mean time in the 3.9–10.0 mmol/L range was 74.1% (10.0%) with CHO counting and 70.5% (11.2%) with qualitative meal-size estimation; mean difference was −3.6% (8.3%; noninferiority P = 0.78). Frequencies of times at <3.9 mmol/L and <3.0 mmol/L were low (<1.6% and <0.2%) in both arms. Automated basal insulin delivery was higher in the qualitative meal-size estimation arm (34.6 vs. 32.6 units/day; P = 0.003).
Though the qualitative meal-size estimation method achieved a high time in range and low time in hypoglycemia, noninferiority was not confirmed.
Introduction
Automated insulin delivery systems use a mathematical dosing algorithm that takes real-time readings from a glucose sensor to titrate the infusion of an insulin pump (1). Current automated insulin delivery systems improve glycemic control compared with pump therapy and sensor-augmented pump therapy (2,3). However, because of delays in subcutaneous insulin absorption (4) compared with meal glucose absorption, users still need to manually initiate preprandial insulin boluses. Moreover, because meal carbohydrate (CHO) content is the major nutritional determinant of postprandial glucose levels in type 1 diabetes (5,6), the user is also required to count and enter the CHO content of meals in the system to determine the prandial insulin boluses.
Despite its benefits (7), CHO counting is a burdensome and error-prone task (8), with an estimation error of around 20% (7,9). Having to count the CHO content of every meal can make people feel restrained, anxious, and less confident, especially when interacting with peers around food (10). Moreover, the relative ease provided by nutrition facts labels encourage the selection of prepackaged, potentially unhealthy, processed foods over whole foods, such as whole grains and fruits (11). Additionally, people with type 1 diabetes tend to standardize their meal choices (10) to facilitate CHO counting, which may further negatively influence dietary choices. Therefore, eliminating or simplifying CHO counting with advanced automated insulin delivery systems is highly desirable (12).
Early automated insulin delivery systems attempted to eliminate the need for CHO counting by relying reactively on glucose-sensor readings to cover meal-related insulin needs. However, this approach resulted in prolonged postprandial hyperglycemia due to delays in insulin absorption (13,14). Other intermediate approaches, such as simple meal announcement (15) (pressing a button) or qualitative meal-size estimation (16) (choosing if the meal is low, medium, high, or very high CHO), have been proposed but, to our knowledge, none have been compared with the gold standard, quantitative CHO counting in outpatient free-living settings when both approaches are combined with an automated insulin delivery system.
Here, we present results of a randomized trial, comparing our automated insulin delivery system with either conventional CHO counting or a simplified, qualitative meal-size estimation in adults with type 1 diabetes over two periods of 3 weeks. The system used the iPancreas platform, which has been used in other studies (17,18), and a previously described model predictive control algorithm (19,20).
Research Design and Methods
Study Design
We conducted an open-label, two-center, randomized, crossover, noninferiority study in type 1 diabetes to compare an automated insulin delivery system with either CHO counting or qualitative meal-size estimation. Each intervention lasted 3 weeks in outpatient, unsupervised, free-living settings with no remote monitoring. The intervention visits were separated by a median of 6 (interquartile range [IQR] = 4, 8) days.
Participants
From July 2020 to December 2021, participants were enrolled at the Montreal Clinical Research Institute and the Research Institute of McGill University Health Centre, Montreal, Canada. Participants were required to be aged ≥18 years, diagnosed with type 1 diabetes for at least 1 year, and using an insulin pump for at least 3 months. Exclusion criteria included HbA1c ≥10.0%; clinically significant nephropathy, neuropathy, and/or retinopathy; an acute macrovascular event within 6 months of screening; pregnancy; severe hypoglycemia within 2 weeks of screening; and diabetes ketoacidosis within 3 months of screening. Participants were required to remain within a travel distance of 3 hours of Montreal during the interventions. Participants provided written informed consent. The study was approved by the local institutions’ ethics committees.
Randomization and Masking
We used blocked randomization to generate allocation sequences, which were disclosed after the admission visit. Participants and investigators were not blinded to the allocation.
Study Procedures
At the admission visit, a research team member (nutritionist or a qualified diabetes educator) asked participants about their usual meals, assessed their CHO counting skills, and trained participants for the simplified qualitative meal-size estimation strategy. After the study admission visit, participants had a run-in period of 12 days with a glucose sensor (Dexcom G5 or G6) using their own insulin pump. Every 3 to 5 days of the run-in period, a member of the team reviewed the participant’s pump and sensor data and adjusted their CHO ratios or basal rates if significant hyperglycemia or hypoglycemia were observed. Some participants who were deemed to have difficulty counting their CHO (per the judgement of the team member during the admission visit) were asked to complete a food diary over 3 days during the run-in period and meet with a study nutritionist to further review their CHO-counting skills.
The automated insulin delivery system was based on the iPancreas system (Oregon Health & Science University, Portland, OR) and our insulin-dosing algorithm (20–22). The system consists of a Dexcom glucose sensor, a noncommercial t:slim TAP3 insulin pump (Tandem Diabetes Care), and a cellphone (Nexus 5; LG Electronics). On the first day of the first intervention, participants were trained on the system use at our clinical research facility or over a video conference call.
The automated insulin delivery system was initialized using participants’ total daily insulin dose, CHO ratios, and programmed basal rates. A new basal rate was calculated every 10 min on the basis of a model predictive-control dosing algorithm (20–22), which used the sensor data as input. The dosing algorithm was identical for both interventions. The computed basal rate was wirelessly communicated to the pump.
The system had one adjustable glucose target, which was set at 6.0 mmol/L for all participants. Participants were made aware of the exercise feature in the system, which would raise the glucose target by 3 mmol/L. The system does not administer automatic boluses. The system switches to open-loop mode (delivering participant’s usual basal rates) if communication with pump and sensor is lost for more than 20 and 30 min, respectively.
In the CHO-counting arm, participants were instructed to manually enter the CHO content of the meals and snacks into the system, which calculated the prandial boluses. Prandial insulin boluses were calculated as the individualized insulin to CHO ratios multiplied by the CHO content of the meals. In the qualitative meal-size estimation arm, participants were instructed to manually announce meals as having low, medium, high, or very high CHO. These categories were defined as <30 g, 30–60 g, 60–90 g, and >90 g CHO, respectively. Their prandial insulin boluses were calculated on the basis of the individualized insulin to CHO ratios multiplied by 15, 35, 65, and 95, respectively. For both interventions, correction boluses were also delivered at mealtimes on the basis of the premeal glucose level and its rate of change. Participants could also deliver manual correction boluses through the system at any time.
There are three reasons behind these ranges for the qualitative meal-size estimation strategy. First, we wanted to cover a range of CHO in which most meals would fall (including adolescents who tend to eat larger meals). Second, we wanted to have no more than four categories because more categories would likely make the task of categorization burdensome. Third, we wanted each category to cover no more than 30 g because CHO-counting error in this range can likely be covered by the automated closed-loop insulin delivery. For the prandial boluses, we aimed to use the lower end of each category out of caution to reduce the risk of hypoglycemia in this first outpatient, supervised assessment of the qualitative meal-size estimation strategy.
For both interventions, sensor alarm thresholds were determined by participants and were kept fixed. Participants were asked to treat hypoglycemia per their standard practice. Participants were contacted on the first evening of both interventions, as well as once per week afterward to discuss any unexpected events or technical problems. Study teams were also on call throughout the interventions to provide technical support. For the call on week 1, a member of the team additionally reviewed the participant’s pump and sensor data and adjusted their CHO ratios or basal rates if significant hyperglycemia or hypoglycemia were observed.
In total, participants used the system for 3 weeks in each study arm (CHO counting and meal-size estimation) separated by a wash-out period of 6 (IQR = 4, 8) days.
Study Outcomes
The primary outcome was the percentage of time that the glucose sensor readings were in the target range (3.9–10.0 mmol/L) (23), with a predefined noninferiority margin of 4%. Secondary outcomes included times spent below and above target range, fasting glucose at 0600 h, and glucose variability (23). Outcomes were calculated for the entire 3-week study period and comprised of all available data, including data from when the system was in open loop mode.
A change of 4% time in target range corresponds to around 60 min/day or 20 min per meal. We considered differences of <20 min per meal to be not clinically significant. This 4% margin is also lower than the international consensus limit of 5% as a clinically meaningful difference in target range (23), though newer international guidelines, published after the design of this study, call for investigators to power studies to detect a minimum of 3% change in time in target range (24).
Statistical Analysis
We anticipated that using qualitative meal-size estimation would yield a time in target range within 4% of CHO counting. We did a power analysis using the formula for the two-sided paired t test with a 5% significance level and SD of 10% (20,25). Assuming a dropout rate of up to 10%, we calculated that 30 participants would provide 80% power.
A linear mixed model was fitted to the data while adjusting for the period effect. To examine for carry-over effect, a model was fitted with the treatment by period interaction term. Residual values were examined for normality and, if skewed, the data were transformed using the square-root function. P < 0.05 was regarded as significant. Results are reported as median (IQR) or mean (SD). Unless otherwise stated, P values are reported for superiority tests.
Results
A total of 34 participants were enrolled in the study. One participant was excluded due to feeling uncomfortable with the technology. Three participants dropped out for reasons including geographical relocation and technical issues with the automated delivery system. The participant flowchart is presented in Supplementary Fig. 1.
A total of 30 participants completed the two interventions and were included in the analysis (n = 20 women; mean age, 44 [17] years; HbA1c 7.4% [0.7%; 57 (8) mmol/mol]; duration of diabetes, 29 [22] years; total daily insulin, 54 (26) units [0.67 (0.22) units/kg]; Table 1). Glucose sensor readings were available 90.5% of the time during the CHO counting interventions and 90.4% of the time during the qualitative meal-size estimation intervention. The system was operational in closed-loop mode 87.2% and 86.8% of the time, respectively. Figure 1 compares the glucose sensor profiles of the two interventions.
Characteristic . | Mean (SD)* . | Range (min–max) . |
---|---|---|
Female sex, n (%) | 20 (67) | NA |
Age (years) | 44 (17) | 21–76 |
Weight (kg) | 79 (16) | 48–121 |
BMI (kg/m2) | 28 (6) | 19–43 |
HbA1c (%) | 7.4 (0.7) | 5.7–8.6 |
HbA1c (mmol/mol) | 57 (8) | 39–70 |
Duration of diabetes (years) | 26 (14) | 6–60 |
Total daily insulin (units) | 54 (26) | 25–137 |
Total daily insulin (units/kg) | 0.67 (0.22) | 0.37–1.18 |
Characteristic . | Mean (SD)* . | Range (min–max) . |
---|---|---|
Female sex, n (%) | 20 (67) | NA |
Age (years) | 44 (17) | 21–76 |
Weight (kg) | 79 (16) | 48–121 |
BMI (kg/m2) | 28 (6) | 19–43 |
HbA1c (%) | 7.4 (0.7) | 5.7–8.6 |
HbA1c (mmol/mol) | 57 (8) | 39–70 |
Duration of diabetes (years) | 26 (14) | 6–60 |
Total daily insulin (units) | 54 (26) | 25–137 |
Total daily insulin (units/kg) | 0.67 (0.22) | 0.37–1.18 |
Max, maximum; min, minimum; NA, not applicable.
Unless otherwise indicated.
The time in 3.9–10.0 mmol/L was 74.1% (SD 10.0%) with CHO counting and 70.5% (11.2%) with qualitative meal-size estimation. The mean difference was −3.6% (8.3%; noninferiority P = 0.78 for a 4% margin; superiority P = 0.018; Table 2). A total of 11 participants (37%) had a difference in time in target range between interventions of <4%, 15 participants (50%) had a higher time in target range by >4% with CHO counting, and four participants (13%) had a higher time in target range by >4% with qualitative meal-size estimation (Supplementary Fig. 2). No treatment by period interaction was found, and no difference was observed due to the order of interventions (data not shown).
Outcomes‡ . | CHO counting (n = 30) . | Meal-size categorization (n = 30) . | Meal-size categorization minus CHO counting; P value¶ . |
---|---|---|---|
24-h Outcomes (08:00–08:00) | |||
Time spent at glucose levels, mmol/L, (%) | |||
3.9–10* | 74.1 (10.0) | 70.5 (11.2) | −3.6 (8.3), 0.018; (−6.5 to −0.6)§, 0.78¥ |
3.9–7.8 | 51.1 (11.6) | 46.6 (10.5) | −4.5 (9.1); 0.012 |
<3.0 | 0.2 (0.0, 0.7) | 0.1 (0.0, 0.5) | 0.0 (−0.2 to 0.1); 0.34 |
<3.9 | 1.4 (0.6, 2.9) | 1.6 (0.6, 2.8) | 0.1 (−0.6 to 0.4); 0.95 |
>7.8 | 48 (12) | 53 (11) | 5 (9); 0.0074 |
>10.0 | 24 (10) | 28 (11) | 4 (8); 0.014 |
>13.9 | 4 (2, 9) | 6 (4 to 9) | 1 (−1 to 4); 0.027 |
Mean glucose (mmol/L) | 8.2 (7.5, 8.9) | 8.6 (8.1, 9.0) | 0.4 (−0.1 to 0.7); 0.0078 |
SD of glucose (mmol/L) | 2.8 (2.6, 3.2) | 3.0 (2.6, 3.4) | 0.2 (0.0, 0.4); 0.019 |
CV of glucose (%) | 34.7 (4.9) | 35.2 (4.8) | 0.5 (3.3); 0.45 |
Basal insulin (units/day) | 26.8 (20.1, 41.2) | 29.6 (20.5, 41.8) | 1.1 (0.4, 3.0); 0.00078 |
Bolus insulin (units/day) | 21.1 (10.9) | 20.8 (10.5) | −0.3 (3.2); 0.59 |
Total insulin/kg | 0.66 (0.19) | 0.68 (0.22) | 0.02 (0.05); 0.12 |
Day outcomes (0800–2300 h) | |||
Time spent at glucose levels, mmol/L (%) | |||
3.9–10.0 | 72.9 (10.5) | 69.8 (11.9) | −3.1 (8.4); 0.052 |
3.9–7.8 | 49.9 (12.5) | 45.8 (11.4) | −4.1 (9.0); 0.017 |
<3.0 | 0.2 (0.0, 0.7) | 0.1 (0.0, 0.4) | 0.0 (−0.3 to 0.1); 0.17 |
<3.9 | 1.7 (0.6, 2.9) | 1.6 (0.6, 2.6) | −0.1 (−0.8 to 0.4); 0.34 |
>7.8 | 49 (13) | 54 (12) | 5 (9); 0.0088 |
>10.0 | 25 (11) | 28 (12) | 4 (8), 0.023 |
>13.9 | 4 (2, 9) | 6 (3, 9) | 1 (−1 to 5); 0.038 |
Mean glucose (mmol/L) | 8.3 (7.4, 9.1) | 8.5 (8.0, 9.1) | 0.4 (−0.3 to 0.7); 0.038 |
SD of glucose (mmol/L) | 2.8 (2.6, 3.2) | 3.0 (2.6, 3.3) | 0.2 (−0.1 to 0.4); 0.073 |
CV of glucose (%) | 35.0 (5.2) | 35.2 (5.0) | 0.1 (3.5); 0.83 |
Basal insulin (units/day) | 19.6 (15.1, 32.5) | 21.7 (16.0, 32.6) | 1.0 (0.1, 2.1); 0.0009 |
Bolus insulin (units/day) | 20.1 (10.1) | 19.8 (9.7) | −0.4 (3.4); 0.62 |
Total insulin/kg | 0.55 (0.15) | 0.56 (0.17) | 0.01 (0.05); 0.26 |
Night outcomes (2300–0800 h) | |||
Time spent at glucose levels, mmol/L (%) | |||
3.9–10.0 | 77.7 (11.9) | 72.7 (13.0) | −5.0 (11.2); 0.014 |
3.9–7.8 | 54.7 (15.3) | 49.2 (15.1) | −5.5 (12.9); 0.028 |
<3.0 | 0.0 (0.0, 0.9) | 0.0 (0.0, 0.1) | 0.0 (−0.3 to 0.0); 0.55 |
<3.9 | 0.7 (0.2, 2.0) | 0.7 (0.3, 1.3) | 0.1 (−0.6 to 0.6); 0.82 |
>7.8 | 45 (16) | 50 (16) | 5 (13); 0.036 |
>10.0 | 22 (12, 25) | 24 (17, 35) | 5 (−1 to 11); 0.014 |
>13.9 | 3 (1, 5) | 5 (2, 7) | 1 (−1 to 3); 0.046 |
Mean glucose (mmol/L) | 8.1 (7.4, 8.6) | 8.4 (7.6, 9.2) | 0.3 (−0.2 to 1.0); 0.025 |
SD of glucose (mmol/L) | 2.6 (0.6) | 2.8 (0.7) | 0.2 (0.5); 0.025 |
CV of glucose (%) | 32.1 (5.8) | 33.3 (6.3) | 1.2 (6.3); 0.32 |
Basal insulin (units/day) | 6.9 (5.2, 8.6) | 7.2 (5.0, 9.6) | 0.2 (−0.2 to 0.9); 0.057 |
Bolus insulin (units/day) | 0.3 (0.1, 1.0) | 0.4 (0.1, 1.1) | 0.1 (−0.1 to 0.2); 0.42 |
Total insulin/kg | 0.11 (0.08, 0.13) | 0.11 (0.08, 0.13) | 0.00 (0.00, 0.01); 0.035 |
Outcomes‡ . | CHO counting (n = 30) . | Meal-size categorization (n = 30) . | Meal-size categorization minus CHO counting; P value¶ . |
---|---|---|---|
24-h Outcomes (08:00–08:00) | |||
Time spent at glucose levels, mmol/L, (%) | |||
3.9–10* | 74.1 (10.0) | 70.5 (11.2) | −3.6 (8.3), 0.018; (−6.5 to −0.6)§, 0.78¥ |
3.9–7.8 | 51.1 (11.6) | 46.6 (10.5) | −4.5 (9.1); 0.012 |
<3.0 | 0.2 (0.0, 0.7) | 0.1 (0.0, 0.5) | 0.0 (−0.2 to 0.1); 0.34 |
<3.9 | 1.4 (0.6, 2.9) | 1.6 (0.6, 2.8) | 0.1 (−0.6 to 0.4); 0.95 |
>7.8 | 48 (12) | 53 (11) | 5 (9); 0.0074 |
>10.0 | 24 (10) | 28 (11) | 4 (8); 0.014 |
>13.9 | 4 (2, 9) | 6 (4 to 9) | 1 (−1 to 4); 0.027 |
Mean glucose (mmol/L) | 8.2 (7.5, 8.9) | 8.6 (8.1, 9.0) | 0.4 (−0.1 to 0.7); 0.0078 |
SD of glucose (mmol/L) | 2.8 (2.6, 3.2) | 3.0 (2.6, 3.4) | 0.2 (0.0, 0.4); 0.019 |
CV of glucose (%) | 34.7 (4.9) | 35.2 (4.8) | 0.5 (3.3); 0.45 |
Basal insulin (units/day) | 26.8 (20.1, 41.2) | 29.6 (20.5, 41.8) | 1.1 (0.4, 3.0); 0.00078 |
Bolus insulin (units/day) | 21.1 (10.9) | 20.8 (10.5) | −0.3 (3.2); 0.59 |
Total insulin/kg | 0.66 (0.19) | 0.68 (0.22) | 0.02 (0.05); 0.12 |
Day outcomes (0800–2300 h) | |||
Time spent at glucose levels, mmol/L (%) | |||
3.9–10.0 | 72.9 (10.5) | 69.8 (11.9) | −3.1 (8.4); 0.052 |
3.9–7.8 | 49.9 (12.5) | 45.8 (11.4) | −4.1 (9.0); 0.017 |
<3.0 | 0.2 (0.0, 0.7) | 0.1 (0.0, 0.4) | 0.0 (−0.3 to 0.1); 0.17 |
<3.9 | 1.7 (0.6, 2.9) | 1.6 (0.6, 2.6) | −0.1 (−0.8 to 0.4); 0.34 |
>7.8 | 49 (13) | 54 (12) | 5 (9); 0.0088 |
>10.0 | 25 (11) | 28 (12) | 4 (8), 0.023 |
>13.9 | 4 (2, 9) | 6 (3, 9) | 1 (−1 to 5); 0.038 |
Mean glucose (mmol/L) | 8.3 (7.4, 9.1) | 8.5 (8.0, 9.1) | 0.4 (−0.3 to 0.7); 0.038 |
SD of glucose (mmol/L) | 2.8 (2.6, 3.2) | 3.0 (2.6, 3.3) | 0.2 (−0.1 to 0.4); 0.073 |
CV of glucose (%) | 35.0 (5.2) | 35.2 (5.0) | 0.1 (3.5); 0.83 |
Basal insulin (units/day) | 19.6 (15.1, 32.5) | 21.7 (16.0, 32.6) | 1.0 (0.1, 2.1); 0.0009 |
Bolus insulin (units/day) | 20.1 (10.1) | 19.8 (9.7) | −0.4 (3.4); 0.62 |
Total insulin/kg | 0.55 (0.15) | 0.56 (0.17) | 0.01 (0.05); 0.26 |
Night outcomes (2300–0800 h) | |||
Time spent at glucose levels, mmol/L (%) | |||
3.9–10.0 | 77.7 (11.9) | 72.7 (13.0) | −5.0 (11.2); 0.014 |
3.9–7.8 | 54.7 (15.3) | 49.2 (15.1) | −5.5 (12.9); 0.028 |
<3.0 | 0.0 (0.0, 0.9) | 0.0 (0.0, 0.1) | 0.0 (−0.3 to 0.0); 0.55 |
<3.9 | 0.7 (0.2, 2.0) | 0.7 (0.3, 1.3) | 0.1 (−0.6 to 0.6); 0.82 |
>7.8 | 45 (16) | 50 (16) | 5 (13); 0.036 |
>10.0 | 22 (12, 25) | 24 (17, 35) | 5 (−1 to 11); 0.014 |
>13.9 | 3 (1, 5) | 5 (2, 7) | 1 (−1 to 3); 0.046 |
Mean glucose (mmol/L) | 8.1 (7.4, 8.6) | 8.4 (7.6, 9.2) | 0.3 (−0.2 to 1.0); 0.025 |
SD of glucose (mmol/L) | 2.6 (0.6) | 2.8 (0.7) | 0.2 (0.5); 0.025 |
CV of glucose (%) | 32.1 (5.8) | 33.3 (6.3) | 1.2 (6.3); 0.32 |
Basal insulin (units/day) | 6.9 (5.2, 8.6) | 7.2 (5.0, 9.6) | 0.2 (−0.2 to 0.9); 0.057 |
Bolus insulin (units/day) | 0.3 (0.1, 1.0) | 0.4 (0.1, 1.1) | 0.1 (−0.1 to 0.2); 0.42 |
Total insulin/kg | 0.11 (0.08, 0.13) | 0.11 (0.08, 0.13) | 0.00 (0.00, 0.01); 0.035 |
Results are reported as median (IQR) or mean (SD) unless otherwise indicated. CV, coefficient of variation.
Outcomes were calculated over the entire 3-week period.
All P values are superiority P values unless otherwise indicated.
Primary outcome.
95% CI.
Noninferiority P value with a 4% margin.
The higher time in range with CHO counting was observed during both the day (0600–0000 h) and night (0000–0600 h). During the day, the time in range was 72.9% (10.5%) with CHO counting and 69.8% (11.9%) with qualitative meal-size estimation (P = 0.052; Table 2). During the night, the time in range was 77.7% (11.9%) with CHO counting and 72.7% (13.0%) with qualitative meal-size estimation (P = 0.014; Table 2). The difference in mean glucose level during the night was highest at the start of the night at 0000 h (qualitative meal-size estimation vs, CHO counting, respectively: 9.3 vs. 8.8 mmol/L; P = 0.02) and diminished toward the morning at 0600 h (7.6 vs. 7.3 mmol/L, respectively; P = 0.17; Fig. 1).
Times spent below 3.9 mmol/L and below 3.0 mmol/L were low in both interventions (for CHO counting vs. qualitative meal-size estimation, respectively: 1.4% [0.6, 2.9] vs. 1.6% [0.6, 2.8] for time below 3.9 mmol/L and 0.2% [0.0, 0.7] vs. 0.1% [0.0, 0.5] for time below 3.0 mmol/L). Times in hypoglycemia were also low during the day and night in both interventions (Table 2).
Bolus insulin delivery was similar between interventions (mean, 21 units/day), whereas automated basal insulin delivery was higher in the qualitative meal-size estimation arm (29.6 [20.5, 41.8] vs. 26.8 [20.1, 41.2] units/day; P = 0.00078; Table 2). This increase in basal insulin delivery was also observed when the outcomes were calculated during the day period and the night period. Bolus insulin delivery was also similar between interventions during the day and night periods (Table 2). There was no difference between the amount of manual correction boluses (outside mealtimes) between the two interventions (for CHO counting vs. meal categorization, respectively: 0.9 [0.6, 2.8] vs. 1.3 [0.6, 2.4] units/day; P = 0.6). Manual correction boluses were also not different during the day and night periods (data not shown).
The mean daily number of meals announced to the system was 3.4 (1.2) meals/day during the CHO counting arm and 3.4 (1.0) meals/day during the qualitative meal-size estimation arm. The mean daily CHOs announced to the system during the CHO counting arm were 129.9 (62) g/day. In the qualitative meal-size estimation arm, 35.5%, 42%, 17.2%, and 5.3% of the meals were announced as having low, medium, high, and very high CHO, respectively. These proportions were the same when the analysis was restricted for breakfast, lunch, and dinner meals (data not shown). In both interventions, 96% of the meals were announced during the day hours.
Supplementary Fig. 3 shows the incremental postprandial glucose levels for breakfast (0600–1100 h), lunch (1110–1600 h), and dinner (1610–2200 h) in both arms, indicating that the difference between the two arms was only observed during breakfast. The mean CHOs announced to the system for breakfast, lunch, and dinner were 39.3 (18.8), 43.7 (19.2), and 44.7 (41) g, respectively.
Figure 2 shows the frequency of announcement of each meal category at the individual level. Throughout the 3-week interventions, 14 participants (47%) did not use the very high CHO meal category, and eight participants (27%) used it only once. Eight participants (27%) used either the low or medium CHO meal category >95% of the time.
There were no severe hypoglycemia, diabetic ketoacidosis, or other serious adverse events throughout the study in either intervention.
Conclusions
In this study, we compared CHO counting strategy with a simplified, qualitative meal-size estimation strategy on the background of closed-loop automated insulin delivery. Even though the qualitative meal-size estimation led to a high time in range and low time in hypoglycemia, it still degraded time in range by 3.6% compared with CHO counting, and noninferiority with a predefined margin of 4% was not confirmed (i.e., the CI of the difference crossed the noninferiority margin). Of note, half the participants had time in range that was better or noninferior (difference, <4%) with the qualitative meal-size estimation strategy, suggesting that this strategy might still be appropriate for a subset of patients.
This is our first outpatient free-living study with the qualitative meal-size estimation strategy, and further improvements to this strategy may be warranted. First, the prandial boluses with the qualitative meal-size estimation were based on an assumed CHO amount at the lower end of each category (e.g., 35 g for the medium category defined as 30–60 g). More aggressive prandial boluses (e.g., using an assumed CHO amount in the middle of each category) may improve the qualitative meal-size estimation strategy, but this needs to be confirmed in future studies. Second, daily automated basal insulin delivery was higher in the qualitative meal-size estimation arm, suggesting that that the algorithm was making efforts to reduce the difference between the two interventions. However, the algorithm’s efforts might have been hindered by internal safety rules that limit its action in case of accumulated “insulin on board,” and these rules may need to be relaxed with the qualitative meal-size estimation strategy.
In the qualitative meal-size estimation strategy, we predefined the four categories and fixed them for all participants. However, 22 participants (73%) used the very high CHO category only once or never. This suggests that allowing the four meal-size categories to be individualized would have permitted participants to define smaller CHO ranges for the categories, which consequently would have allowed more-precise prandial dosing. Whether this individualization leads to improved glucose control warrants further investigation.
Another finer adjustment to the qualitative meal-size estimation strategy would be to allow the meal categories to be defined differently for different times of the day (i.e., breakfast, lunch, and dinner). Moreover, meals’ CHO content is the dominant, but not the only, factor determining postprandial insulin needs, and future simplified meal-bolus estimation strategies could consider other macronutrients (e.g., fat, proteins) that also impact the timing of postprandial glycemic peaks and insulin requirements, even with closed-loop systems (26).
We expected nighttime outcomes to be similar between interventions, given that the two systems had identical basal dosing algorithms but different bolus calculators. However, time in range during the night was higher in the CHO-counting arm. This is unlikely to be related to the differences in the bolus calculator, given that nocturnal correction boluses were the same between interventions and only 4% of the meals were announced during the night. Instead, it is likely that better daytime glucose control extended to better nighttime control because the difference in mean glucose level during the night was highest at the start of the night and diminished toward the morning hours. Earlier studies applying automated insulin delivery at night only showed that better nighttime control extends to better daytime control (2,27). Our study shows that the opposite is also likely to be true.
Alternative solutions to manual CHO counting are being investigated. These include smartphone apps based on image processing of meal pictures and audio processing of meal descriptions. A few have been assessed clinically for their accuracy, including Foodvisor (28) and GoCarb (29) (image processing) and VoiceDiab (30) (audio processing), with promising results. Combining these apps with automated insulin delivery could be a venue for future research (31).
Strengths of this study include the randomized controlled design, the unsupervised settings, and the practical significance of the study hypothesis. However, our study has several limitations that affect its generalizability. First, the study duration of 3 weeks is relatively short. Longer duration might have allowed study participants to be better accustomed to the qualitative meal-size estimation strategy. Second, we excluded participants with high HbA1c and complications, and we did not assess the accuracy of CHO counting, because this was a real-life study. The initial training in CHO counting and the mean baseline HbA1c of 7.4% suggest that our study participants performed CHO counting with reasonable accuracy. The findings of our study would have been more generalizable had we restricted recruitments to participants not proficient with CHO counting, because this group is the intended target population for this strategy. Third, we used a research-based automated insulin delivery system. The use of a commercial system would have made it challenging to implement the qualitative meal-size estimation strategy.
Our study shows that the qualitative meal-size estimation strategy achieves solid glycemic outcomes, though it slightly degrades glucose control compared with CHO counting. Further improvements to this strategy, which might reduce diabetes management burden, are warranted.
Clinical trial reg. no. NCT04031599, clinicaltrials.gov
This article contains supplementary material online at https://doi.org/10.2337/figshare.22595938.
Article Information
Acknowledgments. The authors thank Virginie Messier and Joanna Rutkowski for providing help in the conduct and documentation of the study, design of the automated system, and the data analysis.
Funding. This study was supported by funding from the National Institutes of Health.
Duality of Interest. A.H. received research support from Eli Lilly, Dexcom, Adocia, Tandem, and AgaMatrix; consulting fees from Eli Lilly; and has pending patents in the field of automated insulin delivery. L.L. received consulting fees from Dexcom and received research support from Merck, AstraZeneca, and Sanofi. R.R.-L. received research grants from AstraZeneca, Eli Lilly, Merck, Novo Nordisk, and Sanofi; has been a consultant or member on advisory panels for Abbott, Amgen, AstraZeneca, Boehringer, Carlina Technology, Eli Lilly, Janssen, Medtronic, Merck, Neomed, Novo Nordisk, Roche, Sanofi, and Takeda; received honoraria for conferences from Abbott, AstraZeneca, Eli Lilly, Janssen, Medtronic, Merck, Novo Nordisk, and Sanofi; received in-kind contributions related to automated insulin delivery studies from Animas, Medtronic, and Roche; benefits from unrestricted grants for clinical and educational activities from Eli Lilly, Lifescan, Medtronic, Merck, Novo Nordisk, and Sanofi; and holds intellectual property in the field of type 2 diabetes risk biomarkers, infusion-set catheter life, and automated insulin delivery system. R.R.L., A.H., and L.L. have received purchase fees from Eli Lilly in relation to automated insulin delivery algorithms. No other potential conflicts of interest relevant to this article were reported.
Author Contributions. A.H., L.L., and R.R.-L. supervised the study. R.R.-L. held the funding. A.H., L.L., M.R., and R.R.-L. designed the study. L.L., M.R., N.G.-P., M.D., and R.R.-L. conducted the study. A.J. and M.G. carried out the data analysis including the statistical analyses. All authors read and approved the final version of the manuscript. A.H. and R.R.-L. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.