Using Continuous Glucose Monitoring Values for Bolus Size Calculation in Smart Multiple Daily Injection Systems: No Negative Impact on Post-bolus Glycemic Outcomes Found in Real-World Data

Abstract

Background:

Recent evidence shows that it may be safe to estimate bolus sizes based on continuous glucose monitoring (CGM) rather than blood glucose (BG) values using glycemic trend-adjusted bolus calculators. Users may already be doing this in the real world, though it is unclear whether this is safe or effective for calculators not employing trend adjustment.

Methods:

We assessed real-world data from a smart multiple daily injections (MDIs) device users with a CGM system, hypothesizing that four-hour post-bolus outcomes using CGM values are not inferior to those using BG values. Our data set included 184 users and spanned 18 months with 79 000 bolus observations. We tested differences using logistic regression predicting CGM or BG value usage based on outcomes and confirmed initial results using a mixed model regression accounting for within-subject correlations.

Results:

Comparing four-hour outcomes for bolus events using CGM and BG values revealed no differences using our initial approach (P > .183). This finding was confirmed by our mixed model regression approach in all cases (P > .199), except for times below range outcomes. Higher times below range were predictive of lower odds of CGM-based bolus calculations (OR = 0.987, P < .0001 and OR = 0.987, P = .0276, for time below 70 and 54 mg/dL, respectively).

Conclusions:

We found no differences in four-hour post-bolus glycemic outcomes when using CGM or BG except for time below range, which showed evidence of being lower for CGM. Though preliminary, our results confirm prior findings showing non-inferiority of using CGM values for bolus calculation compared with BG usage in the real world.

Introduction

People with diabetes treated with multiple daily injections (MDIs) must estimate their prandial insulin needs through a bolus size estimation before administering each injection. Calculators that use continuous glucose monitoring (CGM) technology with glycemic trend “arrows” (ie, the rate of change [ROC] in the glucose trace) that adjust bolus dosing have been shown to benefit diabetes management and improve glycemic control.^{1 -6} These CGM calculators reduce the burden of blood glucose (BG) testing, which involves performing a painful fingerstick that may also be inconvenient for persons with diabetes. ⁷ These challenges may lead some individuals to use CGM values in the insulin-dose calculators, as several CGM system apps provide real-time glucose information with trend arrows in addition to having predictive hypoglycemia and hyperglycemia alarms in addition to dosing recommendations.

Recent evidence shows that using CGM values with a bolus calculator capable of adjusting for glycemic trends is not correlated with inferior outcomes compared with one that accounts for BG values alone. ⁸ This was shown in persons using an insulin pump in manual mode based on outcomes within four hours of each bolus. However, the comparative effectiveness of a calculator that does not adjust for CGM trends paired with CGM values compared with the same calculator using BG values as input is unknown. Considering the relative convenience and utility (eg, trending glycemic information with alerts) of CGM, individuals may be tempted to use CGM values for bolus calculation under some circumstances. Though such uses of CGM-based bolus calculators with smart MDI devices are possible and even desirable from the user experience perspective, no evidence is available revealing the potential risks and effectiveness of this practice.

To address this gap, we employed real-world data from users of a smart MDI device and a CGM system. We hypothesized that (1) users used both CGM and BG values for bolus calculations, (2) the outcomes within four hours of a bolus for a bolus calculated using CGM values would not be worse than one calculated using BG values, and (3) similar results could be found for bolus events happening in the daytime (6 am-midnight) and nighttime (midnight-6 am). To test our hypotheses, we employed a combination of descriptive and inferential statistics to show the distribution of outcomes and test for outcome differences. Specifically, we present descriptive statistics on the use of a bolus calculator based on CGM or BG values for combined smart MDI device and Medtronic CGM system users. We also present four-hour post-bolus outcomes and test group differences between CGM and BG bolus events using multiple statistical approaches to confirm the validity of our findings.

Methods

We extracted data spanning over 18 months between January 1, 2021 and July 18, 2022 from a smart MDI device (InPen smart insulin pen; Medtronic, Northridge, California), which included real-world bolus calculation activities (ie, meal therapy mode use: carb counting, meal estimation, fixed dose or unspecified), bolus information (ie, bolus timing and size), BG meter, and CGM system (Guardian Connect; Medtronic) information. The smart MDI device database also contained voluntarily entered demographic information, including age, height, weight, body mass index (BMI), gender, diabetes type, and year of diabetes diagnosis. Study data inclusion criteria dictated the need to have users of both the smart MDI device and the CGM system. Given the narrow focus of our analysis, no exclusion criteria were posed, which helped to limit additional bias that could be introduced via analytical framing. Our initial analysis included data from 184 users that contained 79 000 bolus observations after data cleaning (ie, outlier detection, duplicate, discrepancy and quality checks).

We classified bolus events by matching insulin-dose calculator events happening between 20 minutes before and five minutes after each bolus event. Post-bolus calculator events occurring up to five minutes later were included due to known system limitations that may delay data recording to the database. We focused on bolus events that matched calculator events, which brought the number of observations to 27 951 (N = 127 users). Then, each bolus was paired with any CGM and/or BG glucose value if it was recorded between 15 minutes prior or five minutes after the bolus event, when available; value matches dictated whether an event was classified as a CGM, BG, or unmatched bolus event. The data set containing calculator-matched events only included 10 911 CGM bolus observations (N = 108 users), 6148 BG observations (N = 95 users), and 10 892 unmatched observations (N = 121 users).

To test our hypotheses, we used a combination of descriptive and inferential statistics. First, we used descriptive statistics to summarize the distribution of bolus events using CGM and BG to calculate bolus sizes or not. Second, we calculated glycemic outcome metrics ⁹ within four hours of the bolus event to compare users across groups. The four-hour timeframe was selected due to the uncertainty of mealtimes after a bolus in our observational database and aiming to cover at least two hours post-meal, and also based on prior work framing as a way of providing comparable published results. Specifically, we calculated time in range (TIR) (ie, percentage of time with CGM values within 70 and 180 mg/dL), time above range (TAR) (ie, percentage of time above 180, 250, or 300 mg/dL), and time below range (TBR) (ie, percentage of time below 70 or 54 mg/dL).

Our data set’s observational framing led to multiple bolus observations per user that could not be considered independent, given within-subject correlations. Also, users with sub-optimal glycemic control using MDI therapy may have administered insulin more frequently, ¹⁰ leading to more observation points and a potential bias. To counter this effect and ensure equal weight across users, statistics of outcome averages per patient were summarized and differences between CGM and BG bolus were compared similarly with prior work. ⁸ To assess the differences between these two groups, we employed generalized linear models-based logistic regression predicting whether a bolus was calculated using CGM or BG based on the outcome numbers. We employed this method due to its robustness in the face of missing values and potential sample size mismatches that allowed us to keep a uniform group difference testing methodology across all glycemic outcomes. To ensure the robustness of our findings and that averaging outcomes would not impact our conclusions on group differences, we re-ran group difference tests using the raw data containing all, unaveraged bolus outcomes. This regression analysis used all available data points while accounting for within-subject correlations by building mixed model regressions with random effects per subject. Finally, we explored the impact of a factor known to impact outcomes and assess result robustness by splitting our data set across bolus events happening during the day (6 am-midnight) and during the night (midnight-6 am). This parameter was chosen due to its known impact on glycemic control and usage in prior work. ⁸ For example, nighttime outcomes may differ from daytime outcomes where users have the possibility of checking and adjusting with additional bolus events. ⁸

Data extraction was performed using Google BigQuery (version 1.38.1; Google, Mountain View, California), data preprocessing, exploratory analysis, and descriptive statistics relied on Python (version 3.9.7) and Jupyter Notebooks (version 6.4.5). Inferential statistics were done R version 3.6.130 and RStudio (version 4.2.2) and RStudio (version 1.1.456; Posit, Boston, Massachusetts).

Results

The overall data set included 184 users with 79 000 bolus observations. Out of users with demographic data available (31.4% of data were available due to optional data entry) (Table 1), 26 (63.6%) were female and 15 (34.1%) male; 24 and 23 reported being diagnosed with type 1 and type 2 diabetes, respectively, and reported 20.7 (±15.1) years since diagnosis. Mean age, weight, height, and BMI were 47.9 (±17.4), 91.8 (±27.5), 169.0 (±10.6), and 32.21 (±9.07), respectively. All users reported their therapy mode showing 114 (62.0%) carb counting therapy, 25 (13.6%) fixed-dose therapy, 27 (14.7%) meal estimation therapy, and 18 (9.8%) users reported no specific therapy. Of the 79 000 bolus observations available, 27 951 relied on a bolus calculator operation in the system, whereas 51 049 did not seem to and were classified as non-calculator bolus (Table 2). Units administered ranged between three and 12 units with and mean of 8.8 (±7.7) units for non-calculator bolus events compared with a range of 2.5 to 10 units with a mean of 7.3 (±6.5) units for calculator bolus events. Rate of change values were relatively similar with means of 0.3 (±1.0) and 0.2 (±1.0) mg/dL/min for non-calculator and calculator modes, respectively. Users took an average for 2.76 BG readings per day.

Table 1.

User Demographics.

Variable	Missing	Overall
n		184
Age, mean (SD)	138	47.9 (17.4)
Weight in kg, mean (SD)	145	91.8 (27.5)
Height in cm, mean (SD)	142	169.0 (10.6)
Body mass index, mean (SD)	146	32.2 (9.07)
Gender, n (%)	140	—
Female		28 (63.6)
Male		15 (34.1)
Did not report		1 (2.3)
Diabetes type, n (%)	137	—
Type 1		24 (51.1)
Type 2		23 (48.9)
Years since diabetes diagnosis, mean (SD)	144	20.7 (15.1)
Therapy mode, n (%)	0	—
Carb counting		114 (62.0)
Fixed dose		25 (13.6)
Meal estimation		27 (14.7)
None		18 (9.8)

Table 2.

Calculator, Non-Calculator, Unmatched, BG, and CGM Bolus Event Parameters Descriptive Statistics.

Variable	Missing	Overall	Non-calculator bolus	Calculator	BG entry	CGM entry	Unmatched entry
N		79000	51049	27951	6148	10911	10892
Users		184	184	127	95	108	121
Insulin units, mean (SD)	0	8.3 (7.3)	8.8 (7.7)	7.3 (6.5)	7.0 (5.9)	7.8 (6.9)	7.0 (6.4)
Insulin units, median (Q1, Q3)	0	6.0 (3.0, 12.0)	6.0 (3.0, 12.5)	5.0 (2.5, 10.0)	5.0 (3.0, 9.0)	5.0 (2.5, 11.0)	4.5 (2.5, 9.5)
Glycemic rate of change (ROC), mean (SD)	8289	0.3 (1.0)	0.3 (1.0)	0.2 (1.0)	0.1 (0.9)	0.2 (0.9)	0.3 (1.0)
Glycemic rate of change (ROC), median (Q1, Q3)	8289	0.1 (−0.3, 0.7)	0.1 (−0.3, 0.8)	0.1 (−0.3, 0.6)	0.0 (−0.4, 0.6)	0.1 (−0.3, 0.6)	0.1 (−0.3, 0.7)
Insulin on board (IOB), mean (SD)	0	1.6 (6.1)	2.1 (7.4)	0.6 (1.9)	0.4 (1.4)	0.6 (1.6)	0.8 (2.3)
Insulin on board (IOB), median (Q1, Q3)	0	.0 (0.0, 0.5)	0.0 (0.0, 0.3)	0.0 (0.0, 0.8)	0.0 (0.0, 0.0)	0.0 (0.0, 0.4)	0.0 (0.0, 0.5)

Abbreviations: BG, blood glucose; CGM, continuous glucose monitoring; ROC, rate of change; IOB, insulin on board.

We found bolus occurrences using both CGM and BG values, with a larger prevalence of those using CGM values. Out of the calculator-classified boluses, the most prevalent were events using CGM values with 10 911 occurrences, followed by those using a value that could not be matched to CGM or BG values 10 892, and finally, those matched to a BG value with 6148 occurrences. Mean bolus sizes were relatively close returning 7.8 (±6.9), 7.0 (±5.9), and 7.0 (±6.4) units on average for CGM, BG, and unmatched bolus events. Rate of change values were also relatively close with means and standard deviations of 0.2 (±0.9), 0.1 (±0.9), and 0.3 (±1.0) mg/dL/min, respectively.

Comparing CGM and BG bolus events, we found that four-hour post-bolus times presented similar values for all outcomes. Specifically, there were no statistically significant differences between CGM and BG bolus outcomes in most cases (Table 3). Specifically, TIR was 52.0% and 52.7% for BG and CGM, respectively (P = .905); TBR returned 2.98% and 2.65%, respectively (P = .183) and TAR returned 45.0% and 44.6% (P = .933), respectively. Extreme low values (ie, time below 54 mg/dL) as 0.73% and 0.76% for BG and CGM bolus events, respectively (P = .208). Extreme highs showed average times of 16.5% and 15.6% for BG and CGM at a 250 mg/dL, respectively (P = .614) and 7.3% and 6.4% for BG and CGM at a 300 mg/dL (P = .804), respectively. Our sensitivity analysis testing differences using a mixed model analysis accounting for within-subject correlations revealed no differences for all outcomes except for times below range. Specifically, we found a difference between CGM and BG, showing lower odds of CGM-based bolus calculations for higher times below range (OR = 0.987, P < .0001 and OR = 0.987, P = .0276, for time below 70 and 54 mg/dL, respectively).

Table 3.

Glycemic Outcomes for Events Using Bolus Calculators Using CGM and BG Values.

Outcome	BG entry, n = 6148 Users = 95	CGM entry, n = 10 911 Users = 121	Difference	P value	P value (mixed model)	OR (95% CI) (if significant)
Time in range	52.0 (±22.7)	52.7 (±24.9)	0.7 27 (±2.16)	.905	.199	—
Time below range	2.98 (±6.114)	2.65 (±5.751)	−0.33 (±0.36)	.183	< .0001	0.987 (0.981, 0.992)
Time below 54	0.73 (±1.92)	0.76 (±2.44)	0.03 (±0.51)	.208	.0276	0.987 (0.976, 0.999)
Time above range	45.0 (±23.8)	44.6 (±26.0)	−0.39 (±2.16)	.933	.972	—
Time above 250	16.5 (±18.8)	15.6 (±16.9)	−0.9 (±2.0)	.614	.644	—
Time above 300	7.3 (±12.0)	6.4 (±9.8)	−0.89 (±2.2)	.804	.451	—

Abbreviations: BG, blood glucose; CGM, continuous glucose monitoring; OR, odds ratio; CI, confidence interval.

Outcomes by time of day showed similar outcomes for CGM and BG bolus events for both daytime and nighttime, returning no statistically significant differences (Table 4). Our sensitivity analysis confirmed the absence of significant differences for all outcomes except for time below 70 mg/dL. Our model predicted that an increase in this outcome would lower the odds of CGM-based bolus calculation during both daytime and nighttime (OR = 0.987, P < .0001 and OR = 0.981, P = .034, respectively).

Table 4.

Glycemic Outcomes for Events Using Bolus Calculators Using CGM and BG Values by Day (ie, 6 am-Midnight) or Night (ie, Midnight-6 am).

Bolus sub-group	Outcome	BG entry	CGM entry	Difference	P value	P value (mixed model)	OR (95% CI) (if significant)
Daytime N = 16 225	Time in range	52.2 (±23.0)	52.9 (±25.1)	0.7 (±2.1)	0.832	.164	—
Users = 115	Time below range	2.86 (±6.06)	2.30 (±5.06)	−0.55 (±0.96)	0.48	<.0001	0.987 (0.982, 0.993)
	Time below 54	0.68 (±1.93)	0.57 (±1.89)	−0.11 (±0.04)	0.699	0.092	—
	Time above range	44.9 (±24.1)	44.8 (±26.2)	−0.16 (±2.13)	0.963	.767	—
	Time above 250	16.5 (±18.9)	15.6 (±17.0)	−0.84 (±1.88)	0.74	.470	—
	Time above 300	7.37 (±12.2)	6.42 (±9.91)	−0.94 (±2.29)	0.549	.597	—
Nighttime N = 835	Time in range	50.6 (±27.3)	51.9 (±26.6)	1.3 (±0.73)	0.809	.361	—
Users = 70	Time below range	5.65 (±11.58)	3.47 (±7.77)	−2.18 (±3.80)	0.274	.034	0.981 (0.963, 0.999)
	Time below 54	1.57 (±3.61)	1.4 (±3.75)	−0.17 (±0.15)	0.827	.210	—
	Time above range	43.7 (±28.87)	44.6 (±26.8)	0.9 (±2.1)	0.870	.126	—
	Time above 250	17.1 (±18.8)	15.7 (±19.3)	−1.6 (±0.5)	0.356	.731	—
	Time above 300	5.77 (±8.1)	6.76 (±13.8)	0.99 (±5.7)	0.436	.739	—

Abbreviations: BG, blood glucose; CGM, continuous glucose monitoring; OR, odds ratio; CI, confidence interval.

Discussion

Our analysis showed that individuals can use both CGM and BG values to calculate their bolus sizes in the real world when using smart MDI and CGM systems. Comparing the outcomes in both cases, we found no difference in four-hour post-bolus glycemic outcomes for users using BG and CGM values to calculate their bolus size and our hypothesis testing did not reveal statistically significant differences in most cases. Only slightly improved outcomes for times below 70 and 54 mg/dL when using CGM were evidenced by our mixed model analysis. Splitting daytime and nighttime outcomes, further confirmed the lack of differences for most predictors except for time below 70 mg/dL, which was confirmed in both time segments, with higher odds of CGM-based calculations for lower times below range.

These initial findings point at the necessity for diabetes technology developers to account for usage patterns in the wild via real-world data analysis, and also that unintended technology use may not always be as risky or unsafe when happening consistently in the real world. Though there may be a general awareness and interest in the need for off-label use to reach underserved populations of individuals with diabetes (eg, young children or pregnant women),^11,12 much less knowledge is available on the possible intended use deviation that target population of diabetes-treating technologies. In this article, we show that users may have a functional incentive to deviate from intended use (ie, avoiding additional blood draws necessary to generate BG values and, instead, using CGM values for bolus calculations) and that such incentive is followed in populations who rely on both smart MDI and CGM devices. It is likely that other unintended use patterns may exists in other sub-populations and for other diabetes-treating technologies. Also, our findings did not show differences in four-hour post-bolus outcomes for events where the intended use approach (ie, using BG values for bolus calculation was used) compared with the alternate approach (ie, using/entering CGM values with a calculator that does not account for glucose trends), which may be considered less safe and effective. A possible explanation for the lack of differences in this system may be the ability of the device’s calculator to account for active insulin or insulin on board^13,14 to adjust dosage recommendations, leading to lower over-dosing likelihood and less TBR. Another interesting finding was the relatively low proportion of pre-bolus calculator usage, which may be explained by long-term MDI patients being used to estimating bolus sizes manually or the use of a calculator not integrated to the smart MDI system. This might be addressed by providing patients with a better understanding of the benefits provided by an integrated calculator that accounts for active insulin and records their entries for future review. Though our findings are observational and preliminary in nature, they point at the existence of safe alternate modes of technology usage that may reduce users’ diabetes management burden^15,16 and simplify their interactions with the technology.

Our findings confirmed and expanded prior results found in the literature. Specifically, our results echo recent findings showing the non-inferiority of using CGM values with a bolus calculator compared with a bolus calculator accounting for BG values alone. ⁸ Our results expand on these findings by providing preliminary real-world evidence that this non-inferiority may expand to the usage of CGM values on calculators that do not account for glycemic trends. Though there were differences in study designs, MDI and manual insulin pump bolus sizes and delivery timings are likely to be similar for similar therapeutic needs. Also, the former study assessed outcomes in temporal phases while our study classified each bolus according to calculator usage with CGM or BG values happening at different times. Given that the outcomes of interest covered only four hours after each bolus, it is likely that immediate bolus calculation effects were comparable despite these differences. Finally, the main similarity between the two studies was the intent to treat classification of users that allowed for relatively simple, unbiased comparison between groups and between studies. ¹⁷

Our analysis is exploratory in nature and relies on real-world data, for which it has three core limitations. First, our study is based on observational data, which presents limitations, such as missingness in demographics values, limited medical history (eg, patients could only report one diabetes diagnosis), and different users following different treatment regiments, adding to the heterogeneity of the sample. These limitations are generally expected in most real-world data analyses,^{18 -22} making our approach adequate for this type of exploratory analysis. Also, bolus events follow real-world patterns as chosen by users rather than those of randomized trials. However, our analysis is intended to follow an intention to treat analysis model as done in prior research, ⁸ which mitigates this limitation. Second, user selection from our real-world database was limited to users of smart MDI devices and Medtronic CGM systems, which had two effects. On one hand, it limited the number of eligible subjects with enough data and bolus events with calculator data to 184, which is rather limited for a real-world data analysis. On the other hand, this selection criterion might have introduced specific biases toward users treated with both systems due to non-random medical necessity or insurance considerations constraining users to use Medtronic CGM systems. However, this was the only inclusion criterion used to harvest our data, which was fundamental to its analytical framing, given that data from both MDI and CGM systems were needed. No additional filtering criteria were used for the analysis and all available data were included to mitigate further bias. Finally, our analysis was set up to emulate an intent to treat analysis to reduce bias further. ¹⁷ Though this might have reduced bias by simplifying the analytical process and enabled comparison with prior work, ⁸ it might have also masked nuances relating to user behaviors and bolus adjustments done by users in the field to improve their glycemic control. This issue will be expanded on and explored via sub-group analyses in future work.

Our findings show that there is no difference in four-hour post-bolus glycemic metrics outcomes, or potential for significantly increased time spent below range, when using CGM or BG values to calculate bolus sizes at the population level based on real-world data for users of smart MDI and CGM systems. Our results confirm prior findings ⁸ and expand them by providing preliminary real-world evidence that using CGM values on calculators that do not account for glycemic trends may not be inferior to using a traditional BG-based bolus calculator with BG values. Additional work is needed to confirm these results and expand the analysis to other user populations as well as users using other smart MDI and CGM systems.