Glucose interpretation meaning and action: enhancing type 1 diabetes decision-making with textual descriptions

Abstract

Background:

Findings from our previous study indicate that people with type 1 diabetes mellitus (T1DM) unknowingly misinterpret data displayed on glucose monitoring systems and make inaccurate treatment decisions, which increases the risk of hospitalisation.

Objectives:

This study aims to assess the effectiveness of incorporating textual descriptions in glucose monitoring systems compared to existing systems. The main goal is to minimise the effort required in glucose data interpretation, facilitating better self-management and ultimately improving haemoglobin A1C levels.

Methods:

A two-arm and mixed-methods evaluation was conducted. Participants were randomly allocated to the control arm (existing systems) or the experimental arm (newly developed systems incorporating textual descriptions). In the first part, a task-based usability assessment was conducted to compare performance between the two arms. The second part evaluated participant preferences, agreement with textual descriptions and perceptions of the new systems.

Results:

A total of 86 participants were recruited. The experimental arm achieved an 85.15% total correctness score, compared to 74.38% in the control arm (p < 0.001). The experimental arm particularly outperformed the control arm in the ambiguous tasks, such as compression low. However, despite a higher performance and greater agreement with the textual descriptions, the experimental group exhibited a less favourable perception compared to the control group.

Conclusion:

Incorporating textual description into glucose monitoring systems enhances treatment decision-making for people with T1DM. It suggests that we are on the right path to helping them better understand their glucose data and assist their self-management. Extensive research is required to focus more on the patient-centred approach in information presentation and prioritise it in parallel with other advancements in glucose monitoring technologies.

Plain language summary

Helping people with type 1 diabetes make better treatment decisions: improving glucose monitoring with text descriptions

People with Type 1 Diabetes Mellitus (T1DM) rely on glucose monitoring systems to manage their blood sugar levels. However, our previous research found that many individuals misinterpret the data these systems display, leading to incorrect treatment decisions and an increased risk of hospitalisation. This study explores whether adding simple text descriptions to glucose monitoring systems can help people better understand their glucose data and make more accurate treatment choices. We conducted a study with 86 participants, randomly assigning them to use either an existing glucose monitoring system (control group) or a new system that includes textual descriptions (experimental group). Participants completed tasks to assess their understanding and decision-making, and we also gathered their opinions on the new system. The results showed that participants using the system with text descriptions performed significantly better, with an accuracy score of 85.15%, compared to 74.38% in the control group. The new system was especially helpful in tasks where the data was more difficult to interpret. However, despite the improved accuracy, users of the new system had a less favourable perception of it compared to the control group. These findings suggest that adding textual descriptions to glucose monitoring systems can help people with T1DM make better treatment decisions. While user acceptance remains a challenge, this study shows that we are moving in the right direction. Future research should focus on making these systems both effective and user-friendly to better support individuals in managing their diabetes and improving their overall health.

Introduction

Background

People with type 1 diabetes mellitus (T1DM) must consistently monitor their blood glucose levels to make informed decisions towards achieving and maintaining the in-target glucose range (3.9–10.0 mmol/L).^1,2 Continuous glucose monitor (CGM) and flash glucose monitor (Flash) enable people with T1DM to monitor their glucose levels in real time. CGM automatically provides glucose readings to a receiver or mobile application every 5 min, while Flash requires users to manually scan the sensor with a receiver to view the readings.^3,4

These individuals must make safety-critical decisions based on their glucose levels approximately 180 times, in addition to their daily life decisions, ⁵ and misinterpretation could significantly increase their risk of making erroneous treatment decisions, potentially resulting in hospitalisation. ⁶

Studies have shown that people with T1DM face challenges in understanding and making appropriate treatment decisions, based on their glucose monitoring system, as it is overwhelming and difficult to review.^7,8 A recent study explored the perceived effectiveness of CGM features presenting glucose data among people with T1DM. ⁹ According to their survey findings, participants found CGM features, such as glucose number readings, graphs, arrows and hypoglycaemia alarms, to be highly beneficial. Conversely, results from their interviews suggested that participants still struggle to use these features, which hinders the optimal use of CGM. However, their results were based on self-reported preferences rather than an evaluation of their actual performance in using CGM. Our previous study ¹⁰ (in 2023) revealed that people with T1DM frequently misinterpreted glucose monitoring systems and made inappropriate treatment decisions despite using familiar devices. Nonetheless, participants expressed high confidence in their self-management skills and strongly believed that they were adhering to clinical guidelines. The average interpretation accuracies for both CGM and Flash users were less than 40% and less than 22% in treatment decision-making. Nonetheless, participants expressed high confidence in their self-management skills and strongly believed that they were adhering to clinical guidelines. These findings underscore the need for solutions to reduce their risk of misinterpretation and improve more accurate treatment decision-making.

Related work

Patient-reported needs for text

Several studies have highlighted the importance of adding textual descriptions in supporting effective glucose reading in people with T1DM.

Two studies^11,12 emphasise the importance of integrating textual descriptions with graphical content to improve the effectiveness of diabetes data visualisation. It was reported that people with T1DM expressed the need for complementary textual descriptions. The text, alongside graphical content, combined with actionable suggestions for behavioural interventions, could reduce cognitive effort and facilitate quicker interpretation of glucose data. ¹¹ It was also found that participants responded favourably to the inclusion of simple text in visualisations, as it provided context and delivered a quick, easy-to-understand summary. ¹²

Furthermore, a study revealed that the data CGM provides are difficult to interpret. This was due to the unactionable details and amount of data, which led to uncertainty in performing self-management and discontinuing the use of the device. ⁷

Together, these findings suggest the need to combine graphical and textual descriptions to present complex glucose data in a clear, concise and actionable manner.

Enhancing decisions with text

Textual descriptions in electronic medical tools have been found to aid decision-making more effectively than graph-only presentations. For instance, one study found that adding text improved the accuracy of decision-making among medical staff. ¹³ Another study found that while participants preferred graphical data, their comprehension significantly improved with accompanying textual information. ¹⁴

In non-medical contexts, integrating text with visuals has been shown to enhance decision-making by 44% compared to visuals alone. ¹⁵ Recent research suggests that textual descriptions should be considered a primary alternative to visualisations. ¹⁶ Furthermore, textual elements highlighting key aspects of visualisations have been found to improve participants’ ability to recall, as they often draw more attention than the visuals. ¹⁷

Building on these previous findings, we developed new glucose monitoring system interfaces of Dexcom G6 and FreeStyle Libre 2, incorporating textual descriptions of glucose data and suggested treatment actions. We evaluated the new system interfaces compared to the existing systems to determine whether this integration could reduce the effort required for data interpretation in people with T1DM. The ultimate goal is to enhance their understanding of glucose data and make safer treatment decisions.

Methods

Participant recruitment

A total of 86 participants were recruited by posting an advertisement in online diabetes communities and various advertising platforms provided by the JDRF, the Type 1 Diabetes Research Charity, UK. Inclusion criteria were as follows: aged over 18 years, being diagnosed with T1DM or being a carer of a person diagnosed with T1DM, currently using either a Dexcom or a FreeStyle Libre device, and being able to read and write in English.

Participants were required to complete a written consent form before participating in the study. No Participant Identifiable Information data was collected during this study.

All participants received compensation in the form of a £10 gift card. This study was conducted in 2024 and received an official ethical exemption approval from the Research Governance, Ethics and Integrity Department and the University Research Ethics Committees at the University of Manchester (Reference number: 2024-20659-35490).

Power

To determine the sample size, Power Analysis ¹⁸ on G*Power statistical software (version 3.1.9.6)^19,20 was used. The statistical power was set to default at 0.8, effect size (d) = 0.8 and threshold for significance (α) = 0.05. Therefore, 52 participants were needed, splitting into 2 arms, with 26 participants per arm.

The sample size also aligns with the recommendations of previous work^21,22 that a sample size of 25 participants is generally considered sufficient to yield statistically significant results in comparative usability testing.

Materials

System interfaces development

This study focused on the two most commonly used CGM and Flash devices: the Dexcom G6 (Dexcom, Inc., San Diego, CA) and the FreeStyle Libre 2 (Abbott Diabetes Care Inc., Alameda, CA), which are recommended by the National Health Service (NHS) in the United Kingdom.^23,24 The system interfaces were developed to mimic these two devices using a photo editor tool.

In terms of the interface design, we adhered to established guidelines for designing user interfaces in computerised clinical decision support systems,^25,26 focusing on four key principles: (1) keeping only essential elements, (2) improving readability with suitable font sizes, (3) meaningful colours and strong contrast, consolidating active information in one place and (4) making passive information visible at a glance without scrolling.

Based on these principles, the current glucose level was placed at the top in a large, colour-coded font consistent with the device’s system. Key details, such as the glucose category, sensor error warnings and alert message, were highlighted in bold text. All information appeared on a single screen, eliminating the need to scroll. The content was carefully ordered, showing the current glucose reading first, then the predicted trend, followed by historical data, helping users track changes easily.

We first ran a pilot study with five participants to evaluate the initial set of system interfaces. After analysing the results, we revised them to be more readable by increasing the text size and enhancing the image resolution. Figure 1 illustrates the finalised four system interfaces. The details of each group are outlined below:

Figure 1.

Four glucose monitoring system interfaces: Baseline, Meaning, Actions and Meaning-Actions.

1. Baseline: This system replicated existing CGM and Flash interfaces, appearing exactly as users see them in daily life.

2. Meaning: In addition to the Baseline system, this system added textual descriptions interpreting glucose levels.

3. Action: In addition to the Baseline system, this system included suggested treatment actions for the displayed glucose levels.

4. Meaning-Action: In addition to the Baseline system, this system included both textual interpretations of glucose levels and corresponding treatment suggestions.

Glucose level scenarios

The representative glucose level scenarios displayed in each glucose monitoring system were modified based on our previous study. The glucose monitoring systems were organised into categories, resulting in seven categories, as shown in Table 1.

Table 1.

All glucose scenarios grouped by categories.

Category	Task name	Description
1. Level 1 hyperglycaemia and hypoglycaemia	‘Level 1 Hypoglycaemia’	An interface representing 0.1 mmol/L below in-target glucose range.
	‘Level 1 Hyperglycaemia’	An interface representing 0.1 mmol/L above in-target glucose range.
2. Excessive glucose levels	‘Very High’	An interface for excessively high level: the glucose level was 14.3 mmol/L.
	‘Very Low’	An interface for excessively low level: the glucose level was 3.3 mmol/L.
3. HI/HIGH and LO/LOW readings	‘HIGH’	An interface for dangerously high glucose level: the reading displayed ‘HIGH’ in CGM, meaning that the level is >22.2 mmol/L, and ‘HI’ in Flash, meaning that the level is >27.8 mmol/L.
	‘LOW’	An interface for dangerously low glucose level: the reading display ‘LOW’ in CGM and ‘LO’ in Flash, both represented the glucose level <2.2 mmol/L.
4. In-target range levels	‘In-range 1’	Two interfaces representing in-target range glucose levels.
	‘In-range 2’
5. Low glucose level alerts	‘Alert Banner 1’	Two interfaces: the alert appears when the sensor detects that the glucose level is falling quickly to ⩽3.1 mmol/L in up to 20 min or ⩽3.9 mmol/L within 15 min in advance.
	‘Alert Banner 2’
6. Low glucose level alarms (Dexcom G6 only)	‘Alarm 1’	Two interfaces for low glucose level alarms. It appears when the level is at or below 3.9 mmol/L. This feature is only available in Dexcom G6.
	‘Alarm 2’
7. Sensor error	‘Compression Low 1’	Two interfaces representing the compression of low glucose levels. This scenario occurs when there is pressure on the sensor, making the displayed glucose level drop very sharply within less than 5 min intervals, deviating the reading from the actual glucose level, which is physiologically implausible for blood glucose.27–30
	‘Compression Low 2’

See the full list of system interfaces in Appendix B.

CGM, continuous glucose monitor.

Due to limited space, see the full list of glucose monitoring systems in Appendix B (Table B1).

CGM and flash systems differences

Note that the systems of Dexcom G6 and FreeStyle Libre 2 devices were intended to present identical glucose level scenarios. However, due to differences in their functionalities, some discrepancies were unavoidable:

• Timeline: Dexcom G6’s system displays the timeline in a 1 h window, while FreeStyle Libre 2 uses a 3 h window, resulting in different graph appearances.

• Alert Setting: In scenarios where the glucose monitoring devices detect a rapid change in glucose levels, the Dexcom G6 displayed an alert message at the top of the screen, such as ‘3.1 mmol/L within 20 minutes’. In the FreeStyle Libre 2, a message ‘Glucose going low’ indicating a likely decrease to 3.9 mmol/L within 15 min^31,32 is shown. This difference results in appearance and perceived urgency variations between the two devices, with the Dexcom G6 alert conveying a greater urgency.

• HI/HIGH and LO/LOW Readings: The Dexcom G6 displays ‘HIGH’ and ‘LOW’ when glucose level exceeds 22.2 mmol/L and falls below 2.2 mmol/L. By contrast, the FreeStyle Libre 2 uses ‘HI’ and ‘LO’ for glucose readings above 27.8 mmol/L and below 2.2 mmol/L, respectively.^31,32

• Rate of Change: Both Dexcom G6 and FreeStyle Libre 2 use arrows to show glucose trends, but their meanings differ. For example, a 0.1–0.2 mmol/L per minute rise is shown as two upward arrows on the G6, but only one arrow on the Libre 2.^31,32

Textual description generation process

We created three pre-defined templates based on three glucose scenarios. The first was a Normal scenario, representing standard situations without warnings or errors. The second was an Alert scenario, illustrating cases where the system displayed a warning or alert message. The third was a sensor error scenario, depicting situations where the device showed erroneous readings due to the pressure on the sensor.

Each canned template contained variables that would later be replaced by actual glucose data. At this stage, the templates were populated manually. However, in future studies, an algorithm could be developed to automate this process. Figure 2 illustrates the text generation process, and the details are as follows:

Figure 2.

The process of glucose meaning generation. Three pre-defined templates were developed and populated by hypothetical glucose data.

• Category: Represented different glucose ranges. For example, ‘very high’ represented the glucose level of 13.9–22.2 mmol/L and ‘low’ for 3.4–3.9 mmol/L.

• Glucose Number: Showed the current glucose level. For example, ‘3.8 mmol/L’.

• Trend: Presented the arrow shown in the glucose monitoring systems. The values were derived based on the rate of change specified in the device’s user manual. For example, ‘stable’, ‘falling quickly’ and ‘rising slowly’.

• Graph: Presented the trend of the graph during the latest 1 h. For example, ‘. . . and it has been stable from 1 hour ago’.

• Predicted Level and Time: These variables were used in the scenario where there was an alert message appearing at the top of the screen. The variable represented the displayed message. For example, ‘. . .is predicted to fall to 3.1 mmol/L in 20 minutes’.

Since our primary goal was to create textual descriptions that require minimal interpretation effort, we generated the text to correspond to a fifth- to sixth-grade reading level,^33,34 indicating an ‘easy’ to ‘very easy’ level of readability, as measured by the Flesch-Kincaid Grade Level metric. ³⁵

Study setting

Figure 3 illustrates the study setting, comprising two main parts: quantitative and qualitative studies.

Figure 3.

Study setting: the quantitative part asked participants’ respective actions in each scenario. Second, the qualitative part gathered their opinion toward the new system interface.

Quantitative study

The study was conducted using a survey on the Qualtrics platform. ³⁶ Participants first provided written consent and demographic information, including age, glucose monitoring device, insulin regimen, role (patient or carer), experience with any T1DM structured education and A1C level.

Participants were randomly assigned to one of two arms: the control arm, presented with the Baseline system interface, and the experimental arm, the Meaning system interface. We did not present the experimental arm with the Action and Meaning-Action systems because treatment actions vary significantly based on individuals’ conditions. Participants were shown a system interface representative of their current device (either Dexcom or FreeStyle Libre) to minimise bias related to unfamiliarity with the device. ³⁷ Randomisation was performed through the Qualtrics workflow.

After viewing the assigned system interface, participants were prompted with the open-ended question, ‘What would you do if this screen were yours?’ They were then allowed to provide written responses of unrestricted length.

Qualitative study

This part comprised three sub-parts designed to explore participants’ preferences, level of agreement, perceived usefulness and ease of use regarding the newly developed Meaning-Action system. This system represents the sensor error scenario (Appendix A Figure A1). These sub-parts explore participants’ opinions towards the newly developed systems.

1. Preference Ranking: Participants were asked to rate all four systems from most to least preferred.

2. Agreement Level on the Generated Textual Description: Participants were shown the Meaning-Action system of a sensor error scenario. They were asked to rate their agreement with the text presented. The response was recorded using the 7-point Likert Scale ³⁸ (1 = Strongly disagree, 7 = Strongly agree).

3. Perceived Usefulness and Perceived Ease of Use: Evaluated the extent to which the participants perceived the Meaning-Action system to be useful. A well-known validated measure, Perceived Usefulness and Perceived Ease of Use (PUPEOU) ³⁹ was applied. The response was recorded using the 7-point Likert Scale ³⁸ (1 = Strongly disagree, 7 = Strongly agree).

Data analysis

Response analysis

The survey responses were assessed by (1) evaluating the degree of congruency to standard clinical guidelines, which were primarily based on the National Institute for Health and Care Excellence (NICE), ⁴⁰ Dose Adjustment For Normal Eating (DAFNE) course book, ⁴¹ American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD)^42,43 and (2) participants’ demographical information as their treatment decisions may vary based on their health history and how their glucose levels respond in specific contexts. The co-authors and a clinical diabetologist at the Manchester Royal Infirmary (the third author) were also assisted in the evaluation process. The level of correctness of responses was categorised into three types^44,45: ‘All correct’, ‘Partially correct’ or ‘Incorrect’ based on standard clinical guidelines, taking into account whether the response contained minor mistakes or major mistakes that could lead to medical complications.

The average total correctness was analysed using the weighted average, where the weight reflects each correctness level’s importance. The ‘All correct’, ‘Partially correct’ and ‘Incorrect’ were given the weight of 2, 1 and 0, respectively. ⁴⁶

Statistical analysis

The statistical analysis was conducted using IBM SPSS Statistics software (Version 29.0.0.0). ⁴⁷ In terms of the normality test, the Kolmogorov–Smirnov test was used.^48,49 Statistical differences in level of correctness between control and experimental arms were derived by using the Wilcoxon rank sum test (Mann–Whitney U test),^50,51 due to the non-normally distributed data. An alpha level of 0.05 was used for all statistical tests.

Descriptive statistical analysis, such as averages and frequency distributions, was applied to summarise the responses to the preference ranking, the agreement levels on the generated textual description and the PUPEOU. In addition, Cronbach’s alpha was used to measure the internal consistency of our PUPEOU survey questions. ⁵²

Results

A total of 86 participants responded to the survey. Table 2 shows the demographic of participants. Initially, we aimed for 52 participants in total, as discussed earlier. However, the responses arrived rapidly, resulting in a higher-than-anticipated volume.

Table 2.

Participant demographics in this study.

Demographic	Control arm (n = 44)	Experimental arm (n = 42)
Device group
CGM users	16	14
Flash users	28	28
Insulin regimen
Insulin pen/pump	42	41
Others (hybrid closed loop)	2	1
Role
Patient	42	40
Carer	2	2
Age group
18–30	8	6
31–60	31	30
>60	5	6
T1DM structured education experience
Yes	31	27
No	13	13
Not sure	0	2
A1C (%)
Average	6.9 ± 0.9	6.8 ± 0.8

CGM, continuous glucose monitor; T1DM, type 1 diabetes mellitus.

Quantitative study results

Performance comparison

Overall, the experimental arm outperformed the control arm across all tasks, achieving a total correctness of 74.38%, compared to 85.15% in the control arm. Furthermore, 8.08% of participants in the experimental arm provided incorrect responses. This was lower than the 18.21% observed among participants in the control arm, as illustrated in Figure 4.

Figure 4.

Comparison of different levels of correctness between the control and experimental arms.

Table 3 presents the total correctness scores for the control and experimental arms, categorised by the device groups. Among the CGM group, the experimental arm achieved a score of 80.10% (n = 14), compared to 70.76% (n = 16) in the control arm. Similar to the Flash group, the experimental arm obtained 88.10% (n = 28), whereas the control arm achieved 76.79% (n = 28).

Table 3.

Total correctness score between the control and experimental arms, separated by CGM and Flash groups.

Device group/Arm	Control arm (n = 44)	Experimental arm (n = 42)
CGM group (n = 30)	70.76%	80.10%
Flash group (n = 56)	76.79%	88.10%

CGM, continuous glucose monitor.

Task-based performance

Figures 5 and 6 illustrate the correctness levels for each task in the control and experimental arms, respectively. Two sensor error scenarios (‘Compression Low 1’ and ‘Compression Low 2’ tasks) gained the highest number of participants with incorrect responses in both arms. These tasks corresponded to erroneous readings that resulted in a faulty hypoglycaemic episode.

Figure 5.

Level of correctness presented by each task for the control arm (n = 44). The bar graphs are ordered by the number of incorrect responses.

Figure 6.

Level of correctness presented by each task for the experimental arm (n = 42). The bar graphs are ordered by the number of incorrect responses.

Upon closer observation, Figure 7 highlights that the control arm demonstrated more incorrect responses than the experimental arm in both the ‘Compression Low 1’ and ‘Compression Low 2’ tasks. It was found that participants in the control arm were unaware that these errors were due to sensor error. Consequently, they responded by consuming carbohydrates, which, in reality, could lead to a hyperglycaemic episode:

Figure 7.

Number of participants categorised by levels of correctness and compared between two arms. (Left) presents ‘Compression Low 1’ task. (Right) presents ‘Compression Low 2’ task.

I would be rushing to get fast acting carbs, either jelly babies or small coke tins. – P4 (Flash, control arm)

By contrast, participants in the experimental arm recognised that the reading was faulty and, as a result, opted to wait and perform a finger prick test:

Wait a few mins to see if it resolves, if not, finger prick test BG, act on results. – P26 (CGM, experimental arm)

The ‘Alert Banner 1’ and ‘Alert Banner 2’ tasks gained the second-highest number of incorrect responses. The former task represented a glucose level of 6.6 mmol/L with an alert message indicating ‘3.1 mmol/L within 20 minutes’ (or ‘Glucose going low’ in the Flash system), suggesting a sudden drop in glucose level. If ignored, this could potentially lead to a severe hypoglycaemic episode.

We found that, in the control arm, 25.00% of participants provided incorrect responses, while 16.67% of participants in the experimental arm did so, as shown in Figure 8. It was observed that participants in the control arm chose not to react to the situation and instead opted to wait and see.

Figure 8.

Number of participants categorised by levels of correctness and compared between two arms. (Left) presents ‘Alert Banner 1’ task. (Right) presents ‘Alert Banner 2’ task.

Additionally, one participant misunderstood the issue, presuming it was an error caused by the application:

Wait for the next reading and then decide. – P77 (CGM, control arm)

. . . 6.6 is normal and shouldn’t be coming up as red. . . Finger pricking to double check - seems like an error on app. – P32 (CGM, control arm)

In the experimental arm, participants appeared to be more proactive in addressing the issue. They took action by preparing to consume carbohydrates and closely monitoring their glucose levels.

However, in the ‘Alert Banner 2’ task, participants demonstrated relatively higher correctness compared to the ‘Alert Banner 1’ task. The number of incorrectly responded participants was 9.09% and 7.14% in the control and experimental arms, respectively. This task presented a similar scenario as in the ‘Alert Banner 1’ but with a glucose value of 4.4 mmol/L.

Specifically, in both of the ‘Alert Banner 1’ and ‘Alert Banner 2’ tasks of the Flash group, the interface only displayed the message ‘Glucose going low’. The results indicated that the experimental arm still had fewer incorrect responses and a higher rate of correctness.

The task with the third-highest rate of incorrect responses was the ‘In-range 1’ in both arms. In the experimental arm, 16.67% of participants responded incorrectly, while 9.09% did so in the control arm. Many incorrect responses indicated that participants considered the glucose level of 8.7 mmol/L (displayed in the ‘In-range 1’ task) to be too high for them, leading them to take a correction dose to decrease their blood glucose level:

I target blood sugar below 8,0 . . . I might inject 1 unit or 2 into my shoulder and feed that back into closed loop. – P29 (CGM, control arm)

Inject 8.7 is too high for me. – P32 (Flash, experimental arm)

For the ‘In-range 2’ task, which showed a glucose level of 7.2 mmol/L, we found that fewer participants responded incorrectly compared to the ‘In-range 1’ task. Despite this, the control arm still gained more incorrect responses (9.09%) compared to the experimental arm (2.38%). Moreover, similar to the previous task, we observed incorrect responses where participants indicated that the glucose level was too high for them and chose to take a correction dose:

Since I like my levels to be closer to 6, I would take half a unit to bring it there. – P44 (Flash, control arm)

Qualitative study results

Preference ranking

Figure 9 presents the number of participants in different ranking combinations among four systems.

Figure 9.

Number of participants choosing different ranking combinations.

The results suggested that the majority of participants ranked the Baseline, Meaning, Action and Meaning-Action as their most-liked to least-liked (n = 25/86). Following this, nine participants ranked the Meaning-Action, Action, Meaning and Baseline as their most-liked to least-liked (n = 9/86).

Overall, the number of participants who ranked each system as their most preferred was follows: Baseline had 38 participants (n = 38/86), followed by Meaning (n = 18/86), Action (n = 15/86) and Meaning-Action (n = 15/86).

Agreement level on the textual description

Table 4 shows the agreement level towards the Meaning-Action system. The majority of participants in the control arm ‘Agree’ with the textual description, at 27.27% (n = 12), while most of the participants in the experimental arm ‘Somewhat agree’ at 40.48% (n = 17). The average agreement level of the control arm was 4.80 ± 0.60 (n = 44), while the experimental arm had 5.05 ± 0.76 (n = 42).

Table 4.

Results from the agreement level on the generated textual description displayed in the Meaning-Action system.

Agreement level	Arm
	Both arms (n = 86)	Control (n = 44)	Experimental (n = 42)
Strongly agree	17.44	20.45%	14.29%
Somewhat agree	29.07%	18.18%	40.48%
Agree	26.74%	27.27%	26.19%
Neither agree nor disagree	5.81%	9.09%	2.38%
Somewhat disagree	5.81%	4.55%	7.14%
Disagree	8.14%	11.36%	4.76%
Strongly disagree	6.98%	9.09%	4.76%

Participants who reported to ‘Strongly agree’ or ‘Agree’ provided feedback indicating that the textual descriptions offered accurate suggested actions and were helpful:

It does look like a false reading and those are the actions I would take! – P3 (Flash, experimental arm)

Those who leaned towards disagreeing with the textual description were not convinced that the glucose reading was faulty. Also, some commented that the interface was text-heavy:

Readings don’t tend to be false, when I look at my sensor to how I feel or to what my finger prick might say it is very similar. – P17 (CGM, control arm)

Perceived usefulness and perceived ease of use

Table 5 shows average PUPEOU scores of the Meaning-Action system (see Appendix C Table C1 for full results). Since the scale ranged from 1 to 7, the averages in both arms suggest a more positive opinion from participants regarding the Meaning-Action system. Notably, the control arm provided higher average scores across all statements compared to the experimental arm. Cronbach’s alpha was applied to determine the internal consistency of the survey questions. The results were 0.946, suggesting excellent reliability. ⁵³

Table 5.

PUPEOU survey results from both arms and each arm.

PUPEOU survey questions	Mean ± SD
	Both arms	Control arm	Experimental arm
PUPEOU 1. Using the screen would keep my blood glucose in-range more quickly.	4.43 ± 1.57	4.71 ± 1.72	4.36 ± 1.68
PUPEOU 2. Using the screen would improve my diabetes self-management.	4.44 ± 1.68	4.79 ± 1.81	4.31 ± 1.93
PUPEOU 3. Using the screen would make it easier for my diabetes self-management.	4.49 ± 1.83	4.79 ± 1.92	4.40 ± 2.01
PUPEOU 4. I would find the screen useful in my diabetes self-management.	4.47 ± 1.78	4.76 ± 2.00	4.38 ± 1.90
PUPEOU 5. I would find this screen easy to use.	4.91 ± 1.71	5.31 ± 1.64	4.74 ± 1.88

PUPEOU, Perceived Usefulness and Perceived Ease of Use.

Result analysis

We sought to determine whether there was a statistically significant difference in performance between the control and experimental arms. To achieve this, the Kolmogorov–Smirnov test was first applied to evaluate the normality distribution of the data. The results showed that the data were not normally distributed (D = 0.158, p = 0.001). Thus, the Wilcoxon rank sum test was employed to identify statistical significance.

The Wilcoxon rank sum test yielded a test statistic of W = 1441.500, Z = −4.151 and a p-value of p < 0.001. Since the p-value was less than 0.05, we rejected the null hypothesis and concluded that the experimental arm outperformed the control arm with statistical significance.

Notably, in ambiguous scenarios, such as the sensor error and low glucose level alert tasks, the addition of textual description significantly helped enhance participants’ decision-making. However, the results from both arms differed slightly in the remaining tasks. For example, all participants responded correctly to the ‘Alarm 1’ and ‘Alarm 2’ tasks. These tasks represented glucose levels, prompting immediate glucose intake. Next, in the ‘HIGH’ and ‘LOW’ tasks, the results of both arms did not differ significantly as they responded to these severe glucose levels correctly, with only a few incorrect responses. The pattern observed in this finding indicated that textual descriptions were particularly beneficial in ambiguous scenarios while providing slight improvement in extreme scenarios.

In addition, the total correctness scores between CGM and Flash groups were not statistically significant (p = 0.083). This finding suggests that there were no differences in responses between CGM and Flash users, indicating that the type of device did not influence the correctness of the responses.

Focusing on participants who reported having received T1DM structured education, it was found that those in the experimental arm attained a higher total correctness score of 85.38%, compared to 74.12% in the control arm.

For individuals without experience in attending the T1DM structured education, those in the experimental arm still achieved a higher total correctness score of 85.98%, compared to 75.00% in the control arm. Additionally, this difference was found to be statistically significant (p = 0.003).

In terms of preference, the results were mixed as the Baseline system was ranked as the first choice by 38 participants (44.19%), and a number of 48 participants (55.81%) selected one of the newly developed systems as their most-liked option. This indicates that, overall, the participants might also consider the systems incorporated with textual description to be useful.

We found interesting discrepancies between performance, agreement level and PUPEOU results among the two arms. The experimental arm, which gained better performance scores, appeared to agree more with the generated text in the Meaning-Action system than the control arm. However, their PUPEOU results showed that they scored lower than the control arm. By contrast, the control arm, which underperformed, appeared to agree less with the text but offered higher PUPEOU scores.

Discussion

In this study, we demonstrated that incorporating the textual description of glucose levels enhances decision-making accuracy in treatment actions for people with T1DM.

Overall, participants expressed positive opinions towards the new glucose monitoring systems.

It was observed that participants in the experimental arm made more accurate treatment decisions compared to those in the control arm. This result supports our conjecture that textual descriptions require less interpretation, which helps reduce cognitive effort and improve the decision-making process. A similar outcome was reported in a study where participants who received a textual representation of the A1C test result demonstrated a better understanding of the information, which subsequently led to positive changes in diabetes self-management. ⁵⁴

Our results showed that the experimental arm outperformed the control arm in ambiguous tasks, the sensor error and the low glucose alert, though not across all tasks. In the sensor error tasks, the control arm failed to recognise the faulty hypoglycaemic reading and was not convinced that the sensor could display a false reading, despite this being identified as one of the most common CGM errors in prior research. ³⁰ This is particularly concerning, as failure to identify such errors could lead participants to consume excessive carbohydrates to raise their blood glucose levels, which may result in severe hyperglycaemia episodes. Hence, the addition of textual description helped the experimental arm to be aware of the error and not take unnecessary treatment.

Next, the low glucose alert tasks. The experimental arm scored higher than the control arm, especially in the ‘Alert Banner 1’ task. One potential explanation is that incorporating textual descriptions emphasises the significance of glucose readings rather than relying solely on the alert message. This may also explain why both arms tended to perform better in the ‘Alert Banner 2’ task, as the level of urgency in this task was higher than in ‘Alert Banner 1’. This trend was particularly noticeable among participants in the Flash group, where the alert messages displayed, ‘Glucose going low’, unlike in CGM devices. In fact, our previous study confirmed this finding, demonstrating that alert messages were often overlooked, especially by Flash users. ¹⁰

Prior studies revealed that people with T1DM frequently find the device’s alerts and alarms to be redundant and annoying, leading to alarm fatigue and disengagement from paying attention to them.^9,55,56 The experimental arm’s superior performance suggests that adding textual descriptions enhances participants’ attention toward the alert features, making them more likely to engage with the message and prompting them to take treatment actions. This elevated attention to alerts and alarms can, in turn, improve hypoglycaemia control, as demonstrated in several studies.^57–59

An intriguing finding is that the ‘In-range’ tasks gained a higher rate of incorrect responses despite being straightforward glucose scenarios that required no action. Consistent with the findings from one study, ⁶⁰ our study found that many participants had a preferred glucose range and tended to administer additional correction doses because their desired glucose levels were below 8 mmol/L, often targeting values as low as 6 mmol/L. This preference contrasts with the clinically recommended glucose range of 3.9–10.0 mmol/L, as outlined in the standard guidelines.^1,2 Such tighter glucose targets may increase the risk of hypoglycaemic episodes. ⁶¹ Therefore, further research in this aspect is encouraged to understand this issue better and to prevent inappropriate treatment decisions for people with T1DM.

A significant benefit of providing glucose data textual descriptions is shown among participants who reported not receiving T1DM structured education. According to the NHS National Diabetes Audit, only 7.20% of people who registered with T1DM attended structured education. ⁶² This suggests that many individuals with T1DM lack comprehensive knowledge about their condition, and this knowledge gap has been identified as one of the major barriers to effective self-management of T1DM.^63,64 Our study indicates that providing glucose data textual descriptions can help those without structured education make better treatment decisions in self-management. This is particularly promising, as it suggests that we can support a significant number of people who have not had the opportunity to access formal training.

In terms of preference ranking between types of glucose monitoring systems, the Baseline system was top-ranked. However, the mixed scores among all system interfaces suggest that participants were also interested in a change in the existing glucose monitoring system. This finding aligns with previous studies, which suggested that people with T1DM preferred having a textual description of glucose data, as discussed in the related work section.^11,12,65

This study also revealed a striking paradox: despite achieving higher correctness scores, participants in the experimental arm rated the textual descriptions less favourable than the control arm. This discrepancy aligns with prior research indicating that users may perform better while disliking a system, or conversely, they may prefer a system while performing poorly.^66,67 While textual descriptions can enhance accuracy, they may also introduce cognitive bias, particularly if the information is perceived as complex or unfamiliar. ⁶⁷ This could lead them to favour a familiar visualisation over a newly developed one. The participant feedback supports this concern, with some participants reporting that the text-heavy interface detracted from usability. This suggests that while textual descriptions can be beneficial, their integration must be carefully optimised to avoid overwhelming users.

This study has some limitations. First, the glucose monitoring systems were limited to only the Dexcom G6 and FreeStyle Libre 2 devices, as these are the devices prescribed by the NHS in the UK. Second, the recruitment process via online diabetes communities and the JDRF could cause sampling bias, which could limit the generalisability to the broader T1DM population internationally. Finally, our survey accepted free-text responses, enabling participants to provide unrestricted input. However, these responses may not have accurately represented their actual treatment behaviours in daily life.

Conclusion

This study examined the effectiveness of the glucose monitoring system incorporating textual descriptions in improving decision-making accuracy compared to existing systems for people with T1DM. The findings suggest that explicit text-based glucose data improved participants’ treatment decision-making and received positive opinions overall, with minor unfavourable comments regarding the amount of the text displayed in the new system. This provides compelling evidence that our current research approach is heading in the right direction towards minimising the data interpretation burden for people with T1DM.

It is essential to focus on technological advancements and the accuracy of healthcare devices. However, it is equally important to prioritise research on information visualisation to ensure users receive the most understandable and practical information for managing their conditions. Adopting a patient-centred approach and tailoring visualisation to meet users’ needs are needed as they will significantly enhance the tool’s overall usability and improve the users’ quality of life.

Future work could (1) explore alternative data presentations that best display glucose data while enhancing people’s acceptance, perceived usefulness, and ease of use, incorporating individuals with different diabetes literacy and reading levels, (2) provide contextual information in each glucose monitoring systems to give participants real-life scenarios and explore the results influenced by this modification and (3) evaluate participants’ response time for each interface to gain insights into whether they take more or less reading time and analyse the results with their self-reported preferences.

Appendix A

Figure A1.

Glucose monitoring system interfaces in qualitative study. (a) Glucose monitoring system interface of CGM: (A) Baseline, (B) Meaning, (C) Action and (D) Meaning-Action. (b) Glucose monitoring system interface of Flash: (A) Baseline, (B) Meaning, (C) Action and (D) Meaning-Action.

CGM, continuous glucose monitor.

Appendix B

A full list of glucose monitoring system interfaces.

Table B1.

All glucose monitoring interfaces (adapted from previous study). ¹⁰

CGM, continuous glucose monitor.

Appendix C

Table C1.

Results from the PUPEOU survey. The results are shown in percentages.

Scale	PUPEOU 1 (%)		PUPEOU 2 (%)		PUPEOU 3 (%)		PUPEOU 4 (%)		PUPEOU 5 (%)
	Control	Experimental	Control	Experimental	Control	Experimental	Control	Experimental	Control	Experimental
Strongly agree	6.82	4.76	11.36	7.14	11.36	11.90	11.36	7.14	13.64	11.90
Agree	22.73	21.43	20.45	23.81	25.00	23.81	27.27	23.81	38.64	33.33
Somewhat agree	25.00	28.57	20.45	21.43	18.18	19.05	18.18	23.81	20.45	19.05
Neither agree nor disagree	25.00	16.67	27.27	14.29	25.00	16.67	18.18	16.67	11.36	11.90
Somewhat disagree	6.82	9.52	9.09	11.90	4.55	7.14	9.09	11.90	4.55	9.52
Disagree	6.82	16.67	4.55	16.67	6.82	9.52	4.55	7.14	6.82	4.76
Strongly disagree	6.82	2.38	6.82	4.76	9.09	11.90	11.36	9.52	4.55	9.52

PUPEOU, Perceived Usefulness and Perceived Ease of Use.