Volume 23, Issue 10 p. 2261-2268
ORIGINAL ARTICLE
Open Access

Predictors of diabetes-related distress before and after FreeStyle Libre-1 use: Lessons from the Association of British Clinical Diabetologists nationwide study

Harshal Deshmukh MRCP

Harshal Deshmukh MRCP

Hull University Teaching Hospitals NHS Trust and the University of Hull, Hull, UK

Search for more papers by this author
Emma G. Wilmot MRCP

Emma G. Wilmot MRCP

University Hospitals of Derby and Burton NHS Foundation Trust, Derby, UK

Search for more papers by this author
Robert Gregory DM

Robert Gregory DM

Leicester General Hospital, Leicester, UK

Search for more papers by this author
Dennis Barnes FRCP

Dennis Barnes FRCP

Tunbridge Wells Hospital, Tunbridge Wells, UK

Search for more papers by this author
Parth Narendran MD

Parth Narendran MD

Queen Elizabeth Hospital Birmingham and University of Birmingham, Birmingham, UK

Search for more papers by this author
Simon Saunders MD

Simon Saunders MD

Warrington and Halton Teaching Hospitals NHS Foundation Trust, Warrington, UK

Search for more papers by this author
Niall Furlong MD

Niall Furlong MD

St Helens and Knowsley Teaching Hospitals NHS Trust, St Helens, UK

Search for more papers by this author
Shafie Kamaruddin MRCP

Shafie Kamaruddin MRCP

Darlington Memorial Hospital, Darlington, UK

Search for more papers by this author
Rumaisa Banatwalla MRCP

Rumaisa Banatwalla MRCP

St Peter's Hospital, Chertsey, UK

Search for more papers by this author
Roselle Herring FRCP

Roselle Herring FRCP

Royal Surrey County Hospital, Guildford, UK

Search for more papers by this author
Anne Kilvert MD

Anne Kilvert MD

Northampton General Hospital NHS Trust, Northampton, UK

Search for more papers by this author
Jane Patmore MRCP

Jane Patmore MRCP

Hull University Teaching Hospitals NHS Trust and the University of Hull, Hull, UK

Search for more papers by this author
Chris Walton FRCP

Chris Walton FRCP

Hull University Teaching Hospitals NHS Trust and the University of Hull, Hull, UK

Search for more papers by this author
Robert E. J. Ryder MD

Robert E. J. Ryder MD

City Hospital, Birmingham, UK

Search for more papers by this author
Thozhukat Sathyapalan MD

Corresponding Author

Thozhukat Sathyapalan MD

Correspondence

Prof. Thozhukat Sathyapalan, Academic Endocrinology, Diabetes and Metabolism, University of Hull/Hull and East Yorkshire Hospitals NHS Trust, Hull, UK.

Email: [email protected]

Search for more papers by this author
First published: 17 June 2021
Citations: 6

Abstract

Aim

To identify the baseline demographic and clinical characteristics associated with diabetes-related distress (DRD) and factors associated with improvement in DRD after initiating use of the FreeStyle Libre (FSL) in people living with type 1 diabetes (T1D).

Methods

The study was performed using baseline and follow-up data from the Association of British Clinical Diabetologists nationwide audit of people with diabetes who initiated use of the FSL in the United Kingdom. DRD was assessed using the two-item diabetes-related distress scale (DDS; defined as the average of the two-item score ≥3). People living with T1D were categorized into two groups: those with high DRD, defined as an average DDS score ≥3 and those with lower DRD, defined as a DDS score <3. We used a gradient-boosting machine-learning (GBM) model to identify the relative influence (RI) of baseline variables on average DDS score.

Results

The study population consisted of 9159 patients, 96.6% of whom had T1D. The median (interquartile range [IQR]) age was 45.1 (32-56) years, 50.1% were women, the median (IQR) baseline body mass index was 26.1 (23.2-29.6) kg/m2 and the median (IQR) duration of diabetes was 20 (11-32) years. The two components of the DDS were significantly correlated (r2 = 0.73; P < 0.0001). Higher DRD was prevalent in 53% (4879/9159) of people living with T1D at baseline. In the GBM model, the top baseline variables associated with average DDS score were baseline glycated haemoglobin (HbA1c; RI = 51.1), baseline Gold score (RI = 23.3), gender (RI = 7.05) and fear of hypoglycaemia (RI = 4.96). Follow-up data were available for 3312 participants. The top factors associated with improvement in DDS score following use of the FSL were change in Gold score (RI = 28.2) and change in baseline HbA1c (RI = 19.3).

Conclusions

In this large UK cohort of people living with T1D, diabetes distress was prevalent and associated with higher HbA1c, impaired awareness of hypoglycaemia and female gender. Improvement in glycaemic control and hypoglycaemia unawareness with the use of the FSL was associated with improvement in DRD in people living with T1D.

1 INTRODUCTION

Diabetes-related distress (DRD) is increasingly recognized as an important determinant of suboptimal glycaemic control and complication risk in people living with type 1 diabetes (T1D).1-6 DRD is distinct from depression and anxiety and is a product of dealing with the unrelenting demands and limitations that diabetes imposes on the life of a person with diabetes.1, 7-9 DRD arises from the demands of self-care associated with diabetes and is the product of emotional adjustments.1 DRD in T1D is also distinct from DRD in type 2 diabetes, given the demands of multiple daily insulin injections, frequent blood glucose monitoring, hypoglycaemia, and general burnout due to the incessant management needs of T1D.2-5, 9

There are few population-based studies looking at the prevalence of DRD in people living with T1D. Most of the data are reported from small cross-sectional studies, which are not representative of the whole population of people with T1D, and these show a prevalence of DRD of up to 40%.3-5, 10-13 We14 and others15-17 have recently shown significant improvement in DRD with the use of the FreeStyle Libre (FSL) monitor. The FSL is a novel method of glucose monitoring for people living with diabetes. It consists of a sensor the size of a two-pound coin, is worn on the arm and has a very fine sensing electrode, which is automatically inserted just under the skin when the user applies the sensor to the skin. It measures blood glucose readings in the subcutaneous fluid. In 2017, the FSL flash glucose monitor became available on the NHS Drug Tariff and is used by approximately 40% of people living with T1D in the United Kingdom. There are no data looking at factors associated with improvement in DRD with the use of the FSL. It is important to identify the risk factors associated with DRD and the factors that influence the reduction in DRD resulting from use of the FSL. This could enhance understanding of the potential causes of DRD in people living with T1D and thereby suggest strategies to alleviate it. The objective of the present study was to identify the baseline demographic and clinical characteristics associated with DRD and the factors associated with improvement in DRD with the use of the FSL monitor.

2 METHODS

2.1 Ethical approval

The Association of British Clinical Diabetologists (ABCD) nationwide audit programme has Caldicott Guardian approval. The programme is an audit rather than a research programme. The National Health Service (NHS) encourages audit of clinical practice, and there are guidelines, which were followed, stipulating that contributing centres only collect data from routine clinical practice and that all data collected are anonymized at the point of submission to the central database.

2.2 Data collection

Data for this study were obtained from the nationwide audit of the FSL conducted by the ABCD (http://www.diabetologists-abcd.org.uk/n3/FreeStyle_Libre_Audit.htm). This nationwide audit commenced in November 2017 using paper forms to collect data, which were then entered onto a secure online tool on the NHS information technology network. This network provides maximum security and allows the analysis of anonymized national audit data. The tool has the facility to detect data from the same patient entered on two sites (eg, hospital and primary care) and to merge the data when exported. Data were collected at baseline and follow-up during routine clinical care. Baseline pre-FSL data included demographics, source of FSL funding, previous structured diabetes education completion, glycated haemoglobin (HbA1c) values from the previous 12 months, Gold score18 (used to assess hypoglycaemia awareness), severe hypoglycaemia, paramedic callouts, and hospital admissions due to hypoglycaemia, hyperglycaemia and diabetic ketoacidosis over the previous 12 months. We also collected diabetes-related distress scale (DDS) scores at baseline and follow-up using the two-item diabetes distress screening instrument, the DDS2.19 The DDS2 is a two-item diabetes distress screening instrument asking respondents to rate on a six-point scale the degree to which the following items caused distress: (a) feeling overwhelmed by the demands of living with diabetes; and (b) feeling that I am often failing with my diabetes regimen. "DDS2" refers to the two-item diabetes screening instrument, and DDS-1 and DDS-2 represent the two components of the instrument. A score of ≥3 (moderate distress) discriminated high- from low-distress subgroups and this cut-off provided the highest sensitivity and specificity.19 Hence, we used the cut-off of ≥3 (moderate distress) to discriminate high distress from low distress. At follow-up, we collected data on DDS2, HbA1c and Gold score, along with FSL-specific measures, such as the number of scans per day and time in range. The baseline DDS score and variables relating to resource utilization, such as episodes of severe hypoglycaemia, episodes of hypoglycaemia and hyperglycaemia requiring hospital admissions and paramedic callouts, were measured at the first visit at which FSL use was initiated. Follow-up variables were collected at the first follow-up visit after the patients initiated FSL use.

2.3 Statistical methods

For reporting all of the study outcomes, including glycated haemoglobin (HbA1c), Gold score, paramedic callouts and hospital admissions, we restricted the statistical analysis to those with at least one follow-up data entry. The χ2 test of association was used to compare categorical variables. The Mann-Whitney U-test or t-tests were used to compare continuous variables associated with improvement in DDS score after initiating use of the FSL monitor.

2.4 Machine-learning methods

To identify risk factors for DRD and predictors of change in DDS score following the use of the FSL we used a gradient-boosting machine-learning (GBM) model. Based on the input predictor variables, this machine-learning algorithm consecutively fits new decision trees to provide a more accurate prediction of response. The primary concept of this algorithm is a learning procedure that results in consecutive error fitting, with each decision tree chosen to minimize the loss function.20, 21 The GBM model generates the relative importance (RI) of each variable in the model by identifying whether that variable was selected to split on during the tree-building process and how much the squared error (overall trees) decreased as a result of this variable. Results from the GBM model are presented as different importance levels of each variable, which are reported as an RI value. All analyses were conducted in R4.0.2 with library GBM and CARET (http://www.r-project.org/). To identify the association with baseline DDS score, we included 14 variables (Supplementary Table S1) in the GBM model, and to identify variables associated with post-FSL improvement in DDS score, 24 variables were used in the GBM model. The post-FSL model included a larger number of variables as it consisted of derived variables from follow-up data, such as change in HbA1c (delta HbA1c) and Gold score (delta Gold). An average item score of ≥3 (high and moderate distress) discriminated the high- from the low-distress subgroup, and a categorical variable was created to be used for the GBM model at baseline and at follow-up. In the GBM model, the data were divided into a training (two-thirds of the data) and a testing set, and the results from the training set were used in the testing set to calculate the model accuracy. The model accuracy and the area under the curve (AUC) were evaluated using the testing dataset. The hyperparameters were selected using a grid search and are described in the supplementary material. We selected all baseline and follow-up variables for the model building which could affect DRD. This was based on both a clinical understanding of factors that can influence DRD and prior literature. We report all RI values for all the variables included in the model in the supplementary materials. An essential advantage of using GBM models is that it deals with missing values as containing information rather than missing at random. During tree building, split decisions for every node are found by minimizing the loss function and treating missing values as a separate category. Although there were 3.4% missing data for HbA1c and 14% missing data for Gold score, this is likely to have had a minimal effect on prediction modelling.

Although GBM models have several advantages as compared to traditional regression analysis, the impurity-based feature importance in the GBM model has two disadvantages: 1) it calculates the variable importance based on training data, and 2) it tends to favour high cardinality features. Permutation feature importance (PIMP)22 is an alternative to impurity-based feature importance that does not have these flaws. The PIMP model ranks the variables based on the increase in the model's prediction error after permuting the variable. If shuffling the variable values in the model increases the model error, then the variable is classed as “important” as the model relied on the feature for its prediction, whereas the variable is “unimportant” if shuffling its values leaves the model error unchanged. We also used the random forest model as an alternative method of feature selection to confirm the results of the GBM models. The random forest can be used to find a set of predictors that best explains the variance in the response variable. This analysis was performed with R-packages “randomForest”, “vita”, and “varImp.

The results of the machine-learning algorithm were examined using logistic regression analysis to understand the direction of the effect. The top predictors from the GBM model were included in the logistic regression model to understand the directionality of the effect estimate. For the baseline analysis, the DDS2 was converted into a categorical variable (average of DDS-1 and DDS-2) and average DDS2 score ≥3 (high and moderate distress) discriminated high from low distress. In the post-FSL follow-up model, those with an average DDS2 score ≥3 (high and moderate distress) at baseline and follow-up DDS2 <3 were classed as having transitioned to lower DRD.

3 RESULTS

Table 1 shows the baseline demographic and clinical characteristics of the study population. The study population consisted of 9159 patients, 96.6% of whom had T1D, and the remaining patients used the FSL for other indications such as pregnancy, poorly controlled type 2 diabetes and renal dialysis. The median (interquartile range [IQR]) age of the study population was 45.1 (32-56) years and 50.1% were women. The median (IQR) baseline body mass index (BMI) was 26.1 (23.2-29.6) kg/m2 and the median (IQR) duration of diabetes was 20 (11-32) years. The majority of the study participants had T1D (96.6%), with 23% using insulin pump therapy. The median (IQR) baseline HbA1c was 67.5(58-79.3) mmol/mol and the median (IQR) Gold score was 2 (1-4). The mean DDS-1 and DDS-2 scores across the baseline study population were 3 (2-4). The overall prevalence of DRD (mean of DDS-1 and DDS-2 ≥3) was 53% (4879/9159). Of the 9159 study participants, 3312 had at least one follow-up DDS score with a mean follow-up period of 7.2 (±6.3) months. The baseline demographic and clinical characteristics of those with and without follow-up were similar.

TABLE 1. Baseline demographic and clinical characteristics of the study population
Baseline (n = 915) Follow-up (n = 3312)
Age, median (IQR) years 45.1 (32-56) 47 (34-58)
Sex: female, n (%) 4573 (50.1) 2230 (50.2)
Baseline BMI, median (IQR) kg/m2 26.1 (23.2-29.6) 26 (12-35)
Duration of diabetes, median (IQR) years 20 (11-32) 22 (43.8)
Type 1 diabetes, n (%) 8816 (96%) 4259 (96%)
Insulin pump, n (%) 2157 (23%) 1020 (23%)
White ethnicity, n (%) 7303 (80%) 3549 (80%)
Pre-FSL HbA1c, median (IQR) mmol/mol 67.5 (58-79.3) 66 (57-76.5)
Baseline Gold score, median (IQR) 2 (1-4) 2 (1-4)
Mean DDS-1 score, median (IQR) 3 (2-4) 3 (2-4)
Mean DDS-2 score, median (IQR) 3 (2-4) 3 (2-4)
  • Abbreviations: BMI, body mass index; DDS, diabetes-related distress scale; FSL, FreeStyle Libre monitor; HbA1c, glycated haemoglobin; IQR, interquartile range.

3.1 Factors associated with baseline DDS score

Figure 1 shows the results of the GBM model with the top six baseline variables associated with baseline average DDS score. In the GBM model, the top baseline variables associated with average DDS score were baseline HbA1c (RI = 51.1), baseline Gold score (RI = 23.3), gender (RI = 7.05) and fear of hypoglycaemia (RI = 4.96). Supplementary Table S1 shows the results of linear regression analysis and shows a statistically significant association of higher baseline HbA1c, higher baseline Gold score and female gender with baseline DDS score.

Details are in the caption following the image
Gradient boosting machine-learning model showing the top six baseline variables associated with baseline diabetes-related distress (DRD)

Supplementary Figure S1 shows the results of the GBM model including the top variables associated with DDS-1 and DDS-2. In the GBM model, the top baseline variables associated with DDS-1 were HbA1c (RI = 35.5), Gold score (RI = 22.5) and glucose variability as an indication for initiating FSL use (RI = 15.02). The top baseline variables associated with DDS-2 were HbA1c (RI = 48.7), Gold score (RI = 20.9) and female gender (RI = 5.6). The model accuracy was 0.44 (0.42-0.47) for DDS-1 and 0.44 (0.42-0.46) for DDS-2. The AUC of the baseline DRD model was 0.69. The findings of the machine-learning model were confirmed by using the top six variables in a logistic regression model, and these were significantly associated with DDS score (P <0.05).

Figure 2A shows the mean DDS score in those with hypoglycaemia awareness and hypoglycaemia unawareness at baseline. The DDS score was significantly higher in those with hypoglycaemia unawareness (Gold score ≥4 at baseline; ANOVA: P <0.0001). Figure 2B shows three strata of baseline HbA1c and baseline DDS score in each category. The mean DDS score was lowest in those with baseline HbA1c <47.5 mmol/mol compared to those with HbA1c between 47.5 and 69.4 mmol/mol, and highest in those with baseline HbA1c >69.4 mmol/mol (ANOVA: P <0.0001).

Details are in the caption following the image
Association of baseline Gold score and glycated haemoglobin (HbA1c) with diabetes-related distress (DRD), with (A) showing DDS scores in those with and without hypoglycaemia unawareness at baseline, and (B) showing three strata of baseline HbA1c and baseline DDS scores in each category

Supplementary Figure S5A and B show the top variables associated with DRD using the random forest model and the PIMP model. Both the models showed results comparable to the GBM models, with baseline HbA1c, Gold score, glucose variability and gender being important predictors of DRD.

3.2 Factors associated with improvement in DDS score after initiating FSL use

In those with paired data, DDS-1 score decreased from 2.4 (±1.3) to 2.2 (±1.2; P <0.0001), while the DDS-2 score decreased from 2.4 (±1.3) to 2.2 (±1.3; P <0.0001). In those with paired data, after initiating use of the FSL, the prevalence of moderate to severe distress on the DDS2 reduced from 50% to 26%.

Table 2 shows the results of univariate analysis in those with and without improvement in DDS score following the use of FSL. Improvement in DDS score at follow-up was associated with a lower follow-up HbA1c and a lower follow-up Gold score, and a higher baseline Gold score.

TABLE 2. Comparison of demographic and clinical characteristics in those with and without improvement in diabetes-related distress score following initiation of FreeStyle Libre monitor use
Improvement in DDS following FSL use (970) No change in DDS with FSL use (2342) P
Age, years 47.3 (±15.3) 46.6 (±15.2) 0.22
Sex: female, n (%) 519 (53) 1211 (51) 0.36
Baseline BMI, kg/m2 26.8 (±5.3) 26.6 (±5.1) 0.96
Duration of diabetes, years 25.4 (±14.9) 25.6 (±14.8) 0.9
Pre-FSL HbA1c, mmol/mol 69.0 (±16.3) 67.9 (16.2) 0.07
Post-FSL HbA1c, mmol/mol 62.2 (±12.5) 63.6 (±14.5) 0.004
Baseline Gold score 2.7 (±1.7) 2.5 (±1.6) 0.0004
Post-FSL Gold score 2.1 (±1.3) 2.3 (±1.5) 0.01
Average number of FSL scans in 14 days 12.5 (±13.8) 12.3 (±12.4) 0.72
  • Note: Data are mean (±SD) unless otherwise indicated.
  • Abbreviations: BMI, body mass index; DDS, diabetes-related distress scale; FSL, FreeStyle Libre monitor; HbA1c, glycated haemoglobin.

Figure 3 shows the results of the GBM model, with top variables associated with improvement in average DDS score with the use of FSL at follow-up. The top factors associated with improvement in DDS following initiation of FSL use were change in Gold score (RI = 28.2) and change in baseline HbA1c (RI = 19.3). Supplementary Table S4 shows the results of the linear regression analysis, which showed a statistically significant association of reduction in Gold score and drop in baseline HbA1c with improvement with DDS.

Details are in the caption following the image
Gradient boosting machine-learning model showing the top variables associated with improvement at follow-up in average diabetes-related distress scale (DDS) score after initiation of FreeStyle Libre (FSL) monitor use. BMI, body mass index; HbA1c, glycated haemoglobin

Supplementary Figure S2 shows the results of the GBM model including the top variables associated with improvement in DDS score following initiation of FSL use. For the DDS-1, lower Gold score at follow-up (RI = 39.8), lower follow-up HbA1c (RI = 18.3), lower follow-up Gold score (RI = 11.58), and a higher number of FSL scans per day (RI = 8.65) were the top factors associated with transitioning to a low DDS-1 score at follow-up. For DDS-2 absolute change in Gold score (RI = 38.2) and in HbA1c (RI = 32.5), follow-up Gold score (RI = 12.6) and a higher number of FSL scans per day (RI = 4.7) were top factors associated with transitioning to a low DDS-2 score at follow-up. A reduction in HbA1c at follow-up had a higher influence on the change in DDS-2 "failing" as compared to change in DDS-1 “overwhelmed” at follow-up. The model accuracy for the DDS-1 model was 0.84 (0.81-0.86) and 0.80 (0.78-0.82) for the DDS-2 model. The AUC for the follow-up DRD model was 0.58. The findings of the machine-learning model were confirmed by using the top six variables in a logistic regression model and these were significantly associated with DDS (P <0.05).

Supplementary Figures S6A and B show the top variables associated with improvement in DRD using the random forest model and the PIMP model. Both the models show results comparable to the GBM models, with change in HbA1c and Gold score and a time in target HbA1c over 14 days as top predictors of improvement in DRD.

4 DISCUSSION

We present the results of the largest study looking at the factors associated with DRD in people with TID before and after initiating use of the FSL in the United Kingdom. The prevalence of DRD is high. We showed that improvement in HbA1c and hypoglycaemia awareness is associated with DRD in FSL-eligible people with T1D in the UK population. We also showed that improvements in HbA1c, hypoglycaemia awareness and a higher number of scans per day are associated with improvements in DRD in people living with TID.

Longitudinal population-based and cross-sectional studies in people living with T1D have demonstrated excess rates of depression compared to those in the population without diabetes.1, 5-7 However, numerous studies have shown that DRD is often misinterpreted as depression in the T1D population.7-9 There have been several small cross-sectional studies5, 10, 13, 23 looking at DRD in T1D, which have used several scales to measure DRD. The Diabetes Distress Scale (DDS17), a 17-item self-report measure24 of overall diabetes distress, is the most commonly used measure to identify DRD. This score is used in assessing emotional burden, physical-related distress, regimen-related distress, and interpersonal distress associated with T1D. The DDS1724 has been used in the United States and in European populations to understand and quantify DRD. However, this is a long questionnaire and can be challenging to use in clinical settings and to screen for DRD in large population-based studies. As an alternative to the DDS17, Lawrence et al.19 developed a brief two-question diabetes distress screening instrument, the DDS2. The sensitivity and specificity of this composite instrument is 95% and 85%, respectively, and it can be used easily in clinical settings. In our nationwide study, we utilized this two-item composite instrument to understand the prevalence of DRD in T1D, the factors which affect it at baseline, and predictors of improvement following flash glucose monitoring.

The baseline prevalence of DRD was high in our cohort, with 53% patients having moderate to severe distress, as compared to previous studies, which have reported a DRD prevalence of 40%.25, 26 This is likely to reflect our study population, which was mostly restricted to those who fulfilled the criteria set for NHS funding in the United Kingdom, limiting the generalizability of the conclusions. However, these access criteria have now resulted in more than 40% of people living with diabetes being reimbursed for the FSL, which indicates that the selection criterion will encompass a substantial proportion of people with T1D.

The association of DRD with baseline clinical and demographic characteristics in people with T1D is complicated by the hidden and complex correlation between the baseline clinical and demographic variables. We used GBM, a machine-learning algorithm that has better efficiency in accounting for correlation, missing data and outliers as compared to a standard regression analysis approach.21 We show that improvement in HbA1c has the largest relative influence on DDS-1 (feeling overwhelmed with demands of living with diabetes), while the degree of hypoglycaemia unawareness (Gold score) had the largest relative association with DDS-2 (feeling that I am failing with my diabetes routine). Our study is in agreement with previous studies, which have shown the detrimental effect of higher HbA1c and hypoglycaemia unawareness in people with T1D.4, 13, 23, 25 For example, a longitudinal study of 280 consecutive T1D patients in the UK population showed that DRD showed a significant correlation with HbA1c and Gold score independently and with synergistic effect.13 Another study4 in 450 adolescents with T1D in Australia showed a significant positive correlation between HbA1c and diabetes distress and found that this correlation was stronger than the relationship between HbA1c and depressive symptoms. Interestingly, female gender was associated with DRD at baseline, which is in agreement with a previous study.13 It is possible that women and girls have additional blood glucose variations during menstruation, pregnancy and the post-gestational period.27, 28 These variations are more likely to cause hypoglycaemia unawareness and poor glycaemic control and can contribute to DRD. However, interestingly, female gender was not the top predictor of the improvement in DRD following the use of FSL. This suggests that glucose monitoring is associated with improvement in DRD irrespective of gender. Further studies are needed to investigate the gender-specific causes of DRD in people with diabetes.

Our data also show that baseline BMI, the absolute value of HbA1c following the use of FSL, number of FSL scans per day and time in range were associated with improved DRD after initiating use of the FSL. It is possible that those with a higher baseline BMI also had a higher baseline HbA1c29 and hence had a larger improvement in glycaemic control and therefore it was associated with improvement in DRD. Interestingly, BMI was not the top predictor of baseline DRD and this is consistent with the previous study in people living with T1D.30 Our study also showed that engagement with the FSL and improved time in range and HbA1c were amongst the top factors associated with improvement in DRD. Polonsky et al.31 recently described an association between better time in range with better mood. We are only starting to understand the relationship between time in range and psychological outcomes, and further analyses will be welcome.31

We have recently shown that use of the FSL significantly improved glycaemic control, hypoglycaemia awareness and DRD in people living with T1D.14 This cohort gave us a unique opportunity to assess the effects of baseline clinical and demographic features in FSL users and the associated changes in DRD in people living with T1D using the FSL. We showed that improvements in hypoglycaemia awareness and glycaemic control had the largest influence on improvement in DRD when patients living with T1D used the FSL. We also showed that engagement with FSL as evidenced by the number of FSL scans performed per day was associated with an improvement in DRD. Higher HbA1c and impaired awareness of hypoglycaemia are associated with DRD, and can have a negative impact on self-management, resulting in further DRD.1, 2, 10

The present study had several limitations. The main limitations include the lack of a comparator arm, and the fact that it was a real-world observational cohort study as opposed to a randomized controlled trial, with opportunistic rather than systematically organized data collection. Furthermore, socioeconomic status has been shown to be an important predictor of DRD and this variable was not captured in our nationwide dataset. Despite these limitations, these data represent the largest UK-wide, real-world data looking at the baseline demographic and clinical factors associated with DRD and the effect of flash glucose monitoring on DRD. Further, our GBM models for baseline DRD had moderate accuracy, indicating that our models do not fully capture all the sources of DRD in this population. For the first time, we have shown that the use of flash glucose monitoring can improve DRD by improving hypoglycaemia awareness and glycaemic control.

In summary, this real-world study demonstrates that high HbA1c and impaired awareness of hypoglycaemia are important risk factors for DRD. Our study shows that use of the FSL, particularly with frequent scanning, can improve both glycaemic control and DRD in people living with T1D.

ACKNOWLEDGMENTS

The authors would like to thank all the clinicians and support staff who participated in the nationwide study, listed at https://abcd.care/Resource/ABCD-Freestyle-Libre-Audit-Contributors.

    CONFLICT OF INTEREST

    The ABCD nationwide FSL audit is supported by a grant from Abbott Laboratories. E.G.W. serves on the advisory panel for Abbott Diabetes Care, Dexcom, and Eli Lilly and Company; has received research support from Diabetes UK; and is on the speakers bureau for Abbott Diabetes Care, Dexcom, Eli Lilly and Company, Insulet Corporation, Novo Nordisk, and Sanofi. C.W. has a spouse/partner serving on the advisory panel for Celgene and on the speakers bureau for LEO Pharma and Novartis. R.E.J.R. serves on the advisory panel for Novo Nordisk A/S and on the speakers bureau for BioQuest. T.S. is on the speakers bureau for Novo Nordisk Foundation and reports a relationship with Bristol-Myers Squibb, Eli Lilly and Company, and Sanofi. No other potential conflicts of interest relevant to this article were reported. The FSL audit was independently initiated and performed by ABCD, and the authors remain independent in the analysis and preparation of this report.

    AUTHOR CONTRIBUTIONS

    E.G.W., C.W., R.E.J.R. and T.S. conceived the study. H.D., E.G.W., C.W., R.E.J.R. and T.S. contributed to the data analysis. H.D. wrote the first draft of the manuscript. All of the authors contributed to the writing of the manuscript and made extensive comments, criticism, and changes in the final draft of the paper. All of the authors saw the final version of the manuscript. H.D. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

    PEER REVIEW

    The peer review history for this article is available at https://publons.com/publon/10.1111/dom.14467.

    DATA AVAILABILITY STATEMENT

    The data that support the findings of this study are available from the corresponding author upon reasonable request.