Q-Score: development of a new metric for continuous glucose monitoring that enables stratification of antihyperglycaemic therapies

Background Continuous glucose monitoring (CGM) has revolutionised diabetes management. CGM enables complete visualisation of the glucose profile, and the uncovering of metabolic ‘weak points’. A standardised procedure to evaluate the complex data acquired by CGM, and to create patient-tailored recommendations has not yet been developed. We aimed to develop a new patient-tailored approach for the routine clinical evaluation of CGM profiles. We developed a metric allowing screening for profiles that require therapeutic action and a method to identify the individual CGM parameters with improvement potential. Methods Fifteen parameters frequently used to assess CGM profiles were calculated for 1,562 historic CGM profiles from subjects with type 1 or type 2 diabetes. Factor analysis and varimax rotation was performed to identify factors that accounted for the quality of the profiles. Results We identified five primary factors that determined CGM profiles (central tendency, hyperglycaemia, hypoglycaemia, intra- and inter-daily variations). One parameter from each factor was selected for constructing the formula for the screening metric, (the ‘Q-Score’). To derive Q-Score classifications, three diabetes specialists independently categorised 766 CGM profiles into groups of ‘very good’, ‘good’, ‘satisfactory’, ‘fair’, and ‘poor’ metabolic control. The Q-Score was then calculated for all profiles, and limits were defined based on the categorised groups (<4.0, very good; 4.0–5.9, good; 6.0–8.4, satisfactory; 8.5–11.9, fair; and ≥12.0, poor). Q-Scores increased significantly (P <0.01) with increasing antihyperglycaemic therapy complexity. Accordingly, the percentage of fair and poor profiles was higher in insulin-treated compared with diet-treated subjects (58.4% vs. 9.3%). In total, 90% of profiles categorised as fair or poor had at least three parameters that could potentially be optimised. The improvement potential of those parameters can be categorised as ‘low’, ‘moderate’ and ‘high’. Conclusions The Q-Score is a new metric suitable to screen for CGM profiles that require therapeutic action. Moreover, because single components of the Q-Score formula respond to individual weak points in glycaemic control, parameters with improvement potential can be identified and used as targets for optimising patient-tailored therapies. Electronic supplementary material The online version of this article (doi:10.1186/s12902-015-0019-0) contains supplementary material, which is available to authorized users.


Background
Continuous glucose monitoring (CGM) is a new area in diabetes care and management [1][2][3]. The advantage of CGM is that daily glucose profiles can be visualised completely and precisely, allowing the identification of 'weak points' in glycaemic control. Each CGM record contains a wealth of data, and 48 parameters are currently available for the analysis of glucose profiles [4][5][6][7][8][9][10]. However, the analysis of CGM has not yet been standardised [10][11][12]. Studies by the Juvenile Diabetes Research Foundation Continuous Glucose Monitoring Study Group employed variables that described glucose levels (time/day above or below the target range), variability (standard deviation [SD], coefficient of variation [CV], mean amplitude of glycaemic excursions [MAGE], mean absolute rate of change [MARC]), and summary values for hypo-and hyperglycaemia (area under the curve for glucose [AUC G ], low blood glucose index [LBGI], and high blood glucose index [HBGI]) [5,[13][14][15]. Accordingly experts have suggested the use of parameters that allow the assessment of target range, glucose exposure, glucose variability, and hyper-and hypoglycaemia [10][11][12].
Mean blood glucose (MBG) is frequently used to reveal central glycaemic tendency. For evaluating glucose variability, a variety of parameters have been described, including SD, range, MAGE, the continuous overall net glycaemic action (CONGA), and the mean of daily differences (MODD) [11][12][13][15][16][17][18][19][20]. Hypo-and hyperglycaemic episodes are assessed based on the time spent and AUC G of CGM segments that appear outside the target range, where hypoglycaemia is defined as time outside the glucose target range t G <3. 9 and AUC G <3. 9 , and hyperglycaemia is defined as t G >8. 9 and AUC G >8. 9 . Risk scores for hypo-or hyperglycaemia are based on the LBGI and HBGI, respectively [19,20]. The glycaemic risk assessment for diabetes equation (GRADE) was also developed for assessing glycaemic risk [16]. These parameters are valuable for clinical research. However, they are often impractical for use by the clinician in routine diabetes care.
Recent studies addressed the need for a single metric that allows for the assessment of short-term glycaemic control using CGM, similar to the way in which glycosylated haemoglobin (HbA 1c ) allows for the assessment of long-term glycaemic control [21][22][23]. Rawlings et al. [21] developed a graphical user interface to evaluate CGM profiles (CGM-GUIDE © ) based on quantitative measures of glucose variability. Thomas et al. [22] described the 'Glucose Pentagon' , which combines different summary measures derived from CGM profiles (including parameters describing glycaemic variability) and HbA1c for assessing glycaemic control. Marling et al. [23] developed a 'consensus perceived glycaemic variability metric' that captures the gestalt perceptions of experienced physicians using an automatic algorithm. However, none of these new methods has yet been introduced into routine diabetes care regimens.
The aim of this study was to develop a metric that facilitates objective assessments of glucose profiles and screening for profiles that require therapeutic action. Moreover, in order to allow patient-tailored therapy, we aimed to develop an automated method for identification of improvement potential in a given profile.

Patient data
CGM profiles and self-control data were recorded in earlier studies [24][25][26][27][28], which were approved by the Regional Ethics Review Board of the University of Greifswald (Germany). All included subjects provided informed consent to participate in CGM and data analysis. Data from 1,562 subjects (females/males; 499/ 1,063) with type 1 and type 2 diabetes (n = 48 and n = 1514, respectively) were analysed ( Table 1). The mean age was 65.8 ± 9.0 years (range 39-89); duration of diabetes 10 ± 9.1 years (range 1-51); body mass index (BMI) 30.9 ± 5.4 kg/m 2 (range 18.5-55.4). Subjects received dietbased diabetes therapy (n = 120), oral hypoglycaemic agents (OHA; n = 513), a combination of OHA and insulin (n = 439), or insulin alone (n = 490). The CGM was performed in an outpatient setting under daily-life conditions. The quality of CGM profiles was assessed on three subsequent days, and the measures were averaged for analyses. All CGM profiles were assessed using the following parameters: MBG, median glucose level (median), SD, range, MAGE, CONGA over a 6-h period, MODD, interquartile range (IQR), t G and AUC G above or below the target range from 3.9 to 8.9 mmol/l, risk scores for LBGI and HBGI, and GRADE [16][17][18][19][20].

Factor analysis
The factor analysis [29,30] was conducted with the FACTOR procedure available in PASW Statistics 17 (SPSS Inc., Chicago, IL, USA). Initially, all included parameters were normalised using the z-score and the correlation between all variables was determined. The number of components to be retained was first based on a scree plot. A calculation with an additional factor provided a further independent and interpretable factor. The calculation of the Kaiser-Meyer-Olkin (KMO) measure resulted in a KMO of 0.821, which indicated that the factor model was appropriate and the sampling was highly adequate. A varimax (orthogonal) rotation was used to obtain a set of independent, interpretable factors. The resulting factor pattern was interpreted with the use of factor loadings >0.5.

Categorisation of CGM profiles
A randomly selected subset from all CGM profiles (n = 766) was independently categorised into groups of 'very good' , 'good' ,'satisfactory' ,'fair' , and 'poor' metabolic control, by three diabetes specialists. The specialists had access to both the CGM profiles and the patient records that indicated the diabetes type, diabetes duration, and types of therapy associated with each CGM.

Statistical methods
All analyses were performed with PASW Statistics 17. Results are expressed as mean ± SD or as medians and IQR. Analysis of variance was used to assess differences between groups. The strength of the dependence between two continuous variables was assessed with the Pearson's correlation coefficient and between ordinal variables with Kendall's tau-b correlation. The weighted Cohen's kappa score [31] was used to assess the interrater reliability of the categorisation of the CGM profiles between diabetes specialists and between Q-Score and diabetes specialists. The reliability (concordance of assessments) was measured using the method proposed by Landis and Koch [32]. A P-value <0.05 was considered to indicate statistical significance.

Extraction of factors accounting for the quality of CGM profiles
To identify the criteria determining a glucose profile, we performed a factor analysis. Four factors were identified (Table 2), which accounted for 95% of the common variance in the dataset (38% factor 1; 34% factor 2; 20% factor 3; and 3% factor 4). Factor 1 (central tendency and hyperglycaemia) was associated with positive loadings of MBG, median, t G >8.9 , AUC G >8.9 , GRADE, and HBGI. Factor 2 (within-day variability) was associated with positive loadings of range, SD, IQR, MAGE, CONGA 6-h , and MODD. Factor 3 (hypoglycaemia) was associated with positive loadings of t G <3.9 , AUC G <3.9 , and LBGI. Factor 4 (between-day variability) was associated with a positive loading of MODD.

Construction of the Q-Score
Among these factors, the parameters with loadings >0.5 were highly correlated with each other, based on linear regression functions (data not shown). Therefore, one parameter from each factor could be selected for the construction of the Q-Score. The only exception was the analysis of MBG and the time spent above the target range (t G >8.9 ) from factor 1, where a sigmoid function was found (Additional file 1: Figure S1). The time spent above the target range was highly variable for any given MBG. Therefore, these two parameters were selected from factor 1 for the construction of the Q-Score.
From all factors, we selected one parameter that had a high factor loading, was simple to calculate, and was easy to interpret in the context of a CGM curve for general practitioners. These parameters were the MBG and the time spent above the target range from factor 1; the range from factor 2; the time spent below the target range from factor 3; and the MODD from factor 4 (Additional file 1: Figure S2).
In the proposed Q-Score, all parameters are combined to generate a single measure and should, therefore, have equal weight. To achieve equivalence in the parameters for calculations, the five selected parameters with unequal means and variances were standardised with a z-transformation. The Q-Score was computed as the sum of all five standardised variables. This ensured that all five parameters had an equal impact on the Q-Score.
Then, to ensure positive values, we added a constant equal to 8. The formula for the Q-Score was:  Figure S3. The inter-rater reliability between the diabetes specialists using weighted Cohen's kappa [31] was significantly different from pure chance for all diabetes specialists (0.438 ± 0.019, 0.713 ± 0.016, 0.403 ± 0.018; P <0.001 for all). The categorisations were highly correlated among the specialists (Kendall's tau = 0.671, 0.787 and 0.751; P <0.001), allowing us to average the categories for each patient. Scores of the same 766 CGM profiles, which were categorised by the three diabetes specialists were calculated. A box-plot analysis was used to define the limiting Q-Score values for the CGM-categories defined by the diabetes specialists ( Figure 1A). The Q-Scores for the CGM-categories were as follows: <4.0, very good; 4.0-5.9, good; 6.0-8.4 satisfactory; 8.5-11.9 fair; and ≥12.0 poor. These limits were also applied to define the Q-Score categories as very good, good, satisfactory, fair and poor (Additional file 1: Figure S3). The criteria for the Q-Score categories and the description of the Q-Score categories are shown in Figure 1B.

Reliability of the Q-Score categories
The reliability of Q-Score categories was measured using the linear weighted Cohen's kappa coefficient [31] and concordance was assessed using the scale by Landis and Koch [32]. Overall there was a substantial concordance between the assessment of CGM profiles by the diabetes specialists and the defined Q-Score categories (κ: 0.666 ± 0.010). There was substantial concordance between two diabetes specialists in terms of the Q-Score categories (Physician A κ: 0.759 ± 0.015; Physician B κ: 0.724 ± 0.015), while the third diabetes specialist showed moderate concordance (Physician C κ: 0.519 ± 0.018). Complete concordance in the selected Q-Score categories and the assessment by diabetes specialists was achieved for 59.1% of CGM profiles, a deviation of one level in the categorisation (above or below; for example diabetes specialist assessment as 'very good' and a Q-Score of 'good') in 37.4% of CGM profiles and of two levels in 3.5% of CGM profiles (above or below; for example diabetes specialist assessment as 'very good' and a Q-Score of 'satisfactory').

Application of Q-Score in diabetes care
In the study population (n = 1,562), increases in the Q-Scores corresponded to changes in common parameters used to described glycaemic control (P <0.001) (Additional file 1: Table S1). We investigated whether the Q-Score also increased with the complexity of therapy (Table 3). We found that the Q-Score was lowest for subjects treated with diet (5.0 ± 2.4), increased for those treated with OHAs (6.8 ± 3.1) and OHA + insulin (8.7 ± 3.3), and was highest for subjects treated with insulin alone (9.6 ± 3.6) (Figure 2A). The analysis of the Q-Score distributions ( Figure 2B) revealed significantly more fair or poor profiles in subjects treated with insulin alone compared with those treated with diet. Subjects with good and poor metabolic control were present in all treatment groups. However, the percentage of subjects with very good or good Q-Scores was decreased in insulin-treated subjects compared with those treated with diet (17.1% vs. 73.1%). Conversely, the percentage of people with fair or poor profiles was higher in insulin-treated than in diet-treated subjects (58.4 vs. 9.3%; Figure 2B). This was adequately reflected by the corresponding Q-Scores ( Figure 2A, Table 3). Moreover, the Q-Score increased with rising of HbA1c. In subjects with HbA1c <6.5% (n = 531) the Q-Score was 6.23 ± 2.77 (mean ± SD). In subjects with HbA1c 6.5-6.99% (n = 375) the Q-Score was 7.56 ± 2.93 and in subjects with 7.0-7.49% (n = 322) the Q-Score was 8.62 ± 3.17. High Q-Scores (9.96 ± 3.33) were seen in subjects with HbA1c 7.5-7.99% (n = 155) and further elevated in subjects with HbA1c ≥8.0% (n = 179; 11.72 ± 3.68).

Patient-tailored analysis of CGM profiles
The Q-Score enables the identification of profiles with insufficient metabolic control. Aiming for a patienttailored approach, we developed a method allowing the identification of the factors with improvement potential in a given glucose profile. First, we defined the limits of the improvement potential using the 95 th percentile of each Q-Score parameter of profiles categorised by the diabetes specialists as very good and good. Values below the 95 th percentile were defined as 'appropriate' and values above as 'with improvement potential'. Next, the improvement potential was categorised as 'low' , 'moderate' or 'high'. Values above the 95 th percentile of profiles categorised as satisfactory were defined as 'low'  improvement potential. The limits for 'moderate' or 'great' improvement potential were built, with equal class size (Additional file 1: Table S2).
Overall, more than 80% of profiles categorised as fair or poor had at least three factors to optimise (Additional file 1: Figure S4). In particular, subjects with those profiles would benefit from therapy optimisation. The individual improvement potential is demonstrated by combined illustration of the CGM curve, the improvement potential for each Q-Score parameter, and the statistical data for each factor (Figure 3). Three CGM curves with different Q-Scores are provided as examples. The profile #128830 had a satisfactory Q-Score ( Figure 3A). The analysis of the improvement potential indicates a low improvement potential for the time in the hyperglycaemic range. Case 133657 had a fair Q-Score ( Figure 3B). The improvement potential was 'moderate' for t G <3.9 and 'low' for range and MODD, respectively ( Figure 3B). Case 136516 had a poor Q-Score. This profile shows prolonged hyperglycaemic status, resulting in a high improvement potential for t G >8.9 . A high MBG and increased glycaemic variability were also recorded in this subject, therefore a moderate improvement potential was recorded for central tendency, intra-and inter-daily variability ( Figure 3C).

Discussion
There is a need for a metric to assess short-term glycaemic control, similar to the way in which measuring HbA 1c allows the assessment of long-term glycaemic control. We aimed to develop a score-type metric that provides an overall, understandable assessment of the blood glucose profile. Ideally, this measure would be sensitive to chronic hyperglycaemia, glucose variability, and hypoglycaemia, and be applicable in routine diabetes care for the screening of profiles with insufficient metabolic control.
For the development of the metric, we first intended to identify the factors determining the quality of a CGM profile. We performed the first published factor analysis [29,30] of CGM-variables [4][5][6][7]9,10]. By definition, a factor represents a cluster of highly correlated variables [29,30]. We identified four factors that described CGM profiles: hyperglycaemia, inter-and intra-daily variability, and hypoglycaemia. To verify our findings, we analysed the variables with positive loadings within each factor. For factor 1, we found that two variables were necessary to describe hyperglycaemia: MBG, and the time spent in the hyperglycaemic range. Thus, overall, five variables adequately described the CGM: central tendency, hyperglycaemia, intra-daily variability, hypoglycaemia, and inter-daily variability. These are equivalent to the key metrics (target range, glucose exposure, glucose variability, hypoglycaemia, and hyperglycaemia) suggested by an expert panel for the standardisation of glucose reporting, analysis, and clinical decision making [10]. The new metric, which we called the Q-Score (Q = Quality), was constructed with parameters from these factors or key metrics.
Earlier studies have also sought to identify the most useful CGM-parameters for clinical use [6][7][8]33,34]. Rodbard [7] evaluated methods for assessment of glycaemic control and glycaemic variability. Consistent with our findings, Rodbard [7] observed high correlations among MAGE, SD, IQR, and CONGA, and also concluded that these measures provided essentially the same information. The author defined four groups of methods for characterising glucose variability. Three groups contained parameters summarised by our factor analysis in the factor 'intra-day variability'. The fourth group [7], the MODD, was also identified in our study, belonging to factor 'inter-day variability'. In accordance with our findings, hypoglycaemia, hyperglycaemia, and euglycaemia were identified as parameters of glycemic control [7]. Recently, Fabris et al. [35] analysed a pool of 25 glucose variability indices using the Sparse principal component analysis in a study with 17 subjects diagnosed with type 1 diabetes. The authors identified a subset of 10  different glucose variability indices that are sufficient to preserve more than the 60% of the variance originally explained by all 25 variables [35]. CGM is increasingly being introduced into diabetes care [1]; therefore, there is an increasingly urgent requirement for a metric summarising the quality of shortterm glucose control [21][22][23]. Like us, Rawlings et al. [21] aimed to develop an integrated approach that provides a complete and consistent assessment of glycaemic control. However, these authors followed a different approach and published a graphical user interface for evaluation of CGM profiles based on glucose variability metrics [SD, MODD, CONGA(n), and MAGE] and glycaemic statistics (time spent within thresholds, time spent in hyperglycemic/hypoglycemic conditions, area under the curve, and mean glucose). The interface was tested in a small number of subjects with type 1 diabetes. Marling et al. developed a 'consensus perceived glycaemic variability metric' that captures the gestalt perceptions of experienced physicians using a machine learning algorithm [23]. In tests on 250 CGM profiles from subjects with type 1 diabetes, this metric outperformed mean amplitude of glycaemic excursion, standard deviation, distance travelled, and excursion frequency [23]. Thomas et al. [22] also developed a measure for assessing glycaemic control using CGM profiles. In the 'Pentagon' model [22] they included MBG, AUC G >160 mg/dl , t G >160 mg/dl , SD Glucose , and HbA 1c ; thus, that model included three of the factors identified in the present study; hyperglycaemia, central tendency and intra-daily variability. However, in contrast to the Q-Score, the Pentagon model included HbA 1c . In another study, Thomas et al. [36] reported that the Pentagon model was helpful for assessing individual glycaemic profiles of subjects with type 1 diabetes and for assessing the influence of therapeutic interventions. In addition, they showed that model predictions of the risk of developing late complications were more accurate than HbA 1c predictions. Future studies are necessary to compare the Q-Score to the Pentagon model. These studies [21][22][23] aimed to facilitate interpretations of blood glucose profiles and to allow the identification of 'weak points' in diabetes management, which is consistent with our approach in this study. However, these studies focused on type 1 diabetes, whereas we tested the Q-Score in a large set of CGM profiles obtained from subjects with type 1 and type 2 diabetes. It should be noted that the relatively small number of subjects with type 1 diabetes represents a limitation of our study.
To develop a practical, readily interpreted metric for routine clinical use and screening, we intended that the Q-Score should allow categorisation of glycaemic control from very good to poor. To achieve categorisation in our study, three diabetes specialists independently classified CGM profiles into groups of very good, good, satisfactory, fair, and poor metabolic control. The majority of profiles were derived from subjects with type 2 diabetes. As expected, the results reflected subjective evaluations. However, the high Kendall's tau correlation indicated that the results were consistent. The groups of evaluated CGM profiles were used to define the Q-Score limits between categories of very good, good, satisfactory, fair, and poor glycaemic control. In addition, we conducted a proof-ofprinciple study to show that patients with diabetes could be stratified for treatment based on the Q-Score using the CGM profiles of 1,562 historical subjects. The category of good glycaemic control included the majority of subjects treated with diet, half of those treated with OHA, only a quarter of those treated with OHA + insulin, and less than 20% of those treated with insulin alone. These findings are in accordance with other studies that demonstrate that blood glucose profiles are worsened with increasing therapy complexity (from diet alone to insulin) [25][26][27]37]. Thus, the categorical Q-Score allows the identification of subjects with poor metabolic control who require therapy optimisation. We intend to verify the Q-Score in a largerscale study that includes CGM profiles derived from hospitalised subjects and those with diabetes and coexistent chronic illness.
Patient-tailored diabetes therapy represents the state-of the art in diabetes care and management [38,39]. Therefore, in addition to providing a general assessment of glycaemia, we demonstrate a method for identification of Q-Score parameters that require therapeutic attention and would provide a basis for personalised diabetes therapy. Profiles of people with diabetes categorised as very good or good were used to set the limit for the improvement potential of all Q-Score parameters. This method reveals the parameters that require therapeutic action; for example, the adjustment of the insulin therapy in the case of hypoglycaemia.
Submit your next manuscript to BioMed Central and take full advantage of: