590154
research-article2015
DSTXXX10.1177/1932296815590154Journal of Diabetes Science and TechnologySchmelzeisen-Redeker et al
Original Article
Time Delay of CGM Sensors: Relevance, Causes, and Countermeasures
Journal of Diabetes Science and Technology 2015, Vol. 9(5) 1006–1015 © 2015 Diabetes Technology Society Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1932296815590154 dst.sagepub.com
Günther Schmelzeisen-Redeker, PhD1, Michael Schoemaker, PhD1, Harald Kirchsteiger, PhD2, Guido Freckmann, MD3, Lutz Heinemann, PhD4, and Luigi del Re, PhD2
Abstract Background: Continuous glucose monitoring (CGM) is a powerful tool to support the optimization of glucose control of patients with diabetes. However, CGM systems measure glucose in interstitial fluid but not in blood. Rapid changes in one compartment are not accompanied by similar changes in the other, but follow with some delay. Such time delays hamper detection of, for example, hypoglycemic events. Our aim is to discuss the causes and extent of time delays and approaches to compensate for these. Methods: CGM data were obtained in a clinical study with 37 patients with a prototype glucose sensor. The study was divided into 5 phases over 2 years. In all, 8 patients participated in 2 phases separated by 8 months. A total number of 108 CGM data sets including raw signals were used for data analysis and were processed by statistical methods to obtain estimates of the time delay. Results: Overall mean (SD) time delay of the raw signals with respect to blood glucose was 9.5 (3.7) min, median was 9 min (interquartile range 4 min). Analysis of time delays observed in the same patients separated by 8 months suggests a patient dependent delay. No significant correlation was observed between delay and anamnestic or anthropometric data. The use of a prediction algorithm reduced the delay by 4 minutes on average. Conclusions: Prediction algorithms should be used to provide real-time CGM readings more consistent with simultaneous measurements by SMBG. Patient specificity may play an important role in improving prediction quality. Keywords continuous glucose monitoring, CGM, MARD, accuracy, precision, performance evaluation, performance comparison, time delay Continuous glucose monitoring (CGM) offers people with diabetes the opportunity to improve their blood glucose (BG) control by monitoring glucose levels, even at times when the patient would not do this, for example, during the night. This allows detection of clinically relevant issues that might not be noticed with conventional capillary measurements of BG. CGM measurements also provide trend information based on past measurement results; this allows estimating the change in glycemia for a short period of time in the future. In addition, CGM systems help avoid hypoglycemic events by providing alarms when a predefined glucose threshold is exceeded or even prior to this if a pending hypoglycemia is predicted.1,2 One reason why CGM data should not be used for determining insulin doses by patients with diabetes, according to the officially approved guidelines set by the regulatory authorities,
is that CGM readings often do not correspond exactly to self monitoring of blood glucose (SMBG) measurement results taken at the same time—especially during fast changes in BG. CGM systems have exhibited an impressive progress in their analytical performance from 1 generation to the next. Nevertheless, 1 1
Roche Diabetes Care GmbH, Mannheim, Germany Institute for Design and Control of Mechatronical Systems, Johannes Kepler University, Linz, Austria 3 Institute for Diabetes-Technology GmbH at Ulm University, Germany 4 Science & Co, Düsseldorf, Germany 2
Corresponding Author: Günther Schmelzeisen-Redeker, PhD, Roche Diabetes Care GmbH, Sandhofer Str 116, 68305 Mannheim, Germany. Email: [email protected]
1007
Schmelzeisen-Redeker et al important reason for differences between CGM and SMBG measurement results is the time delay. The aim of this article is to characterize the time delay observed with a novel CGM system and measures to reduce this delay.
CGM systems measure glucose levels in interstitial fluid (ISF) in the subcutaneous adipose tissue and not directly in blood. Once glucose levels change in either compartment there is a difference in measurement results between these 2 compartments before they equalize after some minutes again. This, in turn, results in the well-known fact that results of CGM measurements are shifted in time compared to results of SMBG measurements during rapid swings in glycemia. This delay is the consequence of the process of glucose diffusion across the walls of capillary vessels and through the interstitial space to the sensor. This process requires some time, and the delay can be observed during both rising and decreasing BG values, probably with varying impact.3-5 It is commonly understood that the delay and the rate of change of glucose in the ISF are affected by several physiological factors, for example, by the local blood flow, the local tissue perfusion, the permeability of the ISF, the local convection in the ISF, and so on. The average physiological time delay (ISF glucose to BG) is assumed to be in the range of 5 to 10 minutes. It is of interest to note that the differences in time delays reported for plasma BG and ISF are not conclusive.6 A more recent estimation based on glucose tracer experiments with healthy humans using a microdialysis technique resulted in an average physiologic time delay of 5-6 minutes.7 Other estimates using the same approach for a very small study population report a similar value for patients with type 1 diabetes.8 There is strong evidence for a great variability of the time delay between individual subjects even under reproducible experimental conditions and independently of a specific CGM system.
Technologically Caused Time Delay The measurement of glucose by the CGM system and the associated calculations inside the CGM system before the measurement result is displayed to the patient might add an additional, but usually short, physical time delay. The measurement delay is caused by both the diffusion of glucose through the protecting membranes and/or glucose flux limiting membranes covering the CGM sensor as well as the speed of glucose reactions inside the sensor. This fraction of time delay is sensor-specific, and little has been published thus far about the physical time delay of different CGM sensors.9 However, this delay is assumed to be within the range of a few minutes. The amperometric sensors measure an electric current that is proportional to the glucose levels in the ISF. This current must be transferred into a glucose reading by proper
raw MA30 MA60
30 25
el. current [nA]
Physiologically Caused Time Delay
35
20 15 10 5 0 2500
2600
2700
2800
2900 3000 time [min]
3100
3200
3300
Figure 1. Raw CGM signal (blue) filtered using a 30-minute (red) and 60-minute (green) moving average filter.
calibration using a conventionally measured capillary BG value. The sensor signal is also subject to a number of disturbances which are mainly slowly varying and random. Slowly varying signal disturbances are caused by biofouling10 and are removed and/or reduced by calibration until the measured signal becomes overly deteriorated and the sensor needs to be replaced. The random noise superimposed over the raw data is reduced by averaging signals over a period of time (usually several minutes), and only these mean values are usually reported. Filtering is another possibility for reducing the noise of the sensor signal and use of additional information increases the speed of this compensation.11,12 However, mathematical filters applied in real time to provide reliable glucose information unavoidably generate an algorithmic time delay (since real time filters can only utilize sensor signals from the past; stronger filtering requires sensor signals from a longer time period in the past, which thereby leads to longer time delays).13 In addition, algorithms are employed to smooth the glucose profile over time and to reduce and/or eliminate artifacts (socalled smart sensors).14,15 Here again only little information is available about the impact of all these time-consuming factors on the overall time delay of commercial CGM systems; thus far, values of 3 to 12 minutes have been reported.9,16 The trade-off between variance of the sensor readings and time delay introduced by filtering can be appreciated in a simulation example shown in Figure 1, where the original raw data were filtered using a moving average filter with 2 different settings for the filter length (equivalent to 30 and 60 minutes). It is clear that longer averaging (MA60, green) leads to less noisy signals, but increases the physical time delay considerably. Using large filter lengths with such simple filters is not realistic as the BG level keeps changing. In modern CGM
1008
Total Time Delay As the resulting total time delay potentially affects the ability of patients to encounter a clinically relevant situation, knowledge of this time delay and measures to reduce it are of clinical relevance. To date, the focus of reports about the characteristics of CGM systems has been on their analytical performance and not on other factors such as the time delay of the respective system. The number of publications in which data about the time delays observed with different CGM systems are reported, favorably from head-to-headcomparisons, is quite limited.17,18 For patients with diabetes who use a given CGM system, the combined effect of time delays and measurement errors is of concern as both factors contribute to the measurement differences they observe between the CGM data and BG measurements performed at the same time. As the accuracy of CGM systems has improved significantly over the years, other sources of errors, such as time delays, become more relevant. In the literature a considerable range of overall time delays—between 5 to 40 minutes9,16—has been reported. It is not clear to which extent the reported differences are due to the different CGM systems used in such studies and/or to the experimental conditions used to evaluate this delay. In these publications, the extent of interindividual differences (= differences between patients) or intraindividual differences (= within the same patient over a longer time horizon) with a given CGM system is most often not reported, partially because of the lack of sufficiently long CGM recordings. However, such differences might be substantial (see the Results section). Also, little is known about the differences between the glycemic ranges with respect to the delay; that is, is the delay of the same magnitude in the low and high glycemic range? Also other patient-specific effects might influence the delay on a day-to-day basis (eg, physical exercise and wear time of the CGM system). In other words, not much is known either about the variability of the time delay between patients or the variability over time in the same patient.
Effect of Time Delay on Sensor Performance The following example demonstrates the effect of a time delay on CGM sensor performance. Figure 2 shows a CGM profile (from Zschornack et al19) of a patient with type 1 diabetes together with the respective SMBG values to visualize the magnitude of the time delay. The measure most often used to characterize the measurement performance of CGM systems is the mean absolute relative difference (MARD) (see the Methods section below for a formal definition). For single paired measurements in time, the absolute relative difference (ARD) can be computed. ARD
400 300
[mg/dL]
systems more complex filters are used that have better performance characteristics than moving average, but exhibit the same basic trade-off.
Journal of Diabetes Science and Technology 9(5)
CGM
SMBG
200 100 0
1500
2000
2500
3000
3500
3000
3500
80 60
abs.error (ARD)
mean=14.2
40 20 0
1500
2000
2500 time [min]
Figure 2. CGM recordings and results of glucose measurements in capillary blood (SMBG) (upper panel). Absolute relative difference between CGM and SMBG values and the mean thereof (lower panel).
is shown as the blue line in the lower panel in Figure 2, together with the MARD (red line), which is 14.3% in this case. Assuming that the CGM measurement itself is perfect (has an analytical error of 0%), and also the BG measurement has no analytical error, all differences in the measurement results have to be attributed to the time delay. One option to evaluate the impact of the time delay on the MARD is to shift the 2 time series against each other; so shifting the artificial SMBG values in relation to the CGM measurements by, for example, 10 minutes results in a MARD of 9.5% (Figure 3). This indicates that although there is absolutely no error at all in the measurements per se, the time delay has a high impact on the performance characteristics. A related analysis of the effect of time delays on the CGM performance was also presented in Scuffi et al.6 To separate the effects of time delays from that of analytical error, 1 guideline for the evaluation of CGM systems issued by the Clinical and Laboratory Standard Institute (CLSI)20 suggests procedures to retrospectively remove the time delay from the measurements before evaluating the accuracy. This yields a figure which is closer to the “real” analytical accuracy of the given CGM system. This value should be, to some extent, independent of the specific user. Other such procedures aiming for the same goal have been proposed; however, their major disadvantage is that they are based on postprocessing of measured data and do not provide any help with real-time data.21 This would, however, be necessary for clinical practice.
Measures to Reduce the Impact of the Time Delay Against this background, the question arises as to what measures can be taken to reduce the effect of the time delay on CGM recordings. Roughly speaking, there are 2 main possibilities to deal with a time delay:
1009
Schmelzeisen-Redeker et al
300
400 CGM
SMBG
200
200
100 0
1500
2000
2500
3000
150
3500
100 50
60 abs.error (ARD)
mean=9.5
0 5200
20
1500
2000
2500 time [min]
3000
3500
Figure 3. Absolute relative difference and mean value thereof when assuming no analytical error, only an overall time delay between CGM and SMBG, using the same data as presented in Figure 2.
1. ISF glucose levels are closer to those in the brain than the results of BG measurements.22 As the brain is the clinically relevant organ (glucose levels in the brain are critical for preventing, eg, hypoglycemic symptoms), it represents the ultimate target for blood glucose control. In the long run, it might therefore be possible to switch the target of glycemic control from BG values measured in peripheral blood to values obtained from ISF glucose. The usage of ISF data for clinical decisions would clearly require a paradigm shift. 2. Using mathematical prediction algorithms to prospectively compensate for the time delay would allow the user to receive a CGM based glucose value that is closer (in the time domain) to the SMBG results. Making use of prediction algorithms allows forecasting future glucose values (eg, those that can be expected in the next minutes) based on the information available at the current time and based on a mathematical model. Predictors typically perform well for short-term prediction, that is, up to about 20 minutes. Several algorithms have been published to achieve this goal (see, eg, Naumova et al,23 Kirchsteiger et al,24 Pappada et al,25 Gani et al26). Below, we have highlighted the idea by using a simple autoregressive algorithm27 of the fifth order applied to CGM data, which enables a relatively good prediction for a 10minute horizon (Figure 4). This demonstrates that for short prediction horizons, the accuracy of the prediction is good; however, for obvious reasons this deteriorates rapidly for longer time periods, see the larger errors in the bottom panel of Figure 4 for the 20-minute prediction. Several predicted traces are shown in Figure 5, where the red cross signs indicate predicted values with a 5-minute interval, that is, the
error [mg/dL]
40
0
cgm 10min prediction 20min prediction
250
[mg/dL]
[mg/dL]
300
5300
5400
5500
5600 5700 5800 time [min]
5900
6000
6100
6200
5300
5400
5500
5600
5900
6000
6100
6200
50 0 -50 -100 5200
5700
5800
Figure 4. CGM data and 10-, 20- minute prediction of an autoregressive model of fifth order (top plot) and the prediction errors (bottom plot). Figure shows validation data; the model was trained on the first 2 days of the data set.
300
250
200
150
10-min 20-min 30-min prediction Detail at t=5700
100
50
0 5200
5300
5400
5500
5600 5700 5800 time [min]
5900
6000
6100
6200
Figure 5. Predicted traces shown at some selected points in time for the same data as in Figure 4. For a given point, the red lines/crosses indicate the future values expected after 10/20/30 minutes using the information available until then.
trace starts at the current time at the current CGM value (blue line), and then shows the 5-, 10-, . . . 30-minute prediction. This kind of prediction exclusively uses past information, which means that any measured change happening between the current time the prediction is computed and the future estimate is not taken in account. For instance, a meal between the current time and the end of the prediction horizon will raise the real BG value but cannot be considered by the predictor. Against this background, it is of the utmost importance to develop safe predictors using as much additional information as possible (including meal information, insulin administrations, activity, etc). In addition, a prediction with
1010
Journal of Diabetes Science and Technology 9(5)
Table 1. Descriptive Characteristics of the Study Participants (20 Male, 19 Female).
Age (years) Height (cm) Weight (kg) BMI (kg/m²) Skinfold thickness (mm) Duration of diabetes (years) HbA1c (%)
Mean (SD)
Range
47 (11) 175 (10) 82 (16) 27 (4) 27 (7) 23 (12) 7.8 (1.2)
21-63 156-197 55-116 21.6-41.4 14-39 2-48 5.8-11.7
an incorrect time horizon could be more detrimental than useful. Patient-specific time delays—if it can be confirmed that they exist—can be used as prediction horizon to personalize the CGM sensor and compensate the delay.
Clinical Data To evaluate the different factors that have an impact on the time delay, CGM data were analyzed that were obtained while running a clinical study with patients with type 1 diabetes using a prototype CGM sensor. The design28 and performance19 of this CGM sensor has been reported earlier. The clinical– experimental study was performed in a Clinical Research Center (Institut für Diabetes-Technologie Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm [IDT]) and was approved by an Ethical Committee and the competent authority (Bundesinstitut für Arzneimittel und Medizinprodukte, Germany). The study was performed over 2 years and consisted of 5 consecutive phases. The study design of each phase followed essentially the recommendation of a guideline for CGM evaluation POCTO5A.20 This includes induction of a significant glucose swing into the high and low glucose range by late and slightly overdosed insulin delivery after breakfast on 2 of the 7 days of each study phase. Patients simultaneously wore 2 or 4 CGM sensors. Capillary blood samples were routinely measured in duplicate once per hour during the day and at 3 am at night. For all analyses the mean of the 2 readings was used as the reference value. Plasma glucose levels in the blood samples were measured using a standard home BG meter (Accu-Chek® Aviva BG meter, Roche Diagnostics GmbH). Glucose levels were measured more often in phases with induced glucose swings (15-minute intervals). Details of the study design have been reported elsewhere.19 A total of 37 patients with type 1 diabetes participated in the clinical study (see Table 1 for patient characteristics). Eight of these patients participated in 2 different study phases separated by at least 8 months. The total number of CGM sensor data sets that could be used for the time delay analysis was 108 (35/10 participations with 2/4 sensors in parallel; sensor data of 1 participation excluded because of failure of 1 of the 2 sensors). Data analysis was performed using raw
signals, that is, the values measured directly by the sensor (current in nano Ampere [nA]), processing was done by standard statistical methods. Part of the data analysis was done with retrospectively calibrated sensor signals (in mg/dl but without using a mathematical filter).19
Methods for Time Delay Estimation In the analysis presented in this article, the focus is on the part of the time delay that results from reasons of physiology and sensor design, and not on the algorithmic treatment of the signals. Thus, the distribution of the time delay of the raw electrical current versus BG of the individual CGM sensors during the different study phases was mainly examined. The quality of time delay estimation depends on a number of factors, among them (1) frequency and accuracy of BG measurements; (2) time resolution and quality of sensor raw signals (low noise, no artifacts); (3) extent and rate of change of glucose swings; (4) stability of time lag over time of experiment. The low noise and 1-minute measurement frequency of the sensor, the use of the mean of duplicate BG values, the high frequency, and large range of BG values during induced glucose swings take these limiting factors into account.
Correlation Coefficient Maximization The correlation between 2 stochastic signals x1 (t ) , x2 (t ) is given by the correlation function
φ (τ ) = lim
N →∞
1 N
N
∑x ( k ) x (k + τ ) 1
2
k =1
Intuitively, this function describes the dependency of the 2 signals when shifted against each other in time by τ minutes. A small value of the function indicates less dependency while a high value indicates high dependency (ie, the 2 signals are more similar to each other). In our application, we consider CGM data as the first signal and SMBG measurement results as the second one. Then, we compute the correlation function as defined above forvarious values of τ by retrospectively shifting the CGM trace in time. The value τ* that results in the highest value of the correlation function is the estimated time delay. This method was used with the raw glucose sensor signals.
MARD Minimization The ARD between a CGM data point yCGM (t ) and a reference measurement yref (t ) is given by ARD ( tk ) = 100% and the mean ARD as
yCGM ( tk ) − yref (tk ) yref (tk )
1011
Schmelzeisen-Redeker et al
Table 2. Estimated Time Delay (Minutes) Using the 2 Different Methods Described in the Methods Section.
22 20
Time delay [min]
18
Correlation coefficient
MARD
9.5 3.7 9 4 3 20
8.0 3.4 8 4 2 17
Mean SD Median IQR Minimum Maximum
16 14 12 10 8 6 4 2
Phase 1
Phase 2
Phase 3
Phase 4
Phase 5
Table 3. Estimated Time Delay (Minutes) When Using a Prediction Algorithm for the Same Data Shown in Table 2.
Study Phase
Correlation coefficient, prediction
Figure 6. Box plot of overall time delay distribution during different phases of the clinical study. Squares: medians; rectangles: 25% and 75% quantiles; extensions from rectangles: minimums and maximums; circles: outliers.
MARD =
1 N
N
∑ARD
Mean SD Median IQR Minimum Maximum
5.9 3.7 5 3 –1 19
MARD, prediction 4.3 3.0 4 4.25 –1 12
k
k =1
Similar to the previous method, the MARD can be computed also when first shifting the CGM time series relative to the available reference measurements by τ minutes. By doing this systematically minute by minute, we receive a function MARD(τ), and the minimum of this function determines the estimated time delay. This method was used with the calibrated glucose sensor signals.
Prediction Algorithm The algorithm utilizes a predetermined population average time delay (Δt) and a Kalman filter for real time estimation of the sensor raw signal at time t (It) and its rate of change. Both variables are used for a linear extrapolation (prediction) of the sensor raw signal Δt minutes later (It+ Δt). Instead of It the predicted sensor raw signal is used for calibration and conversion of raw signals into sensor glucose values at time t. In case of the prototype sensor Δt was 8 minutes.
Results Time Delay An analysis of the CGM data sets with respect to differences in the time delays in the different study phases showed no significant differences in the 5 different study phases (ANOVA analysis; P = .55; Figure 6). This allows for the pooling of the data from all phases for statistical analysis. Based on the maximization of the correlation coefficient of the sensor raw signal with respect to BG, the mean (standard deviation [SD]) overall time delay was 9.5 (3.7) minutes. The
median time interval was 9 minutes and the interquartile range (IQR: difference between the 25% lowest and 75% highest values) 4 minutes (Table 2). The estimated time delay based on the minimization of the MARD was on average 1 to 2 minutes smaller (P < .001). The use of a prediction algorithm, as conceptually introduced in the introduction section, led to a mean time delay reduction of the overall time delay—as seen by the patient— by about 4 minutes (Table 3).
Intraindividual Time Delay Sixty-eight CGM data sets were obtained in study phases with patients simultaneously wearing 2 CGM sensors (sensor A and sensor B). The correlation of the overall time delays observed with the 2 sensors in the same patient is given in Figure 7. The coefficient of determination (corrected R2) was .51. This means that 51% of the variability of the time delay results from the patients. In 1 study phase, patients wore 4 CGM sensors simultaneously (10 patients, 40 data sets). The box plot of the time delays of these patients when grouped by the individual patient confirms that the patient per se has a significant effect on the overall time delay (Figure 8; ANOVA P < .01); the corrected R2 was .43, which means that 43% of the time delay variability is explained by the patient effect. Eight patients who participated in 2 study phases were analyzed separately for each study phase (phases 1 and 2), and the estimated delays plotted versus each other are given in Figure 9. The good correlation of the mean time delays of the 2 study phases indicates that the time delay of a given patient is rather constant over 8 months.
1012
Journal of Diabetes Science and Technology 9(5) 18
20 18
16
mean time delay 2 [min]
time delay sensor A [min]
16 14 12 10 8
14
12
10
8
6
6
4 2
2
4
6
8
10
12
14
16
18
20
22
4
4
6
8
10
12
14
16
18
mean Time delay 1 [min]
time delay sensor B [min]
Figure 7. Correlation of the time delays observed in 2 simultaneously worn CGM sensors in the same patient (sensors A and B); red line: linear regression line; each circle represents data from 1 patient.
Figure 9. Correlation of mean time delays of the CGM sensors of patients participating in 2 different study phases (mean time delay 1 and 2); time between study phases was 8 to 12 months.
Discussion 20 18
TimeDelay [min]
16 14 12 10 8 6 4
63
54
61
57
59
51
52
53
62
64
Patient-ID
Figure 8. Box plot of time delay distribution of CGM sensors for 10 patients wearing 4 sensors simultaneously. Squares: medians; rectangles: 25% and 75% quantiles; extensions from rectangles: minimums and maximums; circles: outliers.
Since the results suggest that patient-specific time delays exist, an analysis of the impact of descriptive characteristics (anamnestic and anthropometric data) of these patients on the time delays was performed. The scatter plots of the mean time delays versus all parameters evaluated (weight, height, body mass index, skinfold thickness, HbA1c, years with diabetes) did not show any significant correlation (Figure 10).
Assessment of the accuracy of CGM systems can be done from 2 points of view: either by comparing the (analytical) performance of different CGM systems in a head-to-head experimental situation or by assessing the accuracy as experienced by the patient when comparing CGM with SMBG. In the first case, time delays should be removed from data as suggested by POCT-05A,20 to allow for fair comparison focusing on the analytical performance only. In the second case, the time delay should not be removed, as the patient would not be able to distinguish between measurement errors resulting from poor accuracy or time delay. Time delays can be estimated reliably only during glucose swings; the larger the glucose swings the better. During such swings, BG readings should be obtained as often as possible with a reliable measurement method to enable precise estimation of the time delay. In addition, the BG measurements should be done at least in duplicate to further reduce the BG measurement error. Several factors affect the quality of the time delay estimation. In particular, artifacts such as dropouts could negatively affect time delay estimation as it is difficult to correct the raw signals for such artifacts. The influence of sensor noise can be limited because it can be reduced for the analysis without affecting the delay estimation. It has been reported17 that the time delay is longer during phases of rising rather than during declining BG levels. A time delay analysis taking both phases into account, as it was done here, reflects an intermediate value for the overall time delay.
1013
18
18
16
16
Mean Time Delay [min]
Mean Time Delay [min]
Schmelzeisen-Redeker et al
14 12 10 8 6
14 12 10 8 6
4 150
155
160
165
170
175
180
185
190
195
4 50
200
60
70
80
18
18
16
16
14 12 10 8 6
110
120
14 12 10 8
22
24
26
28
30
32
34
36
38
40
42
4 12
44
14
16
18
20
BMI
22
24
26
28
30
32
34
36
38
40
Skinfold Thickness [mm]
18
18
16
16
Mean Time Delay [min]
Mean Time Delay [min]
100
6
4 20
14 12 10 8 6 4
90
Weight [kg]
Mean Time Delay [min]
Mean Time Delay [min]
Height [cm]
14 12 10 8 6
0
10
20
30
40
Years with Diabetes
50
4
5
6
7
8
9
10
11
12
HBA1c [%]
Figure 10. Correlation of mean time delay including patient height (a), weight (b), BMI (c), skinfold thickness (d), duration of diabetes (e), and HbA1c (f).
These effects can lead to errors in the time delay estimation of an individual CGM sensor. To achieve reliable results on the average time delay of a given CGM sensor type the number of sensors tested in the clinical setting must be sufficiently high. For quality assurance of the time delay estimation a visual inspection of the time course of blood and sensor glucose should be conducted.
Patient-Specific Effects A pooling of time delay data of different study phases is only allowed if there is no difference in average sensor time delays between these phases. The statistical analysis shows that this assumption is fulfilled for the data set used.
1014 Despite the limitations of the accuracy and precision of SMBG measurement results, both of which contribute to the variability of time delays, the correlation of time delays of sensors working simultaneously in the same patient is evident in Figure 8. This indicates that the physiologic time delay varies between patients to a significant extent; approximately 50% of the time delay variability can be explained by the patient effect. In addition, the analysis of the mean time delays of patients who participated in 2 study phases, separated by at least 8 months, indicates that the average time delay of a patient is rather constant at least over this time period. However, the relatively small database of only 8 patients asks for more data to confirm these results. The existence of a patient-specific time delay might be useful for further advances of algorithms to improve the quality of CGM systems. The overall time delay of a given CGM sensor is composed of the physiologic, the sensor-specific physical and the algorithmic time delays. The physical time delay and the variability of the CGM sensors is also determined (to a relatively small extent) by the production processes and the materials used for sensor production. These were identical for all 108 sensors on which this time delay analysis is based. The described methods of time delay estimation for the raw and the retrospectively calibrated sensor signals do not introduce an algorithmic time delay. Therefore, the measured patient-specific effect on time delay can be attributed to differences in the physiologic time delay only. Of particular interest is the root cause for the differences in physiologic time delay between patients. One hypothesis is that the length of transport path of glucose from the capillary blood vessels in adipose tissue to the sensor surface varies from patient to patient due to different sizes of the fat cell diameter. However, no correlation to patient parameters like weight, BMI, or skinfold thickness could be identified in our analysis. Other root causes may be different vascularization and capillary density of adipose tissue, differences in adipose tissue blood flow affecting the speed of blood exchange in the tissue (similar to the Alternative Site Testing phenomenon where a time delay exists between blood obtained from the forearm and blood obtained from the fingertips), and so on. Further research is needed to test these and other possible hypotheses on the origin of the patient-specific time delay.
Consequences for Medical Use Cases The existence of a patient-specific time delay which might vary from a few to up to 20 minutes must be taken into consideration when a patient is to use a CGM sensor. The impact of the time delay depends on the area of use of the CGM system (when it is used for data-logging and retrospective data analysis, the delay is not as relevant as for real-time decision making). All time-critical decisions (hypo- and hyperglycemia warnings, input for artificial pancreas systems, etc)
Journal of Diabetes Science and Technology 9(5) suffer from a high time delay of a given patient. For such patients CGM performance parameters such as MARD would be rather poor even if the sensor would monitor ISF glucose accurately. The effect of high time delays on the quality of diagnostic or therapeutic systems must be carefully and specifically evaluated for the respective application. The variability of the time delay over sensor wear time is moderate, particularly when taking the accuracy of time delay measurement into account. However, the extreme examples of time delay variation over sensor wear time show that the quality of, for example, prediction algorithms for hypo- and hyperglycemia detection and warning might vary over time even if the quality of the sensor signal (low noise, no artifacts) were constant.
Conclusions The time delay of CGM sensors is composed by an intrinsic physiologic property of the approach and by the technology used for CGM measurement. The time delay represents an important factor that must be taken into account for optimal usage of CGM systems. Further technological improvement of CGM sensors will lead to a lower noise (and better analytical performance) as well as in faster internal glucose transport, both of which will reduce the physical time delay. Glucose prediction techniques are the only available approach that allows real-time compensation of time delays in general. However, a precise estimation of the total time delay, preferably patient-specific, is required for good prediction. There is growing evidence that at least part of the time delay is patient-specific. If confirmed, this would enable usage of learning techniques to improve the quality and reliability of glucose prediction. Abbreviations ARD, absolute relative difference; BG, blood glucose; CGM, continuous glucose monitoring; CLSI, Clinical and Laboratory Standard Institute; IDT, Institut für Diabetes-Technologie Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm; IQR, interquartile range; ISF, interstitial fluid; MARD, mean absolute relative difference; SD, standard deviation; SMBG, self-monitoring of blood glucose.
Declaration of Conflicting Interests The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: GS and MS are full-time employees of Roche Diagnostics. GF is general manager of Institut für Diabetes-Technologie Forschungsund Entwicklungsgesellschaft mbH an der Universität Ulm, Ulm, Germany, which carries out studies evaluating BG meters and medical devices for diabetes therapy on behalf of various companies, and received speakers’ honoraria or consulting fees from Abbott, Bayer, Berlin-Chemie, Becton-Dickinson, Dexcom, Menarini Diagnostics, Roche Diagnostics, Sanofi, and Ypsomed. LH is a consultant for a number of companies developing new diagnostic and therapeutic options for diabetes treatment and is a member of a Sanofi advisory
Schmelzeisen-Redeker et al board for biosimilar insulins. He is a partner of Profil Institute for Clinical Research, US and Profil Institut für Stoffwechselkrankheiten, Germany.
Funding The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was (partially) funded by Roche Diagnostics.
References 1. Palerm CC, Bequette BW. Hypoglycemia detection and prediction using continuous glucose monitoring—a study on hypoglycemic clamp data. J Diabetes Sci Technol. 2007;1:624-629. 2. Vashist SK. Continuous glucose monitoring systems: a review. Diagnostics. 2013;3:385-412. 3. Kulcu E, Tamada JA, Reach G, Potts RO, Lesho MJ. Physiological differences between interstitial glucose and blood glucose measured in human subjects. Diabetes Care. 2003;26:2405-2419. 4. Ward WK, Engle JM, Branigan D, Youssef JE, Massoud RG, Castle JR. The effect of rising vs. falling glucose level on amperometric glucose sensor lag and accuracy in type 1 diabetes. Diabet Med. 2012;29:1067-1073. 5. Wentholt IME, Hart AAM, Hoekstra JBL, Devries JH. Relationship between interstitial and blood glucose in type 1 diabetes patients: delay and the push-pull phenomenon revisited. Diabetes Technol Ther. 2007;9:169-175. 6. Scuffi C, Lucarelli F, Valgimigli F. Minimizing the impact of time lag variability on accuracy evaluation of continuous glucose monitoring systems. J Diabetes Sci Technol. 2012;6: 1383-1391. 7. Basu A, Dube S, Slama M, et al. Time lag of glucose from intravascular to interstitial compartment in humans. Diabetes. 2013;62:4083-4087. 8. Basu A, Dube S, Veettil S, et al. Time lag of glucose from intravascular to interstitial compartment in type 1 diabetes. J Diabetes Sci Technol. 2015;9:63-68. 9. Keenan DB, Mastrototaro JJ, Voskanyan G, Steil GM. Delays in minimally invasive continuous glucose monitoring devices: a review of current technology. J Diabetes Sci Technol. 2009;3:1207-1214. 10. Klueh U, Liu Z, Feldman B, et al. Metabolic biofouling of glucose sensors in vivo: role of tissue microhemorrhages. J Diabetes Sci Technol. 2011;5:583-595. 11. Bequette BW. Continuous glucose monitoring: real-time algorithms for calibration, filtering and alarms. J Diabetes Sci Technol. 2010;4:404-418. 12. Bequette BW. Optimal estimation applications to continuous glucose monitoring. Proc Am Control Conf. 2004;958-962.
1015 13. Oppenheim AV, Schafer RW, Oppenheim AV, eds. DiscreteTime Signal Processing. Boston, MA: Pearson; 2009. 14. Del Favero S, Facchinetti A, Sparacino G, Cobelli C. Improving accuracy and precision of glucose sensor profiles: retrospective fitting by constrained deconvolution. IEEE Trans Biomed Eng. 2014;61:1044-1053. 15. Facchinetti A, Sparacino G, Cobelli C. Online denoising method to handle intraindividual variability of signal-to-noise ratio in continuous glucose monitoring. IEEE Trans Biomed Eng. 2011;58:2664 -2671. 16. Rebrin K, Sheppard NF, Steil GM. Use of subcutaneous interstitial fluid glucose to estimate blood glucose: revisiting delay and sensor offset. J Diabetes Sci Technol. 2010;4:1087-1098. 17. Boyne MS, Silver DM, Kaplan J, Saudek CD. Timing of changes in interstitial and venous blood glucose measured with a continuous subcutaneous glucose sensor. Diabetes. 2003;52:2790-2794. 18. Kovatchev BP, Shields D, Breton M. Graphical and numerical evaluation of continuous glucose sensing time lag. Diabetes Technol Ther. 2009;11:139-143. 19. Zschornack E, Schmid C, Pleus S, et al. Evaluation of the performance of a novel system for continuous glucose monitoring. J Diabetes Sci Technol. 2013;7:815-823. 20. CLSI. Performance metrics for continuous interstitial glucose monitoring; approved guideline. POCT05-A. Wayne, PA: Clinical and Laboratory Standards Institute; 2008. 21. Carter G. Coherence and time delay estimation. Proc IEEE. 1987;75:236-255. 22. Nielsen JK, Djurhuus CB, Gravholt CH, et al. Continuous glucose monitoring in interstitial subcutaneous adipose tissue and skeletal muscle reflects excursions in cerebral cortex. Diabetes. 2005;54:1635-1639. 23. Naumova V, Pereverzyev SV, Sampath S. A meta-learning approach to the regulatrized learning—case study: blood glucose prediction. Neural Netw. 2012;33:181-193. 24. Kirchsteiger H, Johansson R, Renard E, Re LD. Continuoustime interval model identification of blood glucose dynamics for type 1 diabetes. Int J Control. 2014;87:1454-1466. 25. Pappada SM, Cameron BD, Rosman PM. Development of a neural network for prediction of glucose concentration in type 1 diabetes patients. J Diabetes Sci Technol. 2008;2:792-801. 26. Gani A, Gribok A, Lu Y, Ward W, Vigersky R, Reifman J. Universal glucose models for predicting subcutaneous glucose concentration in humans. IEEE Trans Info Tech Biomed. 2010;14:157-165. 27. Finan DA, Palerm CC, Doyle FJ III, et al. Identification of empirical dynamic models from type 1 diabetes subject data. Proc Am Control Conf. 2008;2099-2104. 28. Schmelzeisen-Redeker G, Staib A, Strasser M, Müller U, Schoemaker M. Overview of a novel sensor for continuous glucose monitoring. J Diabetes Sci Technol. 2013;7:808-814.
research-article2015
DSTXXX10.1177/1932296815590154Journal of Diabetes Science and TechnologySchmelzeisen-Redeker et al
Original Article
Time Delay of CGM Sensors: Relevance, Causes, and Countermeasures
Journal of Diabetes Science and Technology 2015, Vol. 9(5) 1006–1015 © 2015 Diabetes Technology Society Reprints and permissions: sagepub.com/journalsPermissions.nav DOI: 10.1177/1932296815590154 dst.sagepub.com
Günther Schmelzeisen-Redeker, PhD1, Michael Schoemaker, PhD1, Harald Kirchsteiger, PhD2, Guido Freckmann, MD3, Lutz Heinemann, PhD4, and Luigi del Re, PhD2
Abstract Background: Continuous glucose monitoring (CGM) is a powerful tool to support the optimization of glucose control of patients with diabetes. However, CGM systems measure glucose in interstitial fluid but not in blood. Rapid changes in one compartment are not accompanied by similar changes in the other, but follow with some delay. Such time delays hamper detection of, for example, hypoglycemic events. Our aim is to discuss the causes and extent of time delays and approaches to compensate for these. Methods: CGM data were obtained in a clinical study with 37 patients with a prototype glucose sensor. The study was divided into 5 phases over 2 years. In all, 8 patients participated in 2 phases separated by 8 months. A total number of 108 CGM data sets including raw signals were used for data analysis and were processed by statistical methods to obtain estimates of the time delay. Results: Overall mean (SD) time delay of the raw signals with respect to blood glucose was 9.5 (3.7) min, median was 9 min (interquartile range 4 min). Analysis of time delays observed in the same patients separated by 8 months suggests a patient dependent delay. No significant correlation was observed between delay and anamnestic or anthropometric data. The use of a prediction algorithm reduced the delay by 4 minutes on average. Conclusions: Prediction algorithms should be used to provide real-time CGM readings more consistent with simultaneous measurements by SMBG. Patient specificity may play an important role in improving prediction quality. Keywords continuous glucose monitoring, CGM, MARD, accuracy, precision, performance evaluation, performance comparison, time delay Continuous glucose monitoring (CGM) offers people with diabetes the opportunity to improve their blood glucose (BG) control by monitoring glucose levels, even at times when the patient would not do this, for example, during the night. This allows detection of clinically relevant issues that might not be noticed with conventional capillary measurements of BG. CGM measurements also provide trend information based on past measurement results; this allows estimating the change in glycemia for a short period of time in the future. In addition, CGM systems help avoid hypoglycemic events by providing alarms when a predefined glucose threshold is exceeded or even prior to this if a pending hypoglycemia is predicted.1,2 One reason why CGM data should not be used for determining insulin doses by patients with diabetes, according to the officially approved guidelines set by the regulatory authorities,
is that CGM readings often do not correspond exactly to self monitoring of blood glucose (SMBG) measurement results taken at the same time—especially during fast changes in BG. CGM systems have exhibited an impressive progress in their analytical performance from 1 generation to the next. Nevertheless, 1 1
Roche Diabetes Care GmbH, Mannheim, Germany Institute for Design and Control of Mechatronical Systems, Johannes Kepler University, Linz, Austria 3 Institute for Diabetes-Technology GmbH at Ulm University, Germany 4 Science & Co, Düsseldorf, Germany 2
Corresponding Author: Günther Schmelzeisen-Redeker, PhD, Roche Diabetes Care GmbH, Sandhofer Str 116, 68305 Mannheim, Germany. Email: [email protected]
1007
Schmelzeisen-Redeker et al important reason for differences between CGM and SMBG measurement results is the time delay. The aim of this article is to characterize the time delay observed with a novel CGM system and measures to reduce this delay.
CGM systems measure glucose levels in interstitial fluid (ISF) in the subcutaneous adipose tissue and not directly in blood. Once glucose levels change in either compartment there is a difference in measurement results between these 2 compartments before they equalize after some minutes again. This, in turn, results in the well-known fact that results of CGM measurements are shifted in time compared to results of SMBG measurements during rapid swings in glycemia. This delay is the consequence of the process of glucose diffusion across the walls of capillary vessels and through the interstitial space to the sensor. This process requires some time, and the delay can be observed during both rising and decreasing BG values, probably with varying impact.3-5 It is commonly understood that the delay and the rate of change of glucose in the ISF are affected by several physiological factors, for example, by the local blood flow, the local tissue perfusion, the permeability of the ISF, the local convection in the ISF, and so on. The average physiological time delay (ISF glucose to BG) is assumed to be in the range of 5 to 10 minutes. It is of interest to note that the differences in time delays reported for plasma BG and ISF are not conclusive.6 A more recent estimation based on glucose tracer experiments with healthy humans using a microdialysis technique resulted in an average physiologic time delay of 5-6 minutes.7 Other estimates using the same approach for a very small study population report a similar value for patients with type 1 diabetes.8 There is strong evidence for a great variability of the time delay between individual subjects even under reproducible experimental conditions and independently of a specific CGM system.
Technologically Caused Time Delay The measurement of glucose by the CGM system and the associated calculations inside the CGM system before the measurement result is displayed to the patient might add an additional, but usually short, physical time delay. The measurement delay is caused by both the diffusion of glucose through the protecting membranes and/or glucose flux limiting membranes covering the CGM sensor as well as the speed of glucose reactions inside the sensor. This fraction of time delay is sensor-specific, and little has been published thus far about the physical time delay of different CGM sensors.9 However, this delay is assumed to be within the range of a few minutes. The amperometric sensors measure an electric current that is proportional to the glucose levels in the ISF. This current must be transferred into a glucose reading by proper
raw MA30 MA60
30 25
el. current [nA]
Physiologically Caused Time Delay
35
20 15 10 5 0 2500
2600
2700
2800
2900 3000 time [min]
3100
3200
3300
Figure 1. Raw CGM signal (blue) filtered using a 30-minute (red) and 60-minute (green) moving average filter.
calibration using a conventionally measured capillary BG value. The sensor signal is also subject to a number of disturbances which are mainly slowly varying and random. Slowly varying signal disturbances are caused by biofouling10 and are removed and/or reduced by calibration until the measured signal becomes overly deteriorated and the sensor needs to be replaced. The random noise superimposed over the raw data is reduced by averaging signals over a period of time (usually several minutes), and only these mean values are usually reported. Filtering is another possibility for reducing the noise of the sensor signal and use of additional information increases the speed of this compensation.11,12 However, mathematical filters applied in real time to provide reliable glucose information unavoidably generate an algorithmic time delay (since real time filters can only utilize sensor signals from the past; stronger filtering requires sensor signals from a longer time period in the past, which thereby leads to longer time delays).13 In addition, algorithms are employed to smooth the glucose profile over time and to reduce and/or eliminate artifacts (socalled smart sensors).14,15 Here again only little information is available about the impact of all these time-consuming factors on the overall time delay of commercial CGM systems; thus far, values of 3 to 12 minutes have been reported.9,16 The trade-off between variance of the sensor readings and time delay introduced by filtering can be appreciated in a simulation example shown in Figure 1, where the original raw data were filtered using a moving average filter with 2 different settings for the filter length (equivalent to 30 and 60 minutes). It is clear that longer averaging (MA60, green) leads to less noisy signals, but increases the physical time delay considerably. Using large filter lengths with such simple filters is not realistic as the BG level keeps changing. In modern CGM
1008
Total Time Delay As the resulting total time delay potentially affects the ability of patients to encounter a clinically relevant situation, knowledge of this time delay and measures to reduce it are of clinical relevance. To date, the focus of reports about the characteristics of CGM systems has been on their analytical performance and not on other factors such as the time delay of the respective system. The number of publications in which data about the time delays observed with different CGM systems are reported, favorably from head-to-headcomparisons, is quite limited.17,18 For patients with diabetes who use a given CGM system, the combined effect of time delays and measurement errors is of concern as both factors contribute to the measurement differences they observe between the CGM data and BG measurements performed at the same time. As the accuracy of CGM systems has improved significantly over the years, other sources of errors, such as time delays, become more relevant. In the literature a considerable range of overall time delays—between 5 to 40 minutes9,16—has been reported. It is not clear to which extent the reported differences are due to the different CGM systems used in such studies and/or to the experimental conditions used to evaluate this delay. In these publications, the extent of interindividual differences (= differences between patients) or intraindividual differences (= within the same patient over a longer time horizon) with a given CGM system is most often not reported, partially because of the lack of sufficiently long CGM recordings. However, such differences might be substantial (see the Results section). Also, little is known about the differences between the glycemic ranges with respect to the delay; that is, is the delay of the same magnitude in the low and high glycemic range? Also other patient-specific effects might influence the delay on a day-to-day basis (eg, physical exercise and wear time of the CGM system). In other words, not much is known either about the variability of the time delay between patients or the variability over time in the same patient.
Effect of Time Delay on Sensor Performance The following example demonstrates the effect of a time delay on CGM sensor performance. Figure 2 shows a CGM profile (from Zschornack et al19) of a patient with type 1 diabetes together with the respective SMBG values to visualize the magnitude of the time delay. The measure most often used to characterize the measurement performance of CGM systems is the mean absolute relative difference (MARD) (see the Methods section below for a formal definition). For single paired measurements in time, the absolute relative difference (ARD) can be computed. ARD
400 300
[mg/dL]
systems more complex filters are used that have better performance characteristics than moving average, but exhibit the same basic trade-off.
Journal of Diabetes Science and Technology 9(5)
CGM
SMBG
200 100 0
1500
2000
2500
3000
3500
3000
3500
80 60
abs.error (ARD)
mean=14.2
40 20 0
1500
2000
2500 time [min]
Figure 2. CGM recordings and results of glucose measurements in capillary blood (SMBG) (upper panel). Absolute relative difference between CGM and SMBG values and the mean thereof (lower panel).
is shown as the blue line in the lower panel in Figure 2, together with the MARD (red line), which is 14.3% in this case. Assuming that the CGM measurement itself is perfect (has an analytical error of 0%), and also the BG measurement has no analytical error, all differences in the measurement results have to be attributed to the time delay. One option to evaluate the impact of the time delay on the MARD is to shift the 2 time series against each other; so shifting the artificial SMBG values in relation to the CGM measurements by, for example, 10 minutes results in a MARD of 9.5% (Figure 3). This indicates that although there is absolutely no error at all in the measurements per se, the time delay has a high impact on the performance characteristics. A related analysis of the effect of time delays on the CGM performance was also presented in Scuffi et al.6 To separate the effects of time delays from that of analytical error, 1 guideline for the evaluation of CGM systems issued by the Clinical and Laboratory Standard Institute (CLSI)20 suggests procedures to retrospectively remove the time delay from the measurements before evaluating the accuracy. This yields a figure which is closer to the “real” analytical accuracy of the given CGM system. This value should be, to some extent, independent of the specific user. Other such procedures aiming for the same goal have been proposed; however, their major disadvantage is that they are based on postprocessing of measured data and do not provide any help with real-time data.21 This would, however, be necessary for clinical practice.
Measures to Reduce the Impact of the Time Delay Against this background, the question arises as to what measures can be taken to reduce the effect of the time delay on CGM recordings. Roughly speaking, there are 2 main possibilities to deal with a time delay:
1009
Schmelzeisen-Redeker et al
300
400 CGM
SMBG
200
200
100 0
1500
2000
2500
3000
150
3500
100 50
60 abs.error (ARD)
mean=9.5
0 5200
20
1500
2000
2500 time [min]
3000
3500
Figure 3. Absolute relative difference and mean value thereof when assuming no analytical error, only an overall time delay between CGM and SMBG, using the same data as presented in Figure 2.
1. ISF glucose levels are closer to those in the brain than the results of BG measurements.22 As the brain is the clinically relevant organ (glucose levels in the brain are critical for preventing, eg, hypoglycemic symptoms), it represents the ultimate target for blood glucose control. In the long run, it might therefore be possible to switch the target of glycemic control from BG values measured in peripheral blood to values obtained from ISF glucose. The usage of ISF data for clinical decisions would clearly require a paradigm shift. 2. Using mathematical prediction algorithms to prospectively compensate for the time delay would allow the user to receive a CGM based glucose value that is closer (in the time domain) to the SMBG results. Making use of prediction algorithms allows forecasting future glucose values (eg, those that can be expected in the next minutes) based on the information available at the current time and based on a mathematical model. Predictors typically perform well for short-term prediction, that is, up to about 20 minutes. Several algorithms have been published to achieve this goal (see, eg, Naumova et al,23 Kirchsteiger et al,24 Pappada et al,25 Gani et al26). Below, we have highlighted the idea by using a simple autoregressive algorithm27 of the fifth order applied to CGM data, which enables a relatively good prediction for a 10minute horizon (Figure 4). This demonstrates that for short prediction horizons, the accuracy of the prediction is good; however, for obvious reasons this deteriorates rapidly for longer time periods, see the larger errors in the bottom panel of Figure 4 for the 20-minute prediction. Several predicted traces are shown in Figure 5, where the red cross signs indicate predicted values with a 5-minute interval, that is, the
error [mg/dL]
40
0
cgm 10min prediction 20min prediction
250
[mg/dL]
[mg/dL]
300
5300
5400
5500
5600 5700 5800 time [min]
5900
6000
6100
6200
5300
5400
5500
5600
5900
6000
6100
6200
50 0 -50 -100 5200
5700
5800
Figure 4. CGM data and 10-, 20- minute prediction of an autoregressive model of fifth order (top plot) and the prediction errors (bottom plot). Figure shows validation data; the model was trained on the first 2 days of the data set.
300
250
200
150
10-min 20-min 30-min prediction Detail at t=5700
100
50
0 5200
5300
5400
5500
5600 5700 5800 time [min]
5900
6000
6100
6200
Figure 5. Predicted traces shown at some selected points in time for the same data as in Figure 4. For a given point, the red lines/crosses indicate the future values expected after 10/20/30 minutes using the information available until then.
trace starts at the current time at the current CGM value (blue line), and then shows the 5-, 10-, . . . 30-minute prediction. This kind of prediction exclusively uses past information, which means that any measured change happening between the current time the prediction is computed and the future estimate is not taken in account. For instance, a meal between the current time and the end of the prediction horizon will raise the real BG value but cannot be considered by the predictor. Against this background, it is of the utmost importance to develop safe predictors using as much additional information as possible (including meal information, insulin administrations, activity, etc). In addition, a prediction with
1010
Journal of Diabetes Science and Technology 9(5)
Table 1. Descriptive Characteristics of the Study Participants (20 Male, 19 Female).
Age (years) Height (cm) Weight (kg) BMI (kg/m²) Skinfold thickness (mm) Duration of diabetes (years) HbA1c (%)
Mean (SD)
Range
47 (11) 175 (10) 82 (16) 27 (4) 27 (7) 23 (12) 7.8 (1.2)
21-63 156-197 55-116 21.6-41.4 14-39 2-48 5.8-11.7
an incorrect time horizon could be more detrimental than useful. Patient-specific time delays—if it can be confirmed that they exist—can be used as prediction horizon to personalize the CGM sensor and compensate the delay.
Clinical Data To evaluate the different factors that have an impact on the time delay, CGM data were analyzed that were obtained while running a clinical study with patients with type 1 diabetes using a prototype CGM sensor. The design28 and performance19 of this CGM sensor has been reported earlier. The clinical– experimental study was performed in a Clinical Research Center (Institut für Diabetes-Technologie Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm [IDT]) and was approved by an Ethical Committee and the competent authority (Bundesinstitut für Arzneimittel und Medizinprodukte, Germany). The study was performed over 2 years and consisted of 5 consecutive phases. The study design of each phase followed essentially the recommendation of a guideline for CGM evaluation POCTO5A.20 This includes induction of a significant glucose swing into the high and low glucose range by late and slightly overdosed insulin delivery after breakfast on 2 of the 7 days of each study phase. Patients simultaneously wore 2 or 4 CGM sensors. Capillary blood samples were routinely measured in duplicate once per hour during the day and at 3 am at night. For all analyses the mean of the 2 readings was used as the reference value. Plasma glucose levels in the blood samples were measured using a standard home BG meter (Accu-Chek® Aviva BG meter, Roche Diagnostics GmbH). Glucose levels were measured more often in phases with induced glucose swings (15-minute intervals). Details of the study design have been reported elsewhere.19 A total of 37 patients with type 1 diabetes participated in the clinical study (see Table 1 for patient characteristics). Eight of these patients participated in 2 different study phases separated by at least 8 months. The total number of CGM sensor data sets that could be used for the time delay analysis was 108 (35/10 participations with 2/4 sensors in parallel; sensor data of 1 participation excluded because of failure of 1 of the 2 sensors). Data analysis was performed using raw
signals, that is, the values measured directly by the sensor (current in nano Ampere [nA]), processing was done by standard statistical methods. Part of the data analysis was done with retrospectively calibrated sensor signals (in mg/dl but without using a mathematical filter).19
Methods for Time Delay Estimation In the analysis presented in this article, the focus is on the part of the time delay that results from reasons of physiology and sensor design, and not on the algorithmic treatment of the signals. Thus, the distribution of the time delay of the raw electrical current versus BG of the individual CGM sensors during the different study phases was mainly examined. The quality of time delay estimation depends on a number of factors, among them (1) frequency and accuracy of BG measurements; (2) time resolution and quality of sensor raw signals (low noise, no artifacts); (3) extent and rate of change of glucose swings; (4) stability of time lag over time of experiment. The low noise and 1-minute measurement frequency of the sensor, the use of the mean of duplicate BG values, the high frequency, and large range of BG values during induced glucose swings take these limiting factors into account.
Correlation Coefficient Maximization The correlation between 2 stochastic signals x1 (t ) , x2 (t ) is given by the correlation function
φ (τ ) = lim
N →∞
1 N
N
∑x ( k ) x (k + τ ) 1
2
k =1
Intuitively, this function describes the dependency of the 2 signals when shifted against each other in time by τ minutes. A small value of the function indicates less dependency while a high value indicates high dependency (ie, the 2 signals are more similar to each other). In our application, we consider CGM data as the first signal and SMBG measurement results as the second one. Then, we compute the correlation function as defined above forvarious values of τ by retrospectively shifting the CGM trace in time. The value τ* that results in the highest value of the correlation function is the estimated time delay. This method was used with the raw glucose sensor signals.
MARD Minimization The ARD between a CGM data point yCGM (t ) and a reference measurement yref (t ) is given by ARD ( tk ) = 100% and the mean ARD as
yCGM ( tk ) − yref (tk ) yref (tk )
1011
Schmelzeisen-Redeker et al
Table 2. Estimated Time Delay (Minutes) Using the 2 Different Methods Described in the Methods Section.
22 20
Time delay [min]
18
Correlation coefficient
MARD
9.5 3.7 9 4 3 20
8.0 3.4 8 4 2 17
Mean SD Median IQR Minimum Maximum
16 14 12 10 8 6 4 2
Phase 1
Phase 2
Phase 3
Phase 4
Phase 5
Table 3. Estimated Time Delay (Minutes) When Using a Prediction Algorithm for the Same Data Shown in Table 2.
Study Phase
Correlation coefficient, prediction
Figure 6. Box plot of overall time delay distribution during different phases of the clinical study. Squares: medians; rectangles: 25% and 75% quantiles; extensions from rectangles: minimums and maximums; circles: outliers.
MARD =
1 N
N
∑ARD
Mean SD Median IQR Minimum Maximum
5.9 3.7 5 3 –1 19
MARD, prediction 4.3 3.0 4 4.25 –1 12
k
k =1
Similar to the previous method, the MARD can be computed also when first shifting the CGM time series relative to the available reference measurements by τ minutes. By doing this systematically minute by minute, we receive a function MARD(τ), and the minimum of this function determines the estimated time delay. This method was used with the calibrated glucose sensor signals.
Prediction Algorithm The algorithm utilizes a predetermined population average time delay (Δt) and a Kalman filter for real time estimation of the sensor raw signal at time t (It) and its rate of change. Both variables are used for a linear extrapolation (prediction) of the sensor raw signal Δt minutes later (It+ Δt). Instead of It the predicted sensor raw signal is used for calibration and conversion of raw signals into sensor glucose values at time t. In case of the prototype sensor Δt was 8 minutes.
Results Time Delay An analysis of the CGM data sets with respect to differences in the time delays in the different study phases showed no significant differences in the 5 different study phases (ANOVA analysis; P = .55; Figure 6). This allows for the pooling of the data from all phases for statistical analysis. Based on the maximization of the correlation coefficient of the sensor raw signal with respect to BG, the mean (standard deviation [SD]) overall time delay was 9.5 (3.7) minutes. The
median time interval was 9 minutes and the interquartile range (IQR: difference between the 25% lowest and 75% highest values) 4 minutes (Table 2). The estimated time delay based on the minimization of the MARD was on average 1 to 2 minutes smaller (P < .001). The use of a prediction algorithm, as conceptually introduced in the introduction section, led to a mean time delay reduction of the overall time delay—as seen by the patient— by about 4 minutes (Table 3).
Intraindividual Time Delay Sixty-eight CGM data sets were obtained in study phases with patients simultaneously wearing 2 CGM sensors (sensor A and sensor B). The correlation of the overall time delays observed with the 2 sensors in the same patient is given in Figure 7. The coefficient of determination (corrected R2) was .51. This means that 51% of the variability of the time delay results from the patients. In 1 study phase, patients wore 4 CGM sensors simultaneously (10 patients, 40 data sets). The box plot of the time delays of these patients when grouped by the individual patient confirms that the patient per se has a significant effect on the overall time delay (Figure 8; ANOVA P < .01); the corrected R2 was .43, which means that 43% of the time delay variability is explained by the patient effect. Eight patients who participated in 2 study phases were analyzed separately for each study phase (phases 1 and 2), and the estimated delays plotted versus each other are given in Figure 9. The good correlation of the mean time delays of the 2 study phases indicates that the time delay of a given patient is rather constant over 8 months.
1012
Journal of Diabetes Science and Technology 9(5) 18
20 18
16
mean time delay 2 [min]
time delay sensor A [min]
16 14 12 10 8
14
12
10
8
6
6
4 2
2
4
6
8
10
12
14
16
18
20
22
4
4
6
8
10
12
14
16
18
mean Time delay 1 [min]
time delay sensor B [min]
Figure 7. Correlation of the time delays observed in 2 simultaneously worn CGM sensors in the same patient (sensors A and B); red line: linear regression line; each circle represents data from 1 patient.
Figure 9. Correlation of mean time delays of the CGM sensors of patients participating in 2 different study phases (mean time delay 1 and 2); time between study phases was 8 to 12 months.
Discussion 20 18
TimeDelay [min]
16 14 12 10 8 6 4
63
54
61
57
59
51
52
53
62
64
Patient-ID
Figure 8. Box plot of time delay distribution of CGM sensors for 10 patients wearing 4 sensors simultaneously. Squares: medians; rectangles: 25% and 75% quantiles; extensions from rectangles: minimums and maximums; circles: outliers.
Since the results suggest that patient-specific time delays exist, an analysis of the impact of descriptive characteristics (anamnestic and anthropometric data) of these patients on the time delays was performed. The scatter plots of the mean time delays versus all parameters evaluated (weight, height, body mass index, skinfold thickness, HbA1c, years with diabetes) did not show any significant correlation (Figure 10).
Assessment of the accuracy of CGM systems can be done from 2 points of view: either by comparing the (analytical) performance of different CGM systems in a head-to-head experimental situation or by assessing the accuracy as experienced by the patient when comparing CGM with SMBG. In the first case, time delays should be removed from data as suggested by POCT-05A,20 to allow for fair comparison focusing on the analytical performance only. In the second case, the time delay should not be removed, as the patient would not be able to distinguish between measurement errors resulting from poor accuracy or time delay. Time delays can be estimated reliably only during glucose swings; the larger the glucose swings the better. During such swings, BG readings should be obtained as often as possible with a reliable measurement method to enable precise estimation of the time delay. In addition, the BG measurements should be done at least in duplicate to further reduce the BG measurement error. Several factors affect the quality of the time delay estimation. In particular, artifacts such as dropouts could negatively affect time delay estimation as it is difficult to correct the raw signals for such artifacts. The influence of sensor noise can be limited because it can be reduced for the analysis without affecting the delay estimation. It has been reported17 that the time delay is longer during phases of rising rather than during declining BG levels. A time delay analysis taking both phases into account, as it was done here, reflects an intermediate value for the overall time delay.
1013
18
18
16
16
Mean Time Delay [min]
Mean Time Delay [min]
Schmelzeisen-Redeker et al
14 12 10 8 6
14 12 10 8 6
4 150
155
160
165
170
175
180
185
190
195
4 50
200
60
70
80
18
18
16
16
14 12 10 8 6
110
120
14 12 10 8
22
24
26
28
30
32
34
36
38
40
42
4 12
44
14
16
18
20
BMI
22
24
26
28
30
32
34
36
38
40
Skinfold Thickness [mm]
18
18
16
16
Mean Time Delay [min]
Mean Time Delay [min]
100
6
4 20
14 12 10 8 6 4
90
Weight [kg]
Mean Time Delay [min]
Mean Time Delay [min]
Height [cm]
14 12 10 8 6
0
10
20
30
40
Years with Diabetes
50
4
5
6
7
8
9
10
11
12
HBA1c [%]
Figure 10. Correlation of mean time delay including patient height (a), weight (b), BMI (c), skinfold thickness (d), duration of diabetes (e), and HbA1c (f).
These effects can lead to errors in the time delay estimation of an individual CGM sensor. To achieve reliable results on the average time delay of a given CGM sensor type the number of sensors tested in the clinical setting must be sufficiently high. For quality assurance of the time delay estimation a visual inspection of the time course of blood and sensor glucose should be conducted.
Patient-Specific Effects A pooling of time delay data of different study phases is only allowed if there is no difference in average sensor time delays between these phases. The statistical analysis shows that this assumption is fulfilled for the data set used.
1014 Despite the limitations of the accuracy and precision of SMBG measurement results, both of which contribute to the variability of time delays, the correlation of time delays of sensors working simultaneously in the same patient is evident in Figure 8. This indicates that the physiologic time delay varies between patients to a significant extent; approximately 50% of the time delay variability can be explained by the patient effect. In addition, the analysis of the mean time delays of patients who participated in 2 study phases, separated by at least 8 months, indicates that the average time delay of a patient is rather constant at least over this time period. However, the relatively small database of only 8 patients asks for more data to confirm these results. The existence of a patient-specific time delay might be useful for further advances of algorithms to improve the quality of CGM systems. The overall time delay of a given CGM sensor is composed of the physiologic, the sensor-specific physical and the algorithmic time delays. The physical time delay and the variability of the CGM sensors is also determined (to a relatively small extent) by the production processes and the materials used for sensor production. These were identical for all 108 sensors on which this time delay analysis is based. The described methods of time delay estimation for the raw and the retrospectively calibrated sensor signals do not introduce an algorithmic time delay. Therefore, the measured patient-specific effect on time delay can be attributed to differences in the physiologic time delay only. Of particular interest is the root cause for the differences in physiologic time delay between patients. One hypothesis is that the length of transport path of glucose from the capillary blood vessels in adipose tissue to the sensor surface varies from patient to patient due to different sizes of the fat cell diameter. However, no correlation to patient parameters like weight, BMI, or skinfold thickness could be identified in our analysis. Other root causes may be different vascularization and capillary density of adipose tissue, differences in adipose tissue blood flow affecting the speed of blood exchange in the tissue (similar to the Alternative Site Testing phenomenon where a time delay exists between blood obtained from the forearm and blood obtained from the fingertips), and so on. Further research is needed to test these and other possible hypotheses on the origin of the patient-specific time delay.
Consequences for Medical Use Cases The existence of a patient-specific time delay which might vary from a few to up to 20 minutes must be taken into consideration when a patient is to use a CGM sensor. The impact of the time delay depends on the area of use of the CGM system (when it is used for data-logging and retrospective data analysis, the delay is not as relevant as for real-time decision making). All time-critical decisions (hypo- and hyperglycemia warnings, input for artificial pancreas systems, etc)
Journal of Diabetes Science and Technology 9(5) suffer from a high time delay of a given patient. For such patients CGM performance parameters such as MARD would be rather poor even if the sensor would monitor ISF glucose accurately. The effect of high time delays on the quality of diagnostic or therapeutic systems must be carefully and specifically evaluated for the respective application. The variability of the time delay over sensor wear time is moderate, particularly when taking the accuracy of time delay measurement into account. However, the extreme examples of time delay variation over sensor wear time show that the quality of, for example, prediction algorithms for hypo- and hyperglycemia detection and warning might vary over time even if the quality of the sensor signal (low noise, no artifacts) were constant.
Conclusions The time delay of CGM sensors is composed by an intrinsic physiologic property of the approach and by the technology used for CGM measurement. The time delay represents an important factor that must be taken into account for optimal usage of CGM systems. Further technological improvement of CGM sensors will lead to a lower noise (and better analytical performance) as well as in faster internal glucose transport, both of which will reduce the physical time delay. Glucose prediction techniques are the only available approach that allows real-time compensation of time delays in general. However, a precise estimation of the total time delay, preferably patient-specific, is required for good prediction. There is growing evidence that at least part of the time delay is patient-specific. If confirmed, this would enable usage of learning techniques to improve the quality and reliability of glucose prediction. Abbreviations ARD, absolute relative difference; BG, blood glucose; CGM, continuous glucose monitoring; CLSI, Clinical and Laboratory Standard Institute; IDT, Institut für Diabetes-Technologie Forschungs- und Entwicklungsgesellschaft mbH an der Universität Ulm; IQR, interquartile range; ISF, interstitial fluid; MARD, mean absolute relative difference; SD, standard deviation; SMBG, self-monitoring of blood glucose.
Declaration of Conflicting Interests The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: GS and MS are full-time employees of Roche Diagnostics. GF is general manager of Institut für Diabetes-Technologie Forschungsund Entwicklungsgesellschaft mbH an der Universität Ulm, Ulm, Germany, which carries out studies evaluating BG meters and medical devices for diabetes therapy on behalf of various companies, and received speakers’ honoraria or consulting fees from Abbott, Bayer, Berlin-Chemie, Becton-Dickinson, Dexcom, Menarini Diagnostics, Roche Diagnostics, Sanofi, and Ypsomed. LH is a consultant for a number of companies developing new diagnostic and therapeutic options for diabetes treatment and is a member of a Sanofi advisory
Schmelzeisen-Redeker et al board for biosimilar insulins. He is a partner of Profil Institute for Clinical Research, US and Profil Institut für Stoffwechselkrankheiten, Germany.
Funding The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was (partially) funded by Roche Diagnostics.
References 1. Palerm CC, Bequette BW. Hypoglycemia detection and prediction using continuous glucose monitoring—a study on hypoglycemic clamp data. J Diabetes Sci Technol. 2007;1:624-629. 2. Vashist SK. Continuous glucose monitoring systems: a review. Diagnostics. 2013;3:385-412. 3. Kulcu E, Tamada JA, Reach G, Potts RO, Lesho MJ. Physiological differences between interstitial glucose and blood glucose measured in human subjects. Diabetes Care. 2003;26:2405-2419. 4. Ward WK, Engle JM, Branigan D, Youssef JE, Massoud RG, Castle JR. The effect of rising vs. falling glucose level on amperometric glucose sensor lag and accuracy in type 1 diabetes. Diabet Med. 2012;29:1067-1073. 5. Wentholt IME, Hart AAM, Hoekstra JBL, Devries JH. Relationship between interstitial and blood glucose in type 1 diabetes patients: delay and the push-pull phenomenon revisited. Diabetes Technol Ther. 2007;9:169-175. 6. Scuffi C, Lucarelli F, Valgimigli F. Minimizing the impact of time lag variability on accuracy evaluation of continuous glucose monitoring systems. J Diabetes Sci Technol. 2012;6: 1383-1391. 7. Basu A, Dube S, Slama M, et al. Time lag of glucose from intravascular to interstitial compartment in humans. Diabetes. 2013;62:4083-4087. 8. Basu A, Dube S, Veettil S, et al. Time lag of glucose from intravascular to interstitial compartment in type 1 diabetes. J Diabetes Sci Technol. 2015;9:63-68. 9. Keenan DB, Mastrototaro JJ, Voskanyan G, Steil GM. Delays in minimally invasive continuous glucose monitoring devices: a review of current technology. J Diabetes Sci Technol. 2009;3:1207-1214. 10. Klueh U, Liu Z, Feldman B, et al. Metabolic biofouling of glucose sensors in vivo: role of tissue microhemorrhages. J Diabetes Sci Technol. 2011;5:583-595. 11. Bequette BW. Continuous glucose monitoring: real-time algorithms for calibration, filtering and alarms. J Diabetes Sci Technol. 2010;4:404-418. 12. Bequette BW. Optimal estimation applications to continuous glucose monitoring. Proc Am Control Conf. 2004;958-962.
1015 13. Oppenheim AV, Schafer RW, Oppenheim AV, eds. DiscreteTime Signal Processing. Boston, MA: Pearson; 2009. 14. Del Favero S, Facchinetti A, Sparacino G, Cobelli C. Improving accuracy and precision of glucose sensor profiles: retrospective fitting by constrained deconvolution. IEEE Trans Biomed Eng. 2014;61:1044-1053. 15. Facchinetti A, Sparacino G, Cobelli C. Online denoising method to handle intraindividual variability of signal-to-noise ratio in continuous glucose monitoring. IEEE Trans Biomed Eng. 2011;58:2664 -2671. 16. Rebrin K, Sheppard NF, Steil GM. Use of subcutaneous interstitial fluid glucose to estimate blood glucose: revisiting delay and sensor offset. J Diabetes Sci Technol. 2010;4:1087-1098. 17. Boyne MS, Silver DM, Kaplan J, Saudek CD. Timing of changes in interstitial and venous blood glucose measured with a continuous subcutaneous glucose sensor. Diabetes. 2003;52:2790-2794. 18. Kovatchev BP, Shields D, Breton M. Graphical and numerical evaluation of continuous glucose sensing time lag. Diabetes Technol Ther. 2009;11:139-143. 19. Zschornack E, Schmid C, Pleus S, et al. Evaluation of the performance of a novel system for continuous glucose monitoring. J Diabetes Sci Technol. 2013;7:815-823. 20. CLSI. Performance metrics for continuous interstitial glucose monitoring; approved guideline. POCT05-A. Wayne, PA: Clinical and Laboratory Standards Institute; 2008. 21. Carter G. Coherence and time delay estimation. Proc IEEE. 1987;75:236-255. 22. Nielsen JK, Djurhuus CB, Gravholt CH, et al. Continuous glucose monitoring in interstitial subcutaneous adipose tissue and skeletal muscle reflects excursions in cerebral cortex. Diabetes. 2005;54:1635-1639. 23. Naumova V, Pereverzyev SV, Sampath S. A meta-learning approach to the regulatrized learning—case study: blood glucose prediction. Neural Netw. 2012;33:181-193. 24. Kirchsteiger H, Johansson R, Renard E, Re LD. Continuoustime interval model identification of blood glucose dynamics for type 1 diabetes. Int J Control. 2014;87:1454-1466. 25. Pappada SM, Cameron BD, Rosman PM. Development of a neural network for prediction of glucose concentration in type 1 diabetes patients. J Diabetes Sci Technol. 2008;2:792-801. 26. Gani A, Gribok A, Lu Y, Ward W, Vigersky R, Reifman J. Universal glucose models for predicting subcutaneous glucose concentration in humans. IEEE Trans Info Tech Biomed. 2010;14:157-165. 27. Finan DA, Palerm CC, Doyle FJ III, et al. Identification of empirical dynamic models from type 1 diabetes subject data. Proc Am Control Conf. 2008;2099-2104. 28. Schmelzeisen-Redeker G, Staib A, Strasser M, Müller U, Schoemaker M. Overview of a novel sensor for continuous glucose monitoring. J Diabetes Sci Technol. 2013;7:808-814.
