Florian Reiterer

Significance and Reliability of MARD for the Accuracy of CGM Systems

Background: There is a need to assess the accuracy of continuous glucose monitoring (CGM) systems for several uses. Mean absolute relative difference (MARD) is the measure of choice for this. Unfortunately, it is frequently overlooked that MARD values computed with data acquired during clinical studies do not reflect the accuracy of the CGM system only, but are strongly influenced by the design of the study. Thus, published MARD values must be understood not as precise values but as indications with some uncertainty. Data and Methods: Data from a recent clinical trial, Monte Carlo simulations, and assumptions about the error distribution of the reference measurements have been used to determine the confidence region of MARD as a function of the number and the accuracy of the reference measurements. Results: The uncertainty of the computed MARD values can be quantified by a newly introduced MARD reliability index (MRI), which independently mirrors the reliability of the evaluation. Thus MARD conveys information on the accuracy of the CGM system, while MRI conveys information on the uncertainty of the computed MARD values. Conclusions: MARD values from clinical studies should not be used blindly but the reliability of the evaluation should be considered as well. Furthermore, it should not be ignored that MARD does not take into account the key feature of CGM sensors, the frequency of the measurements. Additional metrics, such as precision absolute relative difference (PARD) should be used as well to obtain a better evaluation of the CGM performance for specific uses, for example, for artificial pancreas.

Deviation analysis of clinical studies as tool to tune and assess performance of diabetes control algorithms

Clinical trials are the commonly accepted proofs of validity of therapeutic approaches in most medical fields. In many cases, a therapy approach is defined a priori and administered by medical personnel during the trial. In the case of type 1 diabetes mellitus (T1DM), the therapy approach typically consists in fixing a rule for the intake of insulin and taking the corresponding decisions during the day according to measurements and inputs by the patient, e.g. the expected carbohydrate intake. As clinical trials are expensive and complex to realize, only few variants can be really tested. However, in view of the large number of possible options, it would be very useful to be able to test a larger number of variants. Against this background, recently, several methods have been proposed in the scientific literature to extrapolate the effect of a modified therapy using real measurements as baseline. The key idea of all these methods consists in splitting the measurements into a controllable part and a “disturbance” component, which is assumed to be independent from the control action, i.e. from the insulin delivery. Of course, this splitting depends on the specific model assumptions, and the evaluation results of modified therapies may change according to the specific assumed model. However, as this paper shows at the example of the comparison between a standard and an adaptive bolus calculator, the results seem to become consistent if a large enough, representative dataset is used.

Identification of diurnal patterns in insulin action from measured CGM data for patients with T1DM

We propose a method to identify diurnal changes in insulin action in patients suffering from type 1 diabetes mellitus (T1DM) based on data recorded by continuous glucose monitoring systems (CGMS). In order to do so the data is fitted using a continuous time transfer function including time dependent terms. The identified values for the insulin needs per gram of carbohydrate were compared with the patient-specific carbohydrate-to-insulin-ratios used for the calculation of the bolus insulin needs. A good agreement between the identified parameters and values determined by diabetologists were found. Furthermore, the diurnal variations in insulin action (as inferred from the changes in the patient-specific carbohydrate-to-insulin-ratios) could be reproduced. The identified models, including the diurnal changes in insulin action and the information on the intra-patient variability, have the potential to be used in future studies for managing the blood glucose level of patients, e.g. in a smart bolus calculator.

Limits to the Evaluation of the Accuracy of Continuous Glucose Monitoring Systems by Clinical Trials

Systems for continuous glucose monitoring (CGM) are evolving quickly, and the data obtained are expected to become the basis for clinical decisions for many patients with diabetes in the near future. However, this requires that their analytical accuracy is sufficient. This accuracy is usually determined with clinical studies by comparing the data obtained by the given CGM system with blood glucose (BG) point measurements made with a so-called reference method. The latter is assumed to indicate the correct value of the target quantity. Unfortunately, due to the nature of the clinical trials and the approach used, such a comparison is subject to several effects which may lead to misleading results. While some reasons for the differences between the values obtained with CGM and BG point measurements are relatively well-known (e.g., measurement in different body compartments), others related to the clinical study protocols are less visible, but also quite important. In this review, we present a general picture of the topic as well as tools which allow to correct or at least to estimate the uncertainty of measures of CGM system performance.

Can We Use Measurements to Classify Patients Suffering from Type 1 Diabetes into Subcategories and Does It Make Sense

We propose two ideas of how recorded signals from continuous glucose monitoring systems could be used to derive information about the patient-specific characteristics of the glucose dynamics for individuals suffering from type 1 diabetes and how these characteristics of the glucose dynamics could be linked to basic patient data (sex, age,...). Ultimately, these relationships could be used in the future in order to classify patients based on these basic patient data. In the first approach a simple transfer function model was used to fit recorded signals from continuous glucose monitoring systems. Using this approach on data from a recent clinical study, a statistically significant relationship between the model parameters and sex, body mass index, weight and age of the corresponding patients could be identified. The observed relationships could be verified with findings in the clinical studies that were documented in the previous publications. In the second approach a moving average filter with a varying filter width was applied on the data and the variance between filtered and unfiltered signal as a function of the filter width was analysed. From the analysed data a relationship between the low blood glucose index and the high frequency content of signals from continuous glucose monitoring systems seems likely.

Alternative Frameworks for Personalized Insulin–Glucose Models

The description of the insulin–glucose metabolism has attracted much attention in the past decades, and several models based on physiology have been proposed. While these models provide a precious insight in the involved processes, they are seldom able to replicate and much less to predict the blood glucose (BG) value arising as a reaction of the metabolism of a specific patient to a given amount of insulin or food at a given time. Data-based models have proven to work better for prediction, but predicted and measured values tend to diverge strongly with increasing prediction horizon. Different approaches, for instance the use of vital signs, have been proposed to reduce the uncertainty, albeit with limited success. The key assumption hidden behind these methods is the existence of a single “correct” model disturbed by some stochastic phenomena. In this chapter, instead, we suggest using a different paradigm and to interpret uncertainty as an unknown part of the process. As a consequence, we are interested in models which yield a similar prediction performance for all measured data of a single patient, even if they do not yield a precise representation of any of them. This chapter summarizes two possible approaches to this end: interval models, which provide a suitable range; and probabilistic models, which provide the probability that the BG lies in predetermined ranges. Both approaches can be used in the framework of automated personalized insulin delivery, e.g., artificial pancreas or adaptive bolus calculators.

Identification of CGM Time Delays and Implications for BG Control in T1DM

Continuous glucose monitoring (CGM) systems are becoming increasingly popular for the management of type 1 diabetes mellitus (T1DM). However, one of the limitations of using CGM information for blood glucose (BG) control is the fact that the CGM sensor measures the glucose not in the blood, but in the interstitial fluid. The current paper shows how a sensor and patient-specific time delay between interstitial glucose (IG) and blood glucose (BG) can be identified from CGM recordings and BG measurements. The resulting time delay values were found to be highly patient-specific and correlated with patient characteristics. Furthermore, it was found that the CGM time delays are a good predictor for the required total daily dose (TDD) of insulin, as well as for the carbohydrate-to-insulin-ratio (CIR). Based on these findings we introduce a method of how insulin therapy in T1DM could possibly be adjusted based on identified CGM time delays.

Analyzing the Potential of Advanced Insulin Dosing Strategies in Patients With Type 2 Diabetes: Results From a Hybrid In Silico Study

Background: The ongoing improvement of continuous glucose monitoring (CGM) sensors and of insulin pumps are paving the way for a fast implementation of artificial pancreas (AP) for type 1 diabetes (T1D) patients. The case for type 2 diabetes (T2D) patients is less obvious since usually some residual beta cell function allows for simpler therapy approaches, and even multiple daily injections (MDI) therapy is not very widespread. However, the number of insulin dependent T2D patients is vastly increasing and therefore a need for understanding chances and challenges of an automated insulin therapy arises. Based on this background, this article analyzes conditions under which the use of more advanced therapeutic approaches, particularly AP, could bring a substantial improvement and should be considered as a viable therapy option. Method: Data of 14 insulin-treated T2D patients on MDI wearing a CGM device and deviation analysis methods were used to estimate the expected improvements in the clinical outcome by using self-monitoring of blood glucose (SMBG) with advanced carbohydrate counting, a full AP or intermediate approaches, either CGM measurements with MDI therapy or SMBG with insulin pump. HbA1C and time in range (70-140 mg/dl, 70-180 mg/dl, respectively) were used as a performance measure. Outcome measures beyond glycemic control (eg, compliance, patient acceptance) have not been analyzed in this study. Results: AP has the potential to improve the condition of many poorly controlled insulin-treated T2D patients. However, as the interpatient variability is much higher than in T1D, a prescreening is recommended to select suitable patients. Conclusions: Clinical criteria need to be developed for inclusion/exclusion of T2D patients for AP related therapies.

Nonlinear approach to virtual trials for insulin dosing systems

Patients with type 1 diabetes mellitus (T1DM) need to supply their body with insulin from external sources in order to manage their blood glucose (BG) concentration and mitigate the long-term effects of a chronically increased BG level. Doing so is challenging and a heavy burden for those patients, which led to efforts of automating (parts of) this task. The trend to automated choice of insulin dosage, e.g. in Artificial pancreas (AP) systems, opens many new possibilities, but also challenges in terms of validation, as the number of freely tunable parameters (e.g. settings in an AP) can be so large that no clinical trial can assess the efficiency of all possible choices. As a consequence, computer model based trials - so called in silico evaluations - are getting increasingly popular. In recent times, several authors have tried to improve the quality of in silico evaluations by using “Deviation Analyses”, a term used to refer to methods that extrapolate the effect of a modified insulin therapy using real measurement data together with simple, linear models of insulin action. However, due to the inherent linear model assumption in all methods proposed so far, large deviations compared to the insulin dosing scheme of the recorded data can lead to unphysiological results, e.g. to negative values in the computed glucose traces. Against this background a new, nonlinear methodology is proposed which effectively avoids the common pitfalls of linear Deviation Analyses approaches, i.e. the constant mode of insulin action.

A probabilistic framework for blood glucose control in diabetes

Diabetes is a chronic disease that frequently requires administration of insulin. The choice of the right amount of insulin is usually done by the patient, but both too large and too small amounts of insulin can have dire and even life threatening effects. Against this background, there is a constant trend in developing systems which help the patient or even completely take over the decision, as the artificial pancreas. Both manual and assisted therapy try to stabilize the blood glucose (BG) value around a given safe target, typically around 100 mg/dl. However, even these systems have to work on the basis of some imprecise information by the patient, in particular on the amount of carbohydrates in the meals. Much work has been done to improve this estimation and to provide better models of the insulin/glucose metabolism in spite of the natural intra-patient variation. Differently from that, this paper proposes a different framework, in which the unavoidable uncertainty is modeled in probabilistic terms and the control goal is defined not in terms of proximity to a specific BG target but as keeping the risk of leaving the euglycemic range under a given threshold. This is achieved by a Markov chain model based on BG regions. The degree of freedom gained by this problem relaxation can be used for other purposes, e.g. the minimization of total insulin intake, as shown in some in silico examples.