Clinical risk measure for variability analysis and real-time prevention of hyper/hypo-glycaemic episodes from continuous glucose monitoring time-series
Year: 2012
Type: dissertation
Abstract: Diabetes is a chronic disease characterized by impairments both in the secretion/action the insulin hormone. As discussed in Chapter 1, insulin acts on the metabolism of glucose, since it enhances glucose uptake by the tissues and suppresses hepatic glucose production. Diabetes is treated with a combination of insulin infusions, drugs and physical exercise. Unfortunately, diabetes therapy is quite difficult to tune individually, and in diabetics glycemia (glucose concentration in blood, BG) often exceeds the normal range (70-180 mg/dl). Conditions of hypoglycemia (BG<70 mg/dl) and hyperglycemia (BG>180 mg/dl) are threatening for the patient in the short and long term, respectively. Severe hypoglycemia causes a suffering of the tissues, and in particular of the brain (because glucose is their most important nutrition factor), and may lead to coma. Hyperglycemia induces complicances such as retinopathies, nephropathies and various cardiovascular diseases. For these reasons, it is important to achieve tight control of the glycemia. Glycemia can be monitored in diabetic patients mainly though 2 strategies. The first is based on few measurements a day performed via fingerpricks (self monitoring blood glucose, SMBG). The second is based on the recently developed continuous glucose monitoring (CGM) devices, which allow collecting a CGM measurement every 1-5 minutes, hence allowing to track the physiological glucose dynamics with finer detail. In order to achieve a good therapy it is crucial to be able to alert the patient promptly in order to allow time to the therapeutic actions (sugar ingestion, insulin administration) to be effective. Moreover, it is important to characterize the individual features of the patient by analyzing the so-called glucose variability. As discussed in Chapter 2, several indexes and transformation of the glucose scale may be used to characterize glucose variability. Along with several indexed based on statistical analysis of the glucose time-series, it has been suggested by Kovatchev (31) that the study of glucose concentration time-series should take into account that the glycemic range is asymmetric and that the distribution of glucose concentration values is skewed within the range. Kovatchev proposed a punctual transformation of the glucose scale (static risk, SR), assigning to each glucose level a specific risk value. Most indexes available in the literature for the analysis of glycemic time-series were developed based on SMBG measurements.Today, the availability of CGM signals opens the possibility to embed glucose dynamics information in the analysis of glucose time-series. In this thesis the concept of clinical risk and of glucose variability will be developed explicitly embedding the trend of the glucose signal in a new ”Dynamic” Risk Function, which will be used as basis for online generation of alerts, and for the definition of new variability indexes. Such indexes will be then included in a multivariate analysis for the parsimonoius description of glucose variability. Part II - Conceptual Design and Implementation In this thesis the Kovatchev’s risk function (Static Risk) will be further developed to be adapted and improved exploiting the natural information brought by CGM signals. In fact, these time-series are sampled more frequently, and allow an estimate of the first time deriva- tive of the glucose signal. In Chapter 3, a Dynamic Risk (DR) function will be defined, which explicitly modulates the risk function through the information of the first-time derivative. In particular, the DR function will be structured in such a way that the actual SR associated with the glucose level is amplified or damped according to the time-derivative: if the glucose signal is heading toward a critical region (hypo or hyperglycemia) the risk is increased, while it is reduced if the glycemia is recovering to safety values. As far as the algorithms for the DR implementation are concerned, in Chapter 4, particular care will be put in the strategy to evaluate the first time derivative, since the measurement noise, always present in the CGM signals, tends to be amplified by the derivative. In this work this task will be carried out by a method based on regularized deconvolution which is suitable for online application, e.g. to be used in algorithms for the generation of hypo and hyperglycemic alerts. 2 Part III - Application of DR for Hypo/Hyperglycemic Alert Generation DR is intrinsically predictive, allowing anticipating threshold crossings and hence alerting the patient with relative temporal gain to take adequate therapeutic actions. To show this, in Chapter 5the DR function will be applied both on simulated and on real CGM data collected by several type of sensors. Results on simulated and real data show that hypoglycemic events can be anticipated of 12 to 18 minutes by simply evaluating the DR of the glycemic time- series. Moreover, in Chapter 6 a hypo alert generation strategy based on the combination of a short-term prediction of the CGM signal via Kalman Filter and DR will allow anticipating the event of more than 20 minutes with relatively small number of false hypo/hyperglycemic alerts. Part IV - Application of DR for the (Parsimonious) Description of Glucose Variability Dozens of indexed evaluated from CGM/SMBG time series have been proposed to characterize and assess glucose variability. None of them explicitly exploit the continuous feature of CGM signals, in fact, none explicitly includes the information of time-derivative as a variability factor. Based on the Dynamic Risk Space (DRS), a phase plot where glucose trajectories (glucose value vs time-derivative) can be effectively shown to evaluate glucose variability and control, several new indexes will be developed in Chapter 7 to analyze potentially useful features of glucose signals. Then, starting from the tens of indexes proposed in the literature and those developed in the present thesis, a multivariate technique, the Sparse Principal Component Analysis will be applied in Chapter 8 to a set of 48 indexes evaluated on a dataset of 60 CGM signals, in order to understand which is the best combination for the description of glucose variability. SPCA indicates that 5 components are relevant for the analysis of the dataset variance and that indexes explicitly considering information about the first time derivative are relevant for the description of glucose variability
Author: Stefania Guerra
Language: English
Subfield: Endocrinology, Diabetes and Metabolism
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W2277216406