Profiling intra-patient type I diabetes behaviors
Introduction
Integrated systems using continuous glucose monitors (CGM) and continuous subcutaneous insulin infusion (CSII) have had a significant impact in type 1 diabetes (T1D) management. The use of CSII-CGM systems was shown to reduce HbA1c without increasing the time spent in hypoglycemia [1]. Subsequently, CSII-CGM systems with automatic features such as automatic suspension of insulin delivery during hypoglycemia and predictive low glucose suspend have shown to reduce nocturnal hypoglycemia without increasing mean glucose substantially [2], [3]. Furthermore, the recent advances in CGM have led to more robust and portable devices which have demonstrated their value in improving the glycemic control working with closed-loop algorithms [4], [5]. The integration of data originating from sensor-based systems and electronic health records combined with smart data analytics methods and powerful user centered approaches enables the shift toward preventive, predictive, personalized, and participatory diabetes care [6]. However, the limited capacity of current solutions to process the data extracted from glucose monitors limits the development of enhanced diabetes management solutions.
Nowadays, companies commercializing CSII-CGM systems offer software platforms that provide tools to upload data from sensors, to share information with physicians and give patient support to diabetes management. These platforms are currently integrating methodologies to support treatment management by analyzing the data generated by the CGM. A notable example is the Medtronic software CareLink Pro/Personal® which provides functionalities to analyze BG values depending on the events occurring at specific periods, assisting patients and physicians with recommendations, which can lead to an improved insulin therapy. However, these features are still unable to detect the majority of scenarios with high impact in the blood glucose variability such as exercise, menstrual period, seasons, diet disturbances, or habits among other factors. Thus, the complexity of generating accurate treatments is accentuated when we have to deal with disturbances that can arise in the same patient. This intra-patient variability has different effects on BG levels altering the patterns usually generated by patients with T1D. The ability to identify such patterns would allow generating temporal profiles, which in turn can assist in identifying the causes of poor glycemic control and aid in the therapy adjustment. With the objective of analyzing similarities among time series we use a hierarchical clustering methodology based on an innovative similarity measure.
In the medical field, clustering methodologies are used in problems such as identifying effective treatments, recognition of diseases and detection of best practices. Broadly speaking, clustering methodologies are used to discover and study the macroscopic structure and relations between objects. The hierarchical clustering family of methods is defined by the method used to compute distances and a linkage criterion. A well-known distance metric is the normalized compression distance (NCD). The NCD is a compression-based similarity distance based on the Kolmogorov complexity [7]. This determines the similarity in terms of information distance between pairs of objects according to the most dominant common features. Previous studies have demonstrated that the NCD is a reliable tool for classification on a number of domains (see Ref. [8]). Furthermore, NCD has been applied successfully in many areas such as classification of genomes [9], protein structure comparison [10], genotyping [11], tumor subclassifications [12] or virus detection [13] among many others. In this study, we are presenting a new methodology that may provide a novel tool capable of extracting information on glucose profile of patients using CGM. This tool can help physicians and patients to detect patterns of poor glycemic control easily, providing useful information for the therapy adjustment.
Section snippets
Materials and methods
We have used a data mining technique aimed to identify time series following different patterns despite the seemingly uncorrelated behavior of the BG series. The clustering process can be summarized as a method that builds a binary tree from individual elements by progressively merging the clusters containing the two closest elements. Thus, we consider a set of N time series to be clustered and a distance matrix, also called dis-similarity matrix, with N*N measurements. The distance measure
In silico experiments: profiling results
Fig. 4 presents the results of clustering tests in the 10 virtual patients through the 3 scenarios. Red nodes represent the basic scenario, light grey nodes the exercise scenario and dark grey nodes the exercise scenario with corrective measures. The results obtained using the clustering methodology display significant clusters of two types of series: well controlled and poorly controlled BG levels regardless of the scenario. For instance, the graphs of Patient 5, 7 and 8 clustered basic
Discussion
The experiments reported in this article have demonstrated the feasibility of clustering different behavior profiles, identifying patterns on different days for the same patient.
Unlike other works [22], [23], where a set of events and possible behavior are predetermined and time series are classified according to patterns, in this study no behavior is predetermined. The system groups days in clusters according to a general similarity criterion. Further analysis determines, for each patient, the
Conclusions
A wide offer of platforms for CSII-CGM systems are currently integrating methodologies to support treatment management by analyzing the data provided by CGMs. The development of software for personal therapy management makes it easy for insulin pump and glucose monitor users to optimize therapy for improving their glucose control. Our methodology fits in this increasing trend of software functionalities where there is a lack of tools for detecting the majority of scenarios with a high impact in
Conflict of interest
No competing financial interests exist.
Acknowledgment
This work was partially supported by the Spanish Government through the grant DPI 2013-46982-C2-R and the People program (Marie Curie Accions) of the European Union Seventh Framework Programme (FP7/2007–2013) with the agreement no. 600388 of the REA and the Agència per a la Competitivitat de l'Empresa (ACCIÓ).
References (23)
- et al.
Postprandial blood glucose control using a hybrid adaptive PD controller with insulin-on-board limitation
Biomed. Signal Process. Control
(2013) - et al.
Intelligent analysis of clinical time series: an application in the diabetes mellitus domain
Artif. Intell. Med
(2000) - et al.
The use and efficacy of continuous glucose monitoring in type 1 diabetes treated with insulin pump therapy: a randomised controlled trial
Diabetologia
(2012) - et al.
Predictive low-glucose insulin suspension reduces duration of nocturnal hypoglycemia in children without increasing ketosis
Diabetes Care
(2015) - et al.
Threshold-based insulin-pump interruption for reduction of hypoglycemia
N. Engl. J. Med
(2013) - et al.
Closed-loop artificial pancreas systems: engineering the algorithms
Diabetes Care
(2014) - et al.
Nocturnal glucose control with an artificial pancreas at a diabetes camp
N. Engl. J. Med
(2013) - et al.
A review of emerging technologies for the management of diabetes mellitus
IEEE Trans. Biomed. Eng
(2015) Three approaches to the quantitative definition of information
Probl. Inf. Transm
(1965)- et al.
Blind optimisation problem instance classification via enhanced universal similarity metric
Memet. Comput
(2014)
An information-based sequence distance and its application to whole mitochondrial genome phylogeny
Bioinformatics
Cited by (16)
Artificial intelligence-based decision support systems for diabetes
2020, Artificial Intelligence in Precision Health: From Concept to ApplicationsAlgorithmic clustering based on string compression to extract P300 structure in EEG signals
2019, Computer Methods and Programs in BiomedicineCitation Excerpt :The NCD has been applied to numerous research areas because of its parameter-free nature, wide applicability and leading efficacy. For example, among others, it has been applied to data mining [25], profiling diabetes behaviors [26], music clustering [27,28], document clustering [29,30], image analysis [31], bird song species classification [32,33] or video activity recognition [34]. Also, this measure has proved to have a high noise resistance [35] which makes it very appealing for our problem.
Statistical Packages and Algorithms for the Analysis of Continuous Glucose Monitoring Data: A Systematic Review
2024, Journal of Diabetes Science and TechnologyA 2-dimensional model framework for blood glucose prediction based on iterative learning control architecture
2023, Medical and Biological Engineering and ComputingA robust model predictive control framework for the regulation of anesthesia process with Propofol
2021, Optimal Control Applications and MethodsLeveraging Artificial Intelligence to Improve Chronic Disease Care: Methods and Application to Pharmacotherapy Decision Support for Type-2 Diabetes Mellitus
2021, Methods of Information in Medicine