Benjamin Lobo, United States of America
University of Virginia School of Data SciencePresenter of 1 Presentation
A DATA-DRIVEN CLASSIFIER OF DAILY CONTINUOUS GLUCOSE MONITORING (CGM) PROFILES
Abstract
Background and Aims
With the proliferation of CGM, massive databases of CGM traces (daily profiles) are constantly growing. A question therefore arises: are all of these profiles substantially different or is there a finite set of distinct daily profiles which sufficiently approximate all possible profiles? We propose a data-driven approach to determine a finite set of “motifs” – representative daily profiles – such that almost any daily profile can be matched to one of the motifs.
Methods
Data: 595 individuals with type 1 or type 2 diabetes (T1D, T2D) participating for 3-6 months in either the International Diabetes Closed-loop (iDCL) Trial or in Dexcom’s DIaMonD study. A set of 226 motifs was constructed by clustering 4,802 (training) profiles from the iDCL Protocol 1 study (T1D) and identifying the motif for each cluster. The representative set of motifs was then tested using profiles from the iDCL Protocol 3 and DIaMonD studies (T1D and T2D), which included a variety of treatment modalities, e.g. daily insulin injections, insulin pumps, and artificial pancreas.
Results
Over 98.8% of the 39,916 testing profiles were successfully classified using the motifs. Each cluster of profiles from the testing data had similar clinical characteristics (e.g., time within or above range) to the corresponding cluster of profiles from the training data.
Conclusions
The finite set of motifs can sufficiently describe almost any daily profile, and the clinical characteristics of each motif are representative of the CGM profiles clustered around it. The motifs can be used for predictive modeling, decision support, or automated closed-loop control.