Parsimonious Description of Glucose Variability in Type 2 Diabetes by Sparse Principal Component Analysis

Background: Increased glucose variability (GV) is considered a risk factor for the development of diabetes complications. To quantify GV, dozens of indices have been developed. In order to limit redundancy, the use of Sparse Principal Component Analysis (SPCA) has been recently assessed in Type 1 Diabetes (T1D), obtaining a parsimonious set of up to 10 indices for describing GV. In this work, we extend the assessment of SPCA to Type 2 Diabetes (T2D) and compare results with those of T1D. Methods: N=27 established GV indices, including SD, MAGE, ADRR and others, are computed on 13 CGM time-series collected by the Guardian RT in T2D subjects and on 16 collected by the SEVEN Plus in T1D subjects. SPCA is used first to determine a reduced data dimension P and, then, to decrease the number of variables from N=27 to M via LASSO estimation of sparse loadings. Results: For both datasets, SPCA selected P=2 principal components (PCs) and M=5 indices for each PC. The subset of indices selected for T2D allowed preserving the 87% of the variance originally explained by all GV metrics, compared to the 67% preserved for T1D. The selected indices are reported in the table. Seven out of the 10 selected GV indices are the same for both datasets. Conclusion: SPCA can be used to extract a parsimonious set of indices describing GV from a large dataset. Some of them seem to be independent on the diabetes type 1 vs 2.