Analysis of Enhanced Prediction Algorithms for Time Lag Compensation in CGM Systems

The goal of prediction in CGM systems is to compensate the time lag between the glucose level in the blood (BG) and the glucose level in the interstitial fluid (IG). Additionally, data-processing performed on raw current signals can introduce additional time-lags. Since the overall time delay based on these factors is of the order of magnitude of about 10 minutes, short-term prediction models are used to deal with this phenomenon in order to increase the overall precision of CGM systems with respect to the reference device. The available data of clinical studies consists of 176 records. Each record contains data of one patient measured over a period of 7 days. In our analysis we compared the performance of a sensor equipped with different prediction algorithms. The following linear and nonlinear prediction models have been considered in this analysis: default built-in two-compartment model of the manufacturer’s state of the art (SOA) algorithm, a two-compartment model with the derivative calculated by continuous wavelet transform (CWT), global and patient-specific autoregressive linear (GARX, PSARX) and nonlinear (GNARX, PNARX) models of different orders. The results displaying the performance of the considered methods in terms of MARD are presented in Table 1. It can be concluded that the global ARX model of order 2, GARX(2), and the CWT-based two-compartment model perform best in terms of MARD. The higher order linear as well as nonlinear models do not lead to any significant performance improvements. There is also no improvement for the patient-specific models.