The prediction of glycemic profiles based on continuous glucose monitoring contributes to enhance diabetes management, helping to prevent hyper and hypoglycemic events or integrated in artificial pancreas. This paper proposes a new set of parameters to assess glucose predictors by integrating accuracy, delay and noise parameters.
The new metrics complement the frequently used root mean square error (RMSE), with the following: the delay-compensated RMSE (RMSEDC), the delay-compensated RMSE into the 25-75% of the extremes (RMSEDC-25/75), the overshooting/undershooting (OS/US); the mean delay (MD25/75), the peak delay (PD); and the noise energy ratio (ENOISEnorm). We have applied the parameters to 5 synthetic test-profiles (TP) that emulate prediction profiles. TP_1-2 are built by adding a constant delay (15&10 min respectively) to an original CGM profile, and artificial deformations at the local maxima (TP_2). TP_3-4-5 correspond to a 10 min delayed CGM profile, adding differet types of noise (20dBW WGN in TP_3, plus extra high&low frequency noise in TP_4-5 respectively).
Figure. Synthetic test profiles
The RMSE is not able to discriminate among TP_1-2-3, having clearly distinct performance (Figure 1) (RMSE:13.0-12.2-13.9), while the proposed metrics reveal their differences (RMSEDC:0-7.0-8.9, RMSEDC-25/75:0-0-9.4). MD25/75 detects the delay in 4 out of 5 TPs and PD reveals if that delay is homogeneous. OS/US analysis allows a critical insight about the absence of false-alarms/false-negatives. ENOISEnorm reveals the presence of an extra medium-low-frequency noise in TP_4-5 vs. TP_3 (ENOISEnorm:4.27-8.09-9.85).
The proposed metrics allow an in-depth characterization of glucose predictors, helpint to select the optimum prediction technique depending on the specific application.