Study subject characteristics are listed in . There was no significant influence of age (p= .345), body mass index (p= .092), years since diagnosis (p= .277), or sex (p= .816) on system accuracy, and the data were homogeneous between clinical sites. There was also no significant influence of insertion site (p= .253), but there was a significant influence of glucose range (p< .0001), with accuracy improving as glucose concentration increases. The average lag time was 9.6 min.
The relationship of normalized sensitivity with implant time ( ) displays some depressed sensitivity values during the early hours. During the first 10 h, 11.3% of the normalized sensitivity values were more than 30% low compared to 3.0% during subsequent times. After the first day, the sensitivity was stable, with an average total increase of 1.8% during the following four days by linear regression analysis (p= .020). In the first day, the CGM values were blanked 14.1% of the time.
Figure 1 Normalized sensitivity as a function of time (n = 9125). Values of normalized sensitivity include BG tests that do not meet the screening limits for calibration. The line is a statistically significant linear regression for days 2–5 (p = .020) (more ...)
In and , the normalized sensitivities were evaluated for BG values that meet the calibration screening requirements (60–400 mg/dl and rate of change of ±3.5 mg/dl/min). The normalized sensitivity as a function of rate ( ) demonstrates little influence of rate on calibration accuracy. An average calibration bias of ±0.8% for rates of ±3.5 mg/dl/min was calculated by linear regression analysis, but the regression was not statistically significant (p= .060). The sensitivity check makes it unlikely that the outlier values in this plot would be used for calibration. The variability in normalized sensitivity decreased as glucose increased ( ), and calibration accuracy would not be degraded in the 300–400 mg/dl glucose range.
Figure 2 Normalized sensitivity for BG tests that meet screening limits for calibration as a function of rate (n = 8584). The line is a nonstatistically significant linear regression (p = .060) and represents a calibration bias of ±0.8% for rates of ±3.5 (more ...)
Figure 3 Normalized sensitivity for BG tests that meet screening limits for calibration as a function of glucose (n = 8584). The variability in calibration decreases with increasing glucose. The sensitivity check makes it very unlikely that the outlier values (more ...)
The Point error grid analysis (EGA), which is the part of the CG-EGA that evaluates individual glucose values, indicated 80.3% in zone A, 18.0% in zone B, and 1.7% in zones C–E ( and ), whereas the Rate EGA, which is the part of the CG-EGA that evaluates the rate of change in glucose, indicated 79.1% in zone A, 16.1% in zone B, and 4.8% in zones C–E. Zone A values are considered clinically accurate, zone B are benign errors, and zones C–E are clinical errors. The combined grids that assess all information provided by CGM rated all data 93.7% “clinically accurate,” 3.6% “benign errors,” and 2.8% “clinical errors.”
Performance Characteristics of the FreeStyle Navigator CGM with the TRUstart Algorithma
Figure 4 Point EGA of the FreeStyle Navigator CGM clinical data. Zone A represents clinically accurate (80.3%); zone B represents benign or no treatment errors (18.0%); zone C represents overcorrection errors (0.2%); zone D represents a potentially dangerous failure (more ...)
The hypoglycemic range was significantly less accurate than the euglycemic and hyperglycemic ranges, with 56.2%, 95.6%, and 94.7% of values deemed “clinically accurate,” respectively.
For the entire glucose range, the mean and median ARD were 14.5% and 10.7%, respectively, and the ISO accuracy was 76.9%. The decrease in ARD and increase in ISO accuracy with increasing glucose indicated improved concordance with higher reference glucose values ( ).
During the time when CGM results were suppressed, the glucose readings exhibited an expected low bias ( ), which resulted in less accurate glucose values: Point EGA zone A 57.5%, zone B 42.0%, and zones C–E 0.6%; mean and median ARD 25.3% and 21.3%, respectively; and ISO accuracy 47.8%. The overall clinical accuracy by CG-EGA, however, was not as severely affected: 91.7% “clinically accurate,” 6.3% “benign errors,” and 2.0% “clinical errors.”
Point EGA of the FreeStyle CGM data during the period when readings were suppressed. The modification of the Point EGA described in was also applied to this figure.
There were no significant differences in clinical accuracy when the first 10 h were compared to subsequent hours with “clinically accurate” values 92.6% and 94.2% (p= .015), respectively ( ). (p values < .006 are statistically significant in this analysis.) When the data were stratified by glucose range, there was one statistically significant difference in the euglycemic range, but the magnitude of the difference was inconsequential, 95.2% versus 96.1% “clinically accurate” (p= .004). There were also no differences in the mean and median ARD when the first 10 h were compared to subsequent hours [mean ARD 13.4% versus 14.9% (p= .267), median ARD 10.0% versus 10.9% (p= .443), respectively].
Performance Characteristics with the TRUstart Algorithm in the First 10 h and Successive Time and Comparison of the TRUstart Algorithm to the First-Generation Algorithma
When TRUstart was not calibrated until after 10 h and compared to the first-generation software version ( ), the clinical accuracy was nearly the same, with 94.6% and 94.8% “clinically accurate” (p= .527). There was no significant difference in mean ARD, with values of 14.9% versus 14.8% (p= .725), or in median ARD, with values of 10.9% versus 10.5% (p= .294), for TRUstart versus the first-generation algorithm, respectively. There were also no significant differences in performance when the data were stratified by glucose range.
The performance analysis of the low alarms shows the trade-off of hypoglycemia detection with false alarms as the alarm setting increases ( ). In the alarm range of 70–85 mg/dl, the detection increased from 60.2% to 86.1%, whereas alarms occurring with the BG >85 mg/dl increased from 19.1% to 23.2%. High alarm performance was similar throughout the range of 140–300 mg/dl. At the alarm threshold, detection ranged from 83.7% to 91.8%, and false alarms were 35.9–70.1%, but when ±20% is allowed for clinical accuracy, detection of BG 20% higher than the alarm setting varied from 97.4% to 100%, and false alarms 20% lower than the alarm setting were between 3.3% and 7.5%. There was a paucity of data with BG levels >300 mg/dl for performing the alarm analysis at 300 mg/dl ( ).
Detection of Hypoglycemia (Glucose < 70 mg/dl) with FreeStyle Navigator CGM Low Alarms at Various Settingsa
Detection of High Glucose Levels between 140 and 300 mg/dl with FreeStyle Navigator CGM High Alarmsa