8.2 Experimental Results for Glucose Correction
8.2.1 Error Correction for Glucose Measured by the Handheld Device
The procedure of error correction for glucose measured by a handheld device is described in Section 6.2. In this section, we evaluate its performance on a specific handheld device that is GlucoDr meter. The plot of paired-differences (also residuals) of glucose values measured by the handheld device minus the values obtained by YSI2300 glucose measurements against changes of collected hematocrit density is shown in Fig. 8.3. From this figure, we can see that there is a dependence of hematocrit on residuals. In addition, using the test statistic for the slope given by
0
, 0
slope
slope
slope slope
t − slope 0
= =
σ (8.3)
and using the P-test, we see that the slope value is significantly different from 0 (p<0.01). Note that tslope and P-test is a measurement of hematocrit dependency.
Therefore, we can conclude that the effect of hematocrit on glucose measurement of handheld device (GlucoDr) is significant which is consistent with the previous
reports.
From Fig. 8.3, we also see that four patterns corresponding to residuals larger than 60.
These patterns can be considered as outliers, which may happen due to significant defect of glucose sensor or special medication that may cause large error on glucose measurement. Thus, we omitted them from the dataset. The remaining data consisting of 187 patterns were used to evaluate our approaches proposed in Chapter 6. Thirty percent of data set was used for training in order to find the function g mapping from hematocrit to residuals, and the remaining 70% was used to evaluate the models. The mapping function g was determined as
( jm) 0.6253 jm 29.1208
g HCT = − HCT + , (8.4)
where is hematocrit estimated by linear methods described in Chapter 5.
From (8.4), (6.2) or (6.9), the error correction for glucose measurements with the handheld device can be done easily.
m
HCTj
The RMSE for handheld device on the test set without error correction is 16.4149, while that after error correction is 13.7418. The paired-differences corresponding to glucose measurements by handheld device without error correction are plotted in Fig. 8.4. The t-test for slope is -3.846 (p-value<0.001) which shows existence of hematocrit dependency.
However, Fig. 8.5 describes the paired-differences corresponding to glucose measurements by handheld device after error correction, and the t-test for slope has
P-value of 0.26. These results show that the effects of hematocrit are reduced after
error correction.
Figure 8.4 The paired-differences of a testing set corresponding to glucose
measurements by handheld device without error correction. The dependency of
hematocrit on residuals is significant.
Figure 8.5 The paired-differences of a testing set corresponding to glucose
measurements by handheld device after error correction. The dependency of
hematocrit on residuals is reduced significantly.
The criterion proposed by the National Committee for Clinical Laboratory Standards (NCCLS) is that error tolerances of ±15mg/dL for glucose levels ≤100 mg/dL and ±20% for glucose levels > 100 mg/dL. At least 95% of glucose meter
measurements should fall within these error tolerances. In our experiments, we evaluate the proposed approaches with several NCCLS-based criteria, which are the error tolerance of ±15mg/dL for glucose levels ≤100mg/dL and p% for glucose levels
>100mg/dL. Table 8.3 presents comparison results (within the error tolerance) of error correction for different values of p.
Table 8.3 Comparison results for different criteria of error tolerance.
p(%) Before error
correction After error
correction
15 91.60% 94.66%
16 94.66% 95.42%
17 94.66% 96.18%
18 95.42% 96.95%
19 96.18% 96.95%
20 96.95% 96.95%
We can see that both approaches, before and after error correction, have 96.95% of glucose measurements within the error tolerances with p of 20%. This satisfies the criteria proposed by the National Committee for Clinical Laboratory Standards. However, the error correction provides the improved performance at levels of error tolerance with p values from 15% to 19%. The comparison of handheld meter with the primary reference instrument YSI2300 for the case of
p=18% is shown in Fig. 8.6, which is a plot for glucose measurements by handheld
device against that by YSI2300. It also shows the improved performance after error
correction with the error tolerance of p=18%.
(a) before error correction
(b) after error correction
Figure 8.6 Comparison of glucose results from handheld meter and the primary
reference instrument, YSI 2300: (a) before error correction and (b) after error
correction.
8.2.2 Error Correction for Glucose Computed Using a Single Transduced Current Point
As described in section 6.3, we can imitate glucose measurements of handheld meter, which glucose value tm is estimated from a single point of the transduced current curve. The proposed equation is given by (6.13) in general and (6.17) for the data set used in our study. Let us call this method as the single-point based measurements.
The estimated glucose values are then corrected by reducing effects of hematocrit as described in section 6.2.
The collected data consisting of 191 patterns was divided into the training set (30%) to find the function approximation of residuals with respect to hematocrit, and the test set to evaluate our proposed methods. The RMSE without error correction was 15.6717, which is slightly smaller than that of handheld meter. The t-test of slope for relationship between hematocrit levels and residuals was -3.65 or P-value
<0.001. This confirms that there was effect of hematocrit levels on glucose measurements. The RMSE after error correction was reduced down to 13.2255 and t- test of slope is -1.53 or P-value >0.1. This shows that the effect of hematocrit was reduced after error correction and thus the RMSE was improved.
The comparison results (within the error tolerance) for different criteria of error tolerance are shown in Table 8.4. We can see that the number of patterns within error tolerance after error correction is significantly improved for the criteria of 15%, 16%
and 17% of p. Both approaches, with and without error correction, have 97.76% of glucose measurements within error tolerances of p=20%, which satisfies the criteria proposed by the National Committee for Clinical Laboratory Standards. In addition,
they are better than the results of error correction for glucose measured by handheld device as shown in the previous section.
Table 8.4 Comparison results on different criteria of error tolerance
p(%) Before error
correction
After error correction
15 90.30% 94.78%
16 92.54% 97.01%
17 94.77% 97.01%
18 97.01% 97.01%
19 97.01% 97.76%
20 97.01% 97.76%
Table 8.5 Comparison results on RMSE of approaches.
Approaches RMSE
Before error correction 16.4149 Commercial
handheld device After error
correction 13.7418 Before error
correction 15.6717 Single point based After error
correction 13.2255 SLFNs 12.0694