In this section, we examine whether inclusion of known or suspected diabetic factors, which are not local predictors in themselves, enhances the local predictive power of the mfERG model. These factors were chosen because they are routinely obtained at each visit and do not require additional laboratory tests. The six variables were age, gender, diabetic eye status (with NPDR or without at baseline: hasRet), duration of diabetes (dmDuration), blood glucose level at the initial visit (bloodGlucose), and diabetes type (type 1 or type 2: diabType). The parameters age, dmDuration, and bloodGlucose were continuous variables on interval scales, whereas gender, hasRet, and diabType were binary (yes/no) variables. In the analysis, female gender, diabetes with baseline retinopathy, and type 1 diabetes were defined as 1, and the counterparts of the three measures were defined as 0.
As is standard, we first examined the association of each variable alone with retinopathy development, using univariate logistic regression.
Table 2 shows that whereas
age,
gender,
diabType, and
bloodGlucose were not significant predictors,
dmDuration and
hasRet had significant power to predict the onset of new retinopathy (
P < 0.05). As expected, the regression coefficients for
dmDuration and
hasRet were positive, indicating that a longer duration of diabetes or the presence of retinopathy at baseline, increased the probability of new retinopathy in the eye within 1 year.
Next, a preliminary multivariate model was established based on the variables that were shown in the univariate analysis to be significantly associated with the occurrence of future retinopathy:
mfergIT,
dmDuration, and
hasRet. The
P value for these variables in the preliminary multivariate model were 0.003, 0.109, and <0.001, respectively. These variables were included in the next stage of model building, because they all have a
P < 0.2. This criterion is chosen based on practical experience that a variable with
P < 0.2 provides some predictive power without adding a significant amount of variation.
31 Variables
bloodGlucose,
age,
gender, and
diabType, which were not significant predictors in the univariate analyses, were added back to the preliminary multivariate model one at a time, in the order of their
P value in the univariate analysis, from low to high (i.e., stronger factors first) to assess their additional contributions.
Age (
P = 0.75),
gender (
P = 0.96), and
diabType (
P = 0.61) did not provide significant information to the model prediction, and so they were excluded from the final model, whereas the variable
bloodGlucose (
P = 0.17) met our criterion of
P = 0.2, and was included. Our final estimated multivariate model to predict the local sites of retinopathy was formulated as
\[\mathrm{log}(\frac{Pr}{1\ {-}\ Pr})\ {=}\ {-}6.78\ {+}\ 0.32\ {\cdot}\ mfergIT\ {+}\ 3.84\ {\cdot}\ hasRet\ {+}\ 0.14\ {\cdot}\ dmDuration\ {+}\ 0.005\ {\cdot}\ bloodGlucose.\]
The odds ratios for all the variables in the final model were greater than 1 (
Table 3 , column 4), indicating that the variables all correlated positively with the development of new retinopathy, though the odds ratio of variable
bloodGlucose does not reach significance. With all other variables fixed, the odds ratio of development of new retinopathy is 1.15 for each 1-year increase in duration of diabetes (
dmDuration). The corresponding odds ratio for 5-year increment in duration is 2.01 (95% confidence interval: 1.09–3.70). The odds ratio for variable
hasRet is large (46.4), suggesting that the presence of diabetic retinopathy is a strong predictor of future retinopathy in diabetic persons, even in a short period. However, caution should be used, given its large confidence interval. After adjustment for the other variables, the local predictor,
mfergIT is still significant, with an odds ratio of 1.38 for a unit increase in the mfERG implicit time
z-score.
Figure 2B shows the ROC curve for the final model, which has an AUC of 0.90 (SE = 0.008;
P < 0.001) compared with an AUC of 0.80 obtained from the model with mfERG only. For a cutoff of
Pr = 0.4, the model provided a sensitivity of 86% and a specificity of 84%, which are substantially higher than the model with
mfergIT as the sole predictor, reflecting that the final multivariate model more accurately predicts new diabetic retinopathy at specific retinal sites.