Table 1 shows the demographic characteristics of the training and test datasets. The mean age (± standard deviation [SD]) was 45.1 ± 18.0 and 52.9 ± 18.3 years in the training and test datasets, respectively. The mean logMAR VA was 0.18 ± 0.32 and 0.31 ± 0.46 in the training and test datasets, respectively.
In the training dataset, foveal sensitivity was significantly correlated with the logMAR VA (
P < 0.001, linear mixed model); the logMAR VA was 1.88–0.051 (± 0.0053) × foveal sensitivity (
Fig. 3A). The MD value was also related to the logMAR VA (
P = 0.002, linear mixed model); the logMAR VA was −0.012–0.017 (± 0.0055) × MD (
Fig. 3B).
The comparison of the TD values in each sector (IN, ON, OUT1, OUT2, OUT3, OUT4, OUT5, and OUT6) in all 93 eyes is shown in a boxplot (
Fig. 4) and
Table 2. There was a significant difference in all comparisons across the sectors (
P < 0.05, Tukey's test and linear mixed model); the TD values decreased significantly from IN toward OUT6.
In the training dataset, 405, 328, 365, 398, 458, 464, 411, and 639 test points were assigned to the regions of IN, ON, OUT1, OUT2, OUT3, OUT4, OUT5, and OUT6, respectively. These values were 273, 247, 262, 326, 348, 371, 333, and 696, respectively, in the testing dataset. With the training dataset, at each sector, the association between the TD values from the values of age, logMAR VA, and the areas of IN and ON was as follows:
The comparison between the absolute prediction error with (7.6 ± 5.6 dB) and without (8.7 ± 15.4 dB) the FAF ring (binarization method) is shown in
Figure 5. There was a significant difference between the two values (
P < 0.001, linear mixed model).
The comparison of the absolute prediction error values with the FAF ring (binarization method) prediction across the eight sectors is shown in
Figure 6 and
Table 3. The IN and ON values were significantly smaller than those at all the six OUT sectors (
P < 0.05, Tukey's test and linear mixed model). The values of OUT1 were significantly smaller than those of OUT2, OUT3, OUT4, and OUT5
(P < 0.05).
Among the 68 test points from 41 eyes (2788 test points), there were 95 test points, which had the absolute prediction error > 20 dB. These test points distributed wide in each area 2.2% in IN (6 test points), 3.2% in ON (8), 2.3% in OUT1 (6), 3.1% in OUT2 (10), 3.4% in OUT3 (12), 4.0% in OUT4 (15), 4.2% in OUT5 (14), and 3.4% in OUT6 (24). There were 10 eyes that had at least one test point with the absolute prediction error > 20 dB. Among these, there were 3 eyes with > 10 such test points (30 test points in 2 eyes and 12 test points in 1 eye). Analyzing the accompanied OCT images of these eyes, it was suggested that the ellipsoid zone was disrupted in the IN area and remained in the OUT areas in these eyes (see a typical case in
Fig. 7). The remaining 7 eyes had 3.3 ± 2.4 (range: from 1 to 8) test points with the absolute prediction error > 20 dB, otherwise nil; such finding was not observed in these eyes.
There was a significant association between absolute prediction error and the values of logMAR VA (coefficient = -2.00, standard error = 0.94, P = 0.039, linear mixed model) and also TD value at each point (coefficient = −0.093, standard error = 0.011, P <0.001), suggesting that the absolute prediction error increased with the deterioration of these visual functions.