Fig 1.
A combination of cross validation and held-out test set was used for the development and evaluation of the models.
Table 1.
Different model architectures trained for 1,000 epochs.
Table 2.
Baseline demographic factors at patient eye level.
Fig 2.
Model development and evaluation.
Model development are shown in Panels A and B, Model evaluation on the test set are shown in Panels C and D. A) Model architectures tested with lowest validation Point-wise Mean Absolute Error (PMAE) shown with each colored line representing one of the ten-fold internal cross validation datasets. B) Data combinations tested with every possible combination of age, gender, eye, and test number (HVF #). Model evaluation using held out test set with Bland-Altman plot (C) between Mean Deviation (MD) of the AI predicted and actual future HVFs, (r = 0.92, Adjusted R2 = 0.84, p < 2.2 x 10−16), color shaded by MD of input (earlier) HVF and sized by time interval. The blue and red lines in (C) represent the mean and 95% CI, respectively. D) Average PMAE for each time interval of held-out test set with 95% confidence intervals as error bars.
Fig 3.
Representative examples from the held-out test set, comparing actual and artificial intelligence (AI) predicted Humphrey Visual Field (HVF).
The input HVF is used for predicting HVF at the respective designated time points. A variety of different starting HVFs are shown ranging from a hemifield defect to very severe glaucoma with the corresponding Point-wise Mean Absolute Error (PMAE).
Fig 4.
Serial comparisons between actual and artificial intelligence (AI) predicted Humphrey Visual Field (HVF).
The single input HVF is used for multiple predictions at different time points. Point-wise mean absolute errors for the left panel from top to bottom were 3.8, 3.8, 5.1, and 4.0 dB respectively, and for the right panel from top to bottom were 2.3, 2.5, 3.2, and 3.2 dB, respectively.
Table 3.
Held-out test-set performance with deep learning compared to linear models using pointwise mean absolute error (PMAE).