Forecasting Future Humphrey Visual Fields Using Deep Learning

INTRODUCTION

• Global indices such as mean deviation and visual field index may miss subtle changes because glaucomatous visual-field loss is often focal.Point-wise linear regression was developed to assess changes at individual test locations.Source: INTRODUCTION · p. 4
• Earlier regression models often assume a constant progression rate, assess test points independently of adjacent points, and require several visual-field tests for accurate prediction.These limitations apply to the described linear and nonlinear models.Source: INTRODUCTION · p. 4
• Deep learning had been applied to ophthalmic classification tasks, but its use for forecasting future findings remained limited.Prior applications included OCT image classification and segmentation and glaucoma diagnosis or classification using disc photographs, OCTs, and visual fields.Source: INTRODUCTION · pp. 4–5
• The study applies a deep-learning generative model to predict future visual fields while preserving spatial information and using minimal baseline visual-field input.The paper frames forecasting as an application beyond classification.Source: INTRODUCTION · p. 5

METHODS

• Consecutive 24-2 Humphrey visual fields from 1998–2018 at the University of Washington were extracted with 54 perimetry values recorded to hundredth-decibel precision.Age, gender, test date, and eye tested were also extracted, and records were de-identified.Source: METHODS · p. 5
• HVFs were divided at the patient level into 80% training and validation data and a 20% held-out test set; ten disjoint patient-level sets were created for 10-fold cross-validation.Within each training, validation, and test set, all possible HVF temporal combinations were binned into 0.5-year intervals.Source: METHODS · p. 5
• The development pipeline selected a model architecture and clinical-input combination before training interval-specific models with transfer learning.The final model weights were frozen at each interval for evaluation on the held-out test set.Source: METHODS · pp. 5–6
• Models mapped an 8x9 HVF input tensor to an 8x9 target HVF tensor using masked mean absolute error; nine architectures were tested, and each model trained 1,000 epochs ten times.The best-performing model was carried forward to subsequent training.Source: METHODS · p. 6
• After selecting the architecture and clinical inputs, models were trained for 0.75- to 5.5-year intervals using weights transferred from earlier intervals.Weights were carried from the first interval and then from the immediately preceding interval; mean predictions across the ten folds were used for held-out test predictions.Source: METHODS · p. 7

RESULTS

• The study extracted 32,443 24-2 HVFs, yielding 1,751,922 perimetry points, without filtering or data cleaning before model training.The HVFs were collected contiguously from 1998 to 2018.Source: RESULTS · p. 7
• CascadeNet-5 had the best network performance among the tested architectures.The model was selected for subsequent training.Source: RESULTS · p. 8
• Including age as an input consistently yielded superior validation performance among the tested clinical-variable combinations.All 16 possible combinations of available predictors were trained.Source: RESULTS · p. 8
• r = 0.92 (adjusted R² = 0.84, p < 2.2 x 10^-16) described the correlation between predicted and actual future mean deviations; the mean difference was +0.41 dB.The correlation is reported in Figure 2C and the mean difference in the Bland–Altman analysis in Figure 2D.Source: RESULTS · p. 8
• For the test set, RMSE was 2.47 dB (95% CI: 2.45 dB to 2.48 dB); another reported value was 3.47 dB (95% CI: 3.45 dB to 3.49 dB).Source: RESULTS · p. 8
• Representative test-set examples showed predictions at different future time points for glaucoma of varying severity and a non-progressing non-glaucomatous hemifield defect.Additional examples used the same input field to predict multiple future time points.Source: RESULTS · p. 8

• The study used more than 1.7 million points from 32,443 consecutive HVFs, including tests regardless of diagnosis or reliability.Source: DISCUSSION · pp. 9, 11
• 2.57 dB was the test-set MAE, and the RMSE was 3.47 dB.Source: DISCUSSION · p. 9
• Earlier approaches could require many tests over years, while constant-rate regression models do not capture varying glaucomatous decay.Source: DISCUSSION · pp. 9–10
• From a single HVF, the model generated pointwise forecasts at half-year intervals up to five years later.Source: DISCUSSION · p. 10
• Examples show appropriate forecasts for both stable non-glaucomatous defects and glaucomatous progression.Source: DISCUSSION · p. 11
• The model used little clinical data, and evaluation came from one academic center with no test-set pairs beyond 5.5 years.The authors identify age as a factor that improved prediction and propose incorporating IOP, medication, and surgical history in future work.Source: DISCUSSION · p. 12

Tables

• Table 1 covers baseline demographics at the patient-eye level, while Table 2 lists model architectures trained for 1,000 epochs.The captions specify table scope but do not provide the underlying cell values.Source: Table 1 · p. 14, Table 2 · p. 15

Figure 1

• The study-design flowchart shows cross-validation combined with a held-out test set for model development and evaluation.The design is presented as a combination of these two evaluation approaches.Source: Figure 1 · p. 16

Figure 2

• r = 0.92 and Adjusted R2 = 0.84 describe the relationship between predicted and actual future HVF mean deviations in held-out evaluation.Figure 2 also presents architecture and data-combination tests, Bland-Altman evaluation, and average MAE by time interval.Source: Figure 2 · p. 17

Figure 3

• Representative held-out examples compare actual and predicted fields across starting conditions from hemifield defects to very severe glaucoma, with corresponding MAEs.The input field is used to predict the field at each designated time point.Source: Figure 3 · p. 18

Figure 4

• Serial examples use a single input HVF to make predictions at multiple time points, with MAEs ranging from 2.3 to 5.1 across the displayed panels.Left-panel MAEs are 3.8, 3.8, 5.1, and 4.0; right-panel MAEs are 2.3, 2.5, 3.2, and 3.2.Source: Figure 4 · p. 19

Supplementary Figure 1

• For each of ten time intervals, predictions from ten models—one per cross-validation set—are averaged before comparison with the actual HVF.The final prediction is the average of the interval-specific model predictions.Source: Figure 1 · p. 20

Supplementary Figure 2

• CascadeNet-5 contains five copies of convolutional layers with 64, 128, and 256 filters, with outputs concatenated and cascaded forward between network blocks.The schematic contrasts this five-copy architecture with one- and two-copy variants.Source: Figure 2 · p. 21