Transformer-Based Deep Learning Prediction of 10-Degree Humphrey Visual Field Tests From 24-Degree Data

Abstract

Purpose: To predict 10-2 Humphrey visual fields (VFs) from 24-2 VFs and associated non-total deviation features using deep learning.

Methods: We included 5189 reliable 24-2 and 10-2 VF pairs from 2236 patients, and 28,409 reliable pairs of macular OCT scans and 24-2 VF from 19,527 eyes of 11,560 patients. We developed a transformer-based deep learning model using 52 total deviation values and nine VF test features to predict 68 10-2 total deviation values. The mean absolute error, root mean square error, and the R² were evaluation metrics. We further evaluated whether the predicted 10-2 VFs can improve the structure–function relationship between macular thinning and paracentral VF loss in glaucoma.

Results: The average mean absolute error and R² for 68 10-2 VF test points were 3.30 ± 0.52 dB and 0.70 ± 0.11, respectively. The accuracy was lower in the inferior temporal region. The model placed greater emphasis on 24-2 VF points near the central fixation point when predicting the 10-2 VFs. The inclusion of nine VF test features improved the mean absolute error and R² up to 0.17 ± 0.06 dB and 0.01 ± 0.01, respectively. Age was the most important 24-2 VF test parameter for 10-2 VF prediction. The predicted 10-2 VFs achieved an improved structure–function relationship between macular thinning and paracentral VF loss, with the R² at the central 4, 12, and 16 locations of 24-2 VFs increased by 0.04, 0.05 and 0.05, respectively (P < 0.001).

Conclusions: The 10-2 VFs may be predicted from 24-2 data.

Translational Relevance: The predicted 10-2 VF has the potential to improve glaucoma diagnosis.

Introduction

Glaucoma is the second leading cause of blindness worldwide and is characterized by progressive vision loss.¹ The irreversible nature of visual impairment caused by glaucoma progression underscores the importance of detecting visual dysfunction in its early stages. Visual field (VF) testing is a principal diagnostic tool for detecting glaucoma and monitoring its progression. Currently, the 24-2 VF, which measures the central 24° of VF using 54 test locations that are each 6° apart, is the most common form of perimetry used in diagnosing and monitoring glaucoma.^2–5 However, studies report that 24-2 VF testing may underestimate damages that are within the central 10° of the fovea.^5–8 Although the macula contains approximately 50% of all retinal ganglion cells (GCs),⁹ only 12 test locations of the 24-2 VF are within the central 10° of the fovea.¹⁰ 

Studies have reported that early glaucoma can affect the macular region resulting in central vision loss, although macula damage is more common in advanced diseases.^4,11,12 The 24-2 VF test often misses VF defects present in the central 10° because 68 locations are tested in the 10-2 test, with only 12 locations tested within the central 10° in the 24-2 test (Fig. 1).⁷ A prior study found that, in the glaucoma group, 16 of the 26 eyes (61.5%) classified as normal based on cluster criteria on 24-2 tests were classified as abnormal on 10-2 VFs.¹³ Hence, it is ideal to test both 24-2 and 10-2 VFs in patients with glaucoma. However, in a busy eye clinic, it is impractical to test both the 10-2 VF and the 24-2 VF on most patients owing to cost and time burdens. Limited to clinical resources, clinicians mainly order 10-2 VF tests when paracentral defects are observed in 24-2 VFs or the patient has advanced VF loss with central sparing in clinical practice. 

Figure 1.

 

Two examples show possible mapping relationships between 24-2 and 10-2 VFs. The color bar corresponds to the VF TD values. The first example shows that (A) upper quadrant VF defects were observed by 24-2 VF at central locations and by (B) corresponding 10-2 VF. The second example shows that (C) no VF defects were observed by 24-2 VF at central locations but observed by (D) corresponding 10-2 VF.

Optical coherence tomography (OCT) is significantly faster to administer (total testing time of 1 or 2 vs. 10–15 minutes) than the 10-2 VF test. In recent years, deep learning techniques have been explored in various studies to estimate 10-2 VFs using OCT scans near the optic nerve head or macula, showcasing the potential for advancements in this area.^14–16 Although these methods have faced limitations owing to the moderate structure–function relationship in glaucoma, which can impact their accuracy and practical application, researchers have proposed innovative approaches that integrate the 24-2 VF data with OCT information to predict 10-2 VFs.^17,18 Although these hybrid models still rely on OCT scans and encounter the challenges posed by the moderate structure–function relationship, they represent important steps forward. In a recent study involving approximately 200 patients, an intriguing approach was explored to predict 10-2 VF test results using only 24-2 VF data in individuals with advanced glaucoma, using a support vector machine.¹⁹ Although the prediction accuracy achieved a mean absolute error (MAE) of 4.0 dB without reporting correlations (R²), this pioneering research highlights the potential for further advancements in this area. Furthermore, considering the widespread use of 10-2 VFs in advanced glaucoma cases, exploring the usefulness of predicted 10-2 VFs in central vision loss assessment could yield valuable insights and enhance clinical practices. 

In this study, we rely on the 24-2 VF total deviation (TD) and VF test parameters to predict 10-2 VFs using an artificial intelligence method called transformer-based deep learning (Fig. 2). R², MAE, and root mean square error (RMSE) are used to assess the 10-2 VF prediction accuracy in a pointwise manner. We quantify the feature importance for predicting 10-2 VFs using Shapley additive explanations.²⁰ We demonstrate that the structure–function correlation for central vision loss is improved by using predicted 10-2 VFs compared with using the central locations within 10° in 24-2 VFs. Specifically, we use the macular GC plus the inner plexiform layer (IPL) to predict the 24-2 TDs for the central 4, 12, and 16 locations compared with predicting the 10-2 TDs at all 68 locations. 

Figure 2.

 

The workflow of model training and evaluation in this work. In step 1, a transformer-based deep learning model was trained to predict 10-2 VF from corresponding 24-2 VF data. In step 2, the 10-2 VF prediction was evaluated in two ways. First, we calculated the pointwise MAE, RMSE, and R² (a measure of correlation) to measure the prediction accuracy by comparing the actual and predicted 10-2 TD values. Second, we performed structure–function correlation analyses by using the macular GC-IPL thickness map to predict 68 TDs of the 10-2 and paracentral (at the 4, 12, and 16 locations closest to fixation) 24-2 VFs, respectively.

Methods

The VF data used for developing the 10-2 VF prediction model using 24-2 VFs were collected and managed by the Glaucoma Research Network, a multicenter consortium including data from Massachusetts Eye and Ear at Harvard Medical School, Wilmer Eye Institute at Johns Hopkins University, New York Eye and Ear Infirmary at Icahn School of Medicine at Mount Sinai, Bascom Palmer Eye Institute at University of Miami, Wills Eye Hospital at Thomas Jefferson University, and Edward S. Harkness Eye Institute at Columbia University. The institutional review boards of each ophthalmic center approved the creation of the database in this retrospective study. The OCT and VF data used for structure–function relationship assessment were from the Massachusetts Eye and Ear Ophthalmic Imaging and Testing Registry, which was approved by the Massachusetts Eye and Ear Institutional Review Board. This study complied with the guidelines outlined in the Declaration of Helsinki. In light of the study's retrospective design, the requirement for informed consent was waived. All deep learning modeling and statistical analyses were performed in Python 3.8 (available at http://www.python.org) on a Linux system. 

Dataset Description

Glaucoma Research Network Dataset

We included reliable Swedish Interactive Thresholding Algorithm VFs from the Glaucoma Research Network dataset measured by the Humphrey Field Analyzer (Carl Zeiss Meditec, Dublin, CA). The reliability criteria were fixation loss of ≤33%, false-negative rates of ≤20%, and false-positive rates of ≤20%, which were consistent with the reliability criteria used in a number of prior studies.^21–24 Pairs of 24-2 and 10-2 VFs tested within 30 days were selected to develop the transformer-based model for 10-2 VF prediction. 

Massachusetts Eye and Ear Dataset

We selected pairs of reliable 24-2 VFs and Cirrus macular OCT scans tested within 30 days. The reliability criteria for 24-2 VFs were fixation loss of ≤33%, false-negative rates of ≤20%, and false-positive rates of ≤20% and the reliability criterion for macular OCT scans was a signal strength of ≥6. The Massachusetts Eye and Ear dataset was used to determine if using predicted 10-2 VFs from the 24-2 VFs enhances the structure–function relationship between paracentral VF loss and GC-IPL measurement. 

Transformer-based Deep Learning Prediction Model

We developed a transformer network-based deep learning model (Fig. 3A) to predict TD values in 10-2 VF (68 in total) from 24-2 VF TDs (52 in total) and nine VF test features of the 24-2 VF test. Our framework consists of a transformer network²⁵ to process the nine VF test features of 24-2 VF, and a multilayer perceptron network to process the 52 TD features of 24-2 VF. The nine 24-2 VF test features are: mean deviation (MD), pattern standard deviation (PSD), MD probability, foveal sensitivity, foveal sensitivity probability, laterality (i.e., right eye or left eye), test duration (i.e., seconds), the time difference (i.e. gap) between 24-2 and 10-2 VF testing, and patient age. Among the nine VF test features, laterality was a categorical feature, and the rest were continuous features. The selection of these VF test features was based on considerations of their availability in the experimental data. However, our model is flexible to include other VF test features (if available) in the training course. Because VF test features including fixation loss, false-negative rates, and false-positive rates have been used for filtering reliable VF tests while composing the dataset, they were excluded during the modeling. The test duration of 24-2 VF can be related to many factors, such as the extent and severity of vision loss. By considering the time gap between 24-2 and 10-2 VF tests, the model can learn about variations in central vision loss associated with time. An encoder was used to encode VF test features into a single feature vector (Fig. 3B). Subsequently, the transformer network takes the feature vector as input to model the correlations among the nine VF test features through the multihead self-attention module (Fig. 3C). To maximize the use of the data, we adopted three-fold cross-validation to train and test the model with patient-level data split. Specifically, VFs were split into three patient-level subsets, where each subset was used iteratively as the testing set, while the other two subsets were combined to train the deep learning model for 10-2 VF prediction. We calculated the pointwise MAE, RMSE, and R² (a measure of correlation) to measure the prediction accuracy by comparing the actual and predicted 10-2 TD values with the test set. Lower MAE and RMSE values indicate more accurate predictions, whereas higher R² values signify better model performance. We also compared the prediction performance of our deep learning model with and without the use of the nine VF test features. Paired Student's t-test was used for the comparison, and the performance improvement was considered significant if P is <0.05. The Shapley additive explanations were used to explain the 10-2 VF prediction by computing the contribution (i.e., feature importance) of each 24-2 VF location to the prediction.²⁰ Shapley additive explanation is a game theoretic approach to explain the output of any machine learning model. It connects optimal credit allocation with local explanations using the classic Shapley values from game theory and their related extensions. 

Figure 3.

 

The proposed transformer-based deep learning model. (A) The deep learning framework for 10-2 VF prediction from 24-2 VF and summary features. It uses a transformer network to process the nine VF test features, and a multilayer perceptron network to process the 52 TD features. The outputs from the two networks are combined for predicting 10-2 VF. (B) The encoding module, which encodes numerical and categorical summary features into a single d-dimensional feature vector. The feature vector is then used as the input of the transformer network. (C) A single transformer layer, which models summary feature correlations with the self-attention mechanism. Norm indicates the normalization operation, and the multihead self-attention captures the correlations among nine VF test features. Finally, the feedforward layer outputs a new feature vector which can be used as the input for the next transformer layer.

Structure–Function Relationship Assessment

We evaluated the structure–function relationships between macular thinning and paracentral VF loss by comparing the associations between OCT GC-IPL maps and predicted 68 10-2 TDs, which were generated by our transformer-based deep learning model from the original 24-2 VFs, as well as the paracentral VF loss observed in the original 24-2 VFs. We hypothesize that the 10-2 VFs, predicted from 24-2 VFs, will show a stronger correlation with the corresponding macular GC-IPL map than the original 24-2 VFs in the paracentral region. The paracentral VF loss in the original 24-2 VFs was measured by TDs at the 4, 12, and 16 locations closest to fixation. The correlation analysis was performed by using the GC-IPL thickness map to predict the TDs of the 10-2 and paracentral 24-2 VFs, respectively. We used a convolutional neural network model (i.e., VGG-16), which is efficient to learn useful information from GC-IPL map to predict the corresponding VF TDs.²⁶ Three-fold cross-validation was used to train and test our deep learning-based structure–function model. MAE, RMSE, and R² were used to compare the difference between actual VF TDs and TDs predicted from the GC-IPL thickness map. Bootstrapping with a t-test was used to compare if using the predicted 10-2 VFs from 24-2 VFs can improve the structure–function relationship between macular damage and paracentral VF loss in glaucoma. 

Results

Participant Characteristics

We included 5189 reliable 24-2 and 10-2 VF pairs tested within 30 days from 2236 patients with an average age of 63.2 ± 15.2 years (Table 1). The average MDs of the 24-2 and 10-2 VFs were −10.2 ± 8.7 dB and −4.9 ± 6.1 dB, respectively. The average time gap between 24-2 and 10-2 VFs was 4.8 ± 9.0 days. For 24-2 VFs, the average TD at the central 4, 12, and 16 locations were −7.0 ± 7.8 dB, −8.0 ± 8.0 dB, and −8.1 ± 8.0 dB, respectively. 

Table 1.

 

Patient Demographics and Baseline Characteristics of the Data Used to Build the Deep Learning Model for 24-2 VF Prediction From 10-2 VF

 

We used 28,409 reliable pairs of macular OCT scans and 24-2 VF from a total of 19,527 eyes of 11,560 patients to evaluate the structure–function correlation between macular OCT scans and VFs (Table 2). Slightly more than one-half of all patients were female (59.1%), and the average age was 58.6 ± 17.7 years. The average GC-IPL thickness was 71.5 ± 13.8 µm. The average 24-2 VF MD was −3.9 ± 5.4 dB. The average 24-2 TD at central 4, 12, and 16 locations were −3.1 ± 6.7 dB, −3.6 ± 6.9 dB, and −3.7 ± 6.9 dB, respectively. 

Table 2.

 

Demographics and Baseline Characteristics of the Patients Included in the Evaluation of Structure–Function Correlations Between Macular Scans and VFs

 

10-2 VF Prediction Performance

The overall MAE and R² for the 68 10-2 VF test points were 3.30 ± 0.52 dB and 0.70 ± 0.11. Among the 68 10-2 VF test points, the MAE ranged from 2.23 to 4.30 dB and the R² ranged from 0.43 to 0.83 (Fig. 4). The prediction accuracy is lower in the inferior temporal region (Fig. 3B), which was previously reported as the less vulnerable zone by Hood and coworkers.¹¹ In the mild glaucoma group (defined as an MD of ≥−6 dB), the average MAE and R² were 2.25 ± 0.47 dB and 0.42 ± 0.10, respectively (Fig. 5). For the moderate glaucoma group (defined as an MD of ≥−12 dB and an MD of <−6 dB), the average MAE and R² were 3.81 ± 0.85 dB and 0.61 ± 0.11 (Fig. 5). In the severe glaucoma group (defined as an MD of <−12 dB), the average MAE and R² were 4.88 ± 0.69 dB and 0.62 ± 0.12, respectively (Fig. 5). As expected, the 24-2 VF locations closer to the fixation point within the central 10° were most important for predicting the 10-2 VFs (Fig. 6A). However, 24-2 VF locations outside the central 10° also moderately contributed to the 10-2 VF prediction (Fig. 6A). Age was the most important 24-2 VF test parameter for 10-2 VF prediction, followed by PSD, test duration, and MD probability which were all similarly weighted by the model (Fig. 6B). 

Figure 4.

 

Pointwise performance of 10-2 VF prediction from 24-2 VF and summary features. (A) Pointwise MAE performance. (B) Pointwise RMSE performance. (C) Pointwise R² performance.

Figure 5.

 

Pointwise performance of 10-2 VF prediction from 24-2 VF and summary features divided in subgroups. (A) Mild glaucoma group defined as 24-2 VF MD of ≥−6 dB. (B) Moderate glaucoma group defined as 24-2 VF MD of ≥−12 dB and MD of <−6 dB. (C) Severe glaucoma group defined as 24-2 VF MD of <−12 dB.

Figure 6.

 

Pointwise importance of VF TD and VF test features for the 10-2 VF prediction based on the Shapley additive explanations analysis.²⁰ (A) The importance of 52 TD features. (B) The importance of nine VF test features.

Impact of VF Test Features

When only using 52 24-2 TD values without including the 9 VF test features, the average MAE, RMSE, and R² for the 10-2 TD prediction were 3.47 ± 0.55 dB, 5.51 ± 0.70 dB, and 0.68 ± 0.10, respectively (Figs. 7A–C). In comparison, combining 52 24-2 TDs with the nine VF test features significantly (P < 0.05) improved MAE, RMSE and R² performance for all 68 10-2 VF test points. The average MAE, RMSE and R² improvements were a decrease of 0.17 ± 0.06 dB (up to 0.35 dB), a decrease of 0.13 ± 0.06 dB (up to 0.27 dB), and an increase of 0.01 ± 0.01 (up to 0.03), respectively (Figs. 7D–F). In two patient examples of 24-2 VFs and actual versus predicted 10-2 VFs, there was no clear evidence of central VF loss in the 24-2 VFs (Fig. 8). Our predicted 10-2 VFs reproduced the superior temporal defect even though the MAE error was relatively high (Fig. 8). 

Figure 7.

 

Comparison between 10-2 VF predictions with and without using the nine VF summary features. (A) Pointwise MAE performance without using the nine summary features. (B) Pointwise RMSE performance without using the nine summary features. (C) Pointwise R² performance without using the nine summary features. (D) MAE improvement with use of the nine summary features. (E) RMSE improvement with use of the nine summary features. (F) R² improvement with use of the nine summary features. Zero indicates no improvement for the respective VF location or that the improvement is insignificant (Student's t-test; P < 0.05 is considered as significant).

Figure 8.

 

Two 10-2 VF prediction examples. The color bar represents the variation of TD values, and each color corresponds to a specific VF TD value. The first row shows the (A) actual 24-2, (B) actual 10-2, and (C) predicted 10-2 VFs for the first example. The second row shows (D) actual 24-2, (E) actual 10-2, and (F) predicted 10-2 VFs for the second example.

Figure 9.

 

The comparison between structure–function correlations of original 24-2 VF and model-predicted 10-2 VF with corresponding macular OCT scans. The correlation analysis was performed by using the GC-IPL thickness map to predict 68 TDs of the 10-2 VF and paracentral TDs (at the 4, 12, and 16 locations closest to fixation) of the 24-2 VF, respectively.

Correlation Between Macular OCT Scans and VFs

The correlation analysis was performed by using the GC-IPL thickness map to predict the 68 TDs of the 10-2 and paracentral (at the 4, 12, and 16 locations closest to fixation) 24-2 VFs, respectively. In other words, two types of structure–function correlation were compared: (1) GC-IPL thickness map predicts the 24-2 VF at central 4, 12, and 16 TDs and (2) GC-IPL thickness map predicts the 10-2 VF, which was previously predicted from the corresponding 24-2 VF. As shown in Table 3, the model-predicted 10-2 VFs were better associated (MAE, 2.62 dB; R², 0.37) with the macular GC-IPL maps with deep learning modeling compared with the original 24-2 TDs at the central four (MAE, 3.08 dB; R², 0.33; P < 0.001), 12 (MAE, 3.26 dB; R², 0.32; P < 0.001) and 16 (MAE, 3.32 dB; R², 0.32; P < 0.001) locations. The improved structure–function relationship between paracentral VF loss and macular damage by using predicted 10-2 VFs from 24-2 VFs shows the predicted 10-2 VFs contained more information about paracentral VF loss than the central locations in 24-2 VFs, highlighting the clinical relevance of our transformer-based 10-2 VF prediction. The Bland–Altman plots demonstrate a moderate disagreement between these two types of structure–function correlations (Fig. 9), indicating that the predicted 10-2 VF exhibits a better correlation with the corresponding GC-IPL thickness map compared with the original 24-2 VF. The correlation analyses in mild, moderate, and severe glaucoma patient groups further demonstrate that the predicted 10-2 VFs improved structure–function correlation over original 24-2 VFs at the central 4, 12, and 16 locations closest to fixation (Table 4). 

Table 3.

 

The Correlation of Original 24-2 VF and Model-predicted 10-2 VF with Corresponding Macular OCT Scans

 

Table 4.

 

The Correlation of Original 24-2 VF and Model-predicted 10-2 VF with Corresponding Macular OCT Scans Stratified by Glaucoma Severity

 

Discussion

In this study, a transformer-based deep learning model was proposed to predict 10-2 VF TDs (n = 68 points) from 24-2 VF TDs (n = 52 points) and nine associated VF test features. Our deep learning model comprises a transformer network to learn feature representation of VF test features and an multilayer perceptron network to learn feature representation of the 24-2 TDs. Considering the inherent test-retest variability, our deep learning model demonstrated robust performance for 10-2 VF TD prediction with an average MAE of 3.30 ± 0.52 dB and R² of 0.70 ± 0.11. We have done an additional experiment using the 8150 baseline 10-2 VFs to predict subsequent follow-up 10-2 VFs. It achieved an average MAE and R² of 2.81 ± 0.29 dB and 0.82 ± 0.06, respectively, which can be used as a reference (such as the maximum achievable performance in predicting 10-2 VFs) to reveal that the achieved 10-2 VF prediction performance in this work is reasonable given that only 12 test locations of the 24-2 VF are within the central 10° of the fovea. These results highlight the model's ability to accurately predict 10-2 VF TD values, even when considering the variability in repeated measurements (Fig. 4). We further demonstrated the clinical relevance of our 10-2 prediction by assessing the structure–function relationship between paracentral VF loss and macular damage. 

We used Shapley additive explanations to quantify the feature importance for 10-2 VF prediction. As expected, 24-2 VF locations closer to the fixation with central 10° were more important than other locations for the 10-2 VF prediction (Fig. 6A).^17,27 The most important VF test feature was age, which may indicate that central vision loss depends on an individual's age. As individuals age, they may be more susceptible to age-related degenerative changes in the retina and optic nerve, which can contribute to central vision impairments. Not surprisingly, PSD and MD probability were associated with the 10-2 VF prediction because they were indicators of glaucoma. Interestingly, the 24-2 VF test duration was also related to the 10-2 VF prediction. This finding may be due to patients with more severe central vision loss taking longer to complete the 24-2 tests, making the test duration highly correlated with the TD values. We demonstrated that adding VF test features on top of the 24-2 TDs can improve 10-2 TD prediction. 

Hood and coworkers previously reported the inferior-temporal region in 10-2 VFs as the less vulnerable zone in the VF.^11,27 The 10-2 VF prediction by our transformer model was also relatively less accurate in this vulnerability zone compared with other regions in the 10-2 VF (Fig. 3B). This result indicates the weaker relationship between vision loss in the less vulnerable zone and overall vision loss manifested in the 24-2 VF. 

In prior studies, deep learning models have used OCT scans predominantly to predict 10-2 VFs.^14–16,28 However, owing to the moderate structure–function correlation in glaucoma, these approaches have encountered challenges in achieving high prediction accuracy. Conversely, some studies have explored the use of 24-2 or 30-2 VFs alone or in combination with OCT scans for predicting 10-2 VFs.^17–19,29 However, these studies had certain limitations. First, they lacked comprehensive quantification of prediction accuracy, with some studies only reporting binary outcomes or providing MAE without R² values,^18,29 which makes it difficult to assess the overall accuracy.^17,19 Second, these studies did not demonstrate whether predicting 10-2 VFs from 24-2 VFs provided additional information about paracentral VF loss, making it less clinically meaningful. More specifically, it is unclear whether the predicted 10-2 VF demonstrates better correlation with the retinal damage at macular region, compared with the corresponding 24-2 VF. Last, the sample sizes in previous studies were relatively small, limiting the coverage of a diverse patient population.^17–19,29 In comparison, our study addresses these limitations by reporting both MAE and R² performance metrics of 3.30 ± 0.52 dB and 0.70 ± 0.11, respectively, which outperforms the reported MAEs in the literature.^17,19 We also showed two examples that demonstrated that the predicted 10-2 VFs by our transformer can capture the paracentral VF loss largely missed in the 24-2 VFs (Fig. 8). We further demonstrated the clinical value of information gain about paracentral VF loss by improving the macular OCT scan's association with predicted 10-2 VFs from 24-2 VFs compared with the central 4, 12, and 16 locations of the 24-2 VFs. More important, for both the transformer model for 10-2 prediction and the VGG-16 model for structure–function relationship assessment, several thousand patients were used for each of the two tasks. 

Our work also has some limitations. First, the pairs of 24-2 and 10-2 VFs from the Glaucoma Research Network have a sample bias toward severe glaucoma, evidenced by the 24-2 MD of −10.2 ± 8.7 dB, although we demonstrated that the 10-2 VF prediction remains relatively accurate with an R² of 0.42 in comparison with the R² of 0.62 for severe glaucoma and R² for all 24-2 VFs regardless of severities. Second, although the R² performance for 10-2 VF prediction is promising, a clinical validation study will be needed to demonstrate how this model can be used by clinicians to make decisions. Specially, although it is possible to predict the pattern of loss (Fig. 8), the extent of scotoma depth varies significantly and is prone to errors, suggesting opportunities for further refinement. Third, we did not combine structural data into our model to predict 10-2 VFs owing to a lack of data. Fourth, our model only used VFs measured by Zeiss Humphrey Field Analyzer, and VFs tested by other devices such as Octopus are not available to use to generalize our model across different testing platforms. Last, we do not have data available for the testing protocol of 24-2C (testing the central 24° light sensitivities with additional testing locations within the central 10°), which can be used to validate the prediction accuracy of our model further. 

In conclusion, we proposed a transformer-based deep learning model for 10-2 VF prediction from 52 TD points and 9 VF test features of 24-2 VF. The experimental findings demonstrate that the deep learning model is capable of predicting 10-2 VFs meaningfully. These predictions showed a stronger correlation with macular OCT scans than the corresponding 24-2 VFs at central locations. The 10-2 VF prediction is interpretable through the generation of feature importance of 24-2 VF TD and VF test features. Our model may be useful for clinicians in deciding when to further test patients with a 10-2 VF in addition to standard 24-2 VF testing protocols. The code for our model is available at the following link: https://github.com/Harvard-Ophthalmology-AI-Lab/VFTransformer. 

Acknowledgments

Supported by NIH R00 EY028631 (M.W.), Research to Prevent Blindness International Research Collaborators Award (M.W.), Alcon Young Investigator Grant (M.W.), and NIH P30 EY003790 (M.W., T.E.). 

Disclosure: M. Shi, None; A. Lokhande, None; Y. Tian, None; Y. Luo, None; M. Eslami, None; S. Kazeminasab, None; T. Elze, Genentech (F); L.Q. Shen, Firecyte Therapeutics (C), Abbvie (C); L.R. Pasquale, Eyenovia-Advisory Board Member (C), Twenty-Twenty (C), Skye Biosciences (C); S.R. Wellik, None; C.G. De Moraes, None; J.S. Myers, None; N. Zebardast, None; D.S. Friedman, None; M.V. Boland, None; M. Wang, Genentech (F)