Forecasting Risk of Future Rapid Glaucoma Worsening Using Early Visual Field, Optical Coherence Tomography and Clinical Data

Abstract

Purpose : We assess whether we can forecast future rapid visual field (VF) worsening using deep learning models (DLM) trained on baseline VF, OCT, and clinical data. Additionally, we study the impact of adding additional VF data (beyond baseline) on model performance.

Methods : We included eyes that met the following: 1) followed for glaucoma or suspect status 2) had at least five reliable VFs over time (VF_1, VF_2, VF_3, VF_4, VF₅) 3) had one reliable baseline OCT scan (OCT₁) and one set of baseline clinical measurement: age, gender, BCVA, IOP (Clinical₁) at the time of VF₁.
We designed a DLM to detect eyes at risk for future rapid VF worsening (mean deviation (MD) slope <-1 dB/yr across all five VFs). The input to the DLM consisted of spatially oriented total deviation values from VF₁ (+/- VF₂ and VF₃ in some models) and spatially oriented RNFL thickness values from OCT₁ (Figure 1A). We input this VF/OCT stack into a vision transformer feature extractor. The output of the feature extractor was concatenated with Clinical₁ and put through a dense linear classifier to make a final prediction for that eyes’ risk of future rapid VF worsening (Figure 1B). We used 80% of data for training, 10% for validation, and 10% for testing.
We compared the performance of models with differing inputs by computing area under receiver operating curve (AUC) in the test set. Specifically, we trained models with following inputs: Model (1) VF₁; (2) VF₁+ Clinical₁; (3) VF₁+ OCT₁; (4) VF₁+ Clinical₁+ OCT₁; (5) VF_{1 +} VF₂ + Clinical₁ + OCT₁; (6) VF₁ + VF₂ + VF_3, + Clinical₁ + OCT₁

Results : We included a total of 4,537 unique eyes from 2,963 patients. Mean and (SD) for: age | VF MD at baseline | MD slope, was: 65.9 years (12.4) | -3.4 dB (4.3) | -0.21 dB/yr (1.0). 518 (11.4%) of eyes underwent rapid VF worsening. Model 6 most accurately forecasted rapid worsening with an AUC (95% CI) of 0.85 (0.78, 0.93). Remaining models in descending order of performance and their respective AUC [95% CI] were: Model 5 (0.81 [0.73 to 0.89]), Model 4 (0.75 [0.66, 0.83]), Model 2 [0.75 [0.66, 0.83], Model 3 (0.71 [0.62, 0.80]), Model 1 (0.68 [0.59, 0.77]).

Conclusions : DLMs trained on multimodal data from early visits can forecast future rapid worsening with AUC >0.8. Deployment of such models in clinical practice may allow us to stratify high from low-risk glaucoma patients early in the disease course.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1