Opportunities for Improving Glaucoma Clinical Trials via Deep Learning-Based Identification of Patients with Low Visual Field Variability.

PURPOSE: Develop and evaluate the performance of a deep learning model (DLM) that forecasts eyes with low future visual field (VF) variability, and study the impact of using this DLM on sample size requirements for neuroprotective trials. DESIGN: Retrospective cohort and simulation study. METHODS: We included 1 eye per patient with baseline reliable VFs, OCT, clinical measures (demographics, intraocular pressure, and visual acuity), and 5 subsequent reliable VFs to forecast VF variability using DLMs and perform sample size estimates. We estimated sample size for 3 groups of eyes: all eyes (AE), low variability eyes (LVE: the subset of AE with a standard deviation of mean deviation [MD] slope residuals in the bottom 25th percentile), and DLM-predicted low variability eyes (DLPE: the subset of AE predicted to be low variability by the DLM). Deep learning models using only baseline VF/OCT/clinical data as input (DLM1), or also using a second VF (DLM2) were constructed to predict low VF variability (DLPE1 and DLPE2, respectively). Data were split 60/10/30 into train/val/test. Clinical trial simulations were performed only on the test set. We estimated the sample size necessary to detect treatment effects of 20% to 50% in MD slope with 80% power. Power was defined as the percentage of simulated clinical trials where the MD slope was significantly worse from the control. Clinical trials were simulated with visits every 3 months with a total of 10 visits. RESULTS: A total of 2817 eyes were included in the analysis. Deep learning models 1 and 2 achieved an area under the receiver operating characteristic curve of 0.73 (95% confidence interval [CI]: 0.68, 0.76) and 0.82 (95% CI: 0.78, 0.85) in forecasting low VF variability. When compared with including AE, using DLPE1 and DLPE2 reduced sample size to achieve 80% power by 30% and 38% for 30% treatment effect, and 31% and 38% for 50% treatment effect. CONCLUSIONS: Deep learning models can forecast eyes with low VF variability using data from a single baseline clinical visit. This can reduce sample size requirements, and potentially reduce the burden of future glaucoma clinical trials. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Deep learning-based identification of eyes at risk for glaucoma surgery.

To develop and evaluate the performance of a deep learning model (DLM) that predicts eyes at high risk of surgical intervention for uncontrolled glaucoma based on multimodal data from an initial ophthalmology visit. Longitudinal, observational, retrospective study. 4898 unique eyes from 4038 adult glaucoma or glaucoma-suspect patients who underwent surgery for uncontrolled glaucoma (trabeculectomy, tube shunt, xen, or diode surgery) between 2013 and 2021, or did not undergo glaucoma surgery but had 3 or more ophthalmology visits. We constructed a DLM to predict the occurrence of glaucoma surgery within various time horizons from a baseline visit. Model inputs included spatially oriented visual field (VF) and optical coherence tomography (OCT) data as well as clinical and demographic features. Separate DLMs with the same architecture were trained to predict the occurrence of surgery within 3 months, within 3-6 months, within 6 months-1 year, within 1-2 years, within 2-3 years, within 3-4 years, and within 4-5 years from the baseline visit. Included eyes were randomly split into 60%, 20%, and 20% for training, validation, and testing. DLM performance was measured using area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC). Shapley additive explanations (SHAP) were utilized to assess the importance of different features. Model prediction of surgery for uncontrolled glaucoma within 3 months had the best AUC of 0.92 (95% CI 0.88, 0.96). DLMs achieved clinically useful AUC values (> 0.8) for all models that predicted the occurrence of surgery within 3 years. According to SHAP analysis, all 7 models placed intraocular pressure (IOP) within the five most important features in predicting the occurrence of glaucoma surgery. Mean deviation (MD) and average retinal nerve fiber layer (RNFL) thickness were listed among the top 5 most important features by 6 of the 7 models. DLMs can successfully identify eyes requiring surgery for uncontrolled glaucoma within specific time horizons. Predictive performance decreases as the time horizon for forecasting surgery increases. Implementing prediction models in a clinical setting may help identify patients that should be referred to a glaucoma specialist for surgical evaluation.

Assessment of linear regression of peripapillary optical coherence tomography retinal nerve fiber layer measurements to forecast glacuoma trajectory.

Linear regression of optical coherence tomography measurements of peripapillary retinal nerve fiber layer thickness is often used to detect glaucoma progression and forecast future disease course. However, current measurement frequencies suggest that clinicians often apply linear regression to a relatively small number of measurements (e.g., less than a handful). In this study, we estimate the accuracy of linear regression in predicting the next reliable measurement of average retinal nerve fiber layer thickness using Zeiss Cirrus optical coherence tomography measurements of average retinal nerve fiber layer thickness from a sample of 6,471 eyes with glaucoma or glaucoma-suspect status. Linear regression is compared to two null models: no glaucoma worsening, and worsening due to aging. Linear regression on the first M ≥ 2 measurements was significantly worse at predicting a reliable M+1st measurement for 2 ≤ M ≤ 6. This range was reduced to 2 ≤ M ≤ 5 when retinal nerve fiber layer thickness measurements were first "corrected" for scan quality. Simulations based on measurement frequencies in our sample-on average 393 ± 190 days between consecutive measurements-show that linear regression outperforms both null models when M ≥ 5 and the goal is to forecast moderate (75th percentile) worsening, and when M ≥ 3 for rapid (90th percentile) worsening. If linear regression is used to assess disease trajectory with a small number of measurements over short time periods (e.g., 1-2 years), as is often the case in clinical practice, the number of optical coherence tomography examinations needs to be increased.

The Impact of Achieving Target Intraocular Pressure on Glaucomatous Retinal Nerve Fiber Layer Thinning in a Treated Clinical Population.

PURPOSE: To estimate the effect of being below and above the clinician-set target intraocular pressure (IOP) on rates of glaucomatous retinal nerve fiber layer (RNFL) thinning in a treated real-world clinical population. DESIGN: Retrospective cohort study. METHODS: A total of 3256 eyes (1923 patients) with ≥5 reliable optical coherence tomography scans and 1 baseline visual field test were included. Linear mixed-effects modeling estimated the effects of the primary independent variables (mean target difference [measured IOP - target IOP] and mean IOP, mm Hg) on the primary dependent variable (RNFL slope, µm/y) while accounting for additional confounding variables (age, biological sex, race, baseline RNFL, baseline pachymetry, and disease severity). A spline term accounted for differential effects when above (target difference >0 mm Hg) and below (target difference ≤0 mm Hg) target pressure. RESULTS: Eyes below and above target had significantly different mean RNFL slopes (-0.44 vs -0.71 µm/y, P < .001). Each 1 mm Hg increase above target had a 0.143 µm/y faster rate of RNFL thinning (P < .001). Separating by disease severity, suspect, mild, moderate, and advanced glaucoma had 0.135 (P = .002), 0.116 (P = .009), 0.203 (P = .02), and 0.65 (P = .22) µm/y faster rates of RNFL thinning per 1 mm Hg increase, respectively. CONCLUSIONS: Being above the clinician-set target pressure is associated with more rapid RNFL thinning in suspect, mild, and moderate glaucoma. Faster rates of thinning were also present in advanced glaucoma, but statistical significance was limited by the lower sample size of eyes above target and the optical coherence tomography floor effect.

Predicting Visual Field Worsening with Longitudinal OCT Data Using a Gated Transformer Network.

PURPOSE: To identify visual field (VF) worsening from longitudinal OCT data using a gated transformer network (GTN) and to examine how GTN performance varies for different definitions of VF worsening and different stages of glaucoma severity at baseline. DESIGN: Retrospective longitudinal cohort study. PARTICIPANTS: A total of 4211 eyes (2666 patients) followed up at the Johns Hopkins Wilmer Eye Institute with at least 5 reliable VF results and 1 reliable OCT scan within 1 year of each reliable VF test. METHODS: For each eye, we used 3 trend-based methods (mean deviation [MD] slope, VF index slope, and pointwise linear regression) and 3 event-based methods (Guided Progression Analysis, Collaborative Initial Glaucoma Treatment Study scoring system, and Advanced Glaucoma Intervention Study [AGIS] scoring system) to define VF worsening. Additionally, we developed a "majority of 6" algorithm (M6) that classifies an eye as worsening if 4 or more of the 6 aforementioned methods classified the eye as worsening. Using these 7 reference standards for VF worsening, we trained 7 GTNs that accept a series of at least 5 as input OCT scans and provide as output a probability of VF worsening. Gated transformer network performance was compared with non-deep learning models with the same serial OCT input from previous studies-linear mixed-effects models (MEMs) and naive Bayes classifiers (NBCs)-using the same training sets and reference standards as for the GTN. MAIN OUTCOME MEASURES: Area under the receiver operating characteristic curve (AUC). RESULTS: The M6 labeled 63 eyes (1.50%) as worsening. The GTN achieved an AUC of 0.97 (95% confidence interval, 0.88-1.00) when trained with M6. Gated transformer networks trained and optimized with the other 6 reference standards showed an AUC ranging from 0.78 (MD slope) to 0.89 (AGIS). The 7 GTNs outperformed all 7 MEMs and all 7 NBCs accordingly. Gated transformer network performance was worse for eyes with more severe glaucoma at baseline. CONCLUSIONS: Gated transformer network models trained with OCT data may be used to identify VF worsening. After further validation, implementing such models in clinical practice may allow us to track functional worsening of glaucoma with less onerous structural testing. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found after the references.

Forecasting Risk of Future Rapid Glaucoma Worsening Using Early Visual Field, OCT, and Clinical Data.

PURPOSE: To assess whether we can forecast future rapid visual field (VF) worsening using deep learning models (DLMs) trained on early VF, OCT, and clinical data. DESIGN: A retrospective cohort study. SUBJECTS: In total, 4536 eyes from 2962 patients. Overall, 263 (5.80%) eyes underwent rapid VF worsening (mean deviation slope less than -1 dB/year across all VFs). METHODS: We included eyes that met the following criteria: (1) followed for glaucoma or suspect status; (2) had at least 5 longitudinal reliable VFs (VF1, VF2, VF3, VF4, and VF5); and (3) had 1 reliable baseline OCT scan (OCT1) and 1 set of baseline clinical measurements (clinical1) at the time of VF1. We designed a DLM to forecast future rapid VF worsening. The input consisted of spatially oriented total deviation values from VF1 (including or not including VF2 and VF3 in some models) and retinal nerve fiber layer thickness values from the baseline OCT. We passed this VF/OCT stack into a vision transformer feature extractor, the output of which was concatenated with baseline clinical data before putting it through a linear classifier to predict the eye's risk of rapid VF worsening across the 5 VFs. We compared the performance of models with differing inputs by computing area under the curve (AUC) in the test set. Specifically, we trained models with the following inputs: (1) model V: VF1; (2) VC: VF1+ Clinical1; (3) VO: VF1+ OCT1; (4) VOC: VF1+ Clinical1+ OCT1; (5) V2: VF1 + VF2; (6) V2OC: VF1 + VF2 + Clinical1 + OCT1; (7) V3: VF1 + VF2 + VF3; and (8) V3OC: VF1 + VF2 + VF3 + Clinical1 + OCT1. MAIN OUTCOME MEASURES: The AUC of DLMs when forecasting rapidly worsening eyes. RESULTS: Model V3OC best forecasted rapid worsening with an AUC (95% confidence interval [CI]) of 0.87 (0.77-0.97). Remaining models in descending order of performance and their respective AUC (95% CI) were as follows: (1) model V3 (0.84 [0.74-0.95]), (2) model V2OC (0.81 [0.70-0.92]), (3) model V2 (0.81 [0.70-0.82]), (4) model VOC (0.77 [0.65-0.88]), (5) model VO (0.75 [0.64-0.88]), (6) model VC (0.75 [0.63-0.87]), and (7) model V (0.74 [0.62-0.86]). CONCLUSIONS: Deep learning models can forecast future rapid glaucoma worsening with modest to high performance when trained using data from early in the disease course. Including baseline data from multiple modalities and subsequent visits improves performance beyond using VF data alone. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Comparing the Accuracy of Peripapillary OCT Scans and Visual Fields to Detect Glaucoma Worsening.

PURPOSE: To compare the accuracy of detecting moderate and rapid rates of glaucoma worsening over a 2-year period with different numbers of OCT scans and visual field (VF) tests in a large sample of glaucoma and glaucoma suspect eyes. DESIGN: Descriptive and simulation study. PARTICIPANTS: The OCT sample comprised 12 150 eyes from 7392 adults with glaucoma or glaucoma suspect status followed up at the Wilmer Eye Institute from 2013 through 2021. The VF sample comprised 20 583 eyes from 10 958 adults from the same database. All eyes had undergone at least 5 measurements over follow-up from the Zeiss Cirrus OCT or Humphrey Field Analyzer. METHODS: Within-eye rates of change in retinal nerve fiber layer (RNFL) thickness and mean deviation (MD) were measured using linear regression. For each measured rate, simulated measurements of RNFL thickness and MD were generated using the distributions of residuals. Simulated rates of change for different numbers of OCT scans and VF tests over a 2-year period were used to estimate the accuracy of detecting moderate (75th percentile) and rapid (90th percentile) worsening for OCT and VF. Accuracy was defined as the percentage of simulated eyes in which the true rate of worsening (the rate without measurement error) was at or less than a criterion rate (e.g., 75th or 90th percentile). MAIN OUTCOME MEASURES: The accuracy of diagnosing moderate and rapid rates of glaucoma worsening for different numbers of OCT scans and VF tests over a 2-year period. RESULTS: Accuracy was less than 50% for both OCT and VF when diagnosing worsening after a 2-year period. OCT accuracy was 5 to 10 percentage points higher than VF accuracy at detecting moderate worsening and 10 to 15 percentage points higher for rapid worsening. Accuracy increased by more than 17 percentage points when using both OCT and VF to detect worsening, that is, when relying on either OCT or VF to be accurate. CONCLUSIONS: More frequent OCT scans and VF tests are needed to improve the accuracy of diagnosing glaucoma worsening. Accuracy greatly increases when relying on both OCT and VF to detect worsening. FINANCIAL DISCLOSURE(S): The author(s) have no proprietary or commercial interest in any materials discussed in this article.

A deep learning model incorporating spatial and temporal information successfully detects visual field worsening using a consensus based approach.

Glaucoma is a leading cause of irreversible blindness, and its worsening is most often monitored with visual field (VF) testing. Deep learning models (DLM) may help identify VF worsening consistently and reproducibly. In this study, we developed and investigated the performance of a DLM on a large population of glaucoma patients. We included 5099 patients (8705 eyes) seen at one institute from June 1990 to June 2020 that had VF testing as well as clinician assessment of VF worsening. Since there is no gold standard to identify VF worsening, we used a consensus of six commonly used algorithmic methods which include global regressions as well as point-wise change in the VFs. We used the consensus decision as a reference standard to train/test the DLM and evaluate clinician performance. 80%, 10%, and 10% of patients were included in training, validation, and test sets, respectively. Of the 873 eyes in the test set, 309 [60.6%] were from females and the median age was 62.4; (IQR 54.8-68.9). The DLM achieved an AUC of 0.94 (95% CI 0.93-0.99). Even after removing the 6 most recent VFs, providing fewer data points to the model, the DLM successfully identified worsening with an AUC of 0.78 (95% CI 0.72-0.84). Clinician assessment of worsening (based on documentation from the health record at the time of the final VF in each eye) had an AUC of 0.64 (95% CI 0.63-0.66). Both the DLM and clinician performed worse when the initial disease was more severe. This data shows that a DLM trained on a consensus of methods to define worsening successfully identified VF worsening and could help guide clinicians during routine clinical care.

Evidence-Based Guidelines for the Number of Peripapillary OCT Scans Needed to Detect Glaucoma Worsening.

PURPOSE: To estimate the number of OCT scans necessary to detect moderate and rapid rates of retinal nerve fiber layer (RNFL) thickness worsening at different levels of accuracy using a large sample of glaucoma and glaucoma-suspect eyes. DESIGN: Descriptive and simulation study. PARTICIPANTS: Twelve thousand one hundred fifty eyes from 7392 adult patients with glaucoma or glaucoma-suspect status followed up at the Wilmer Eye Institute from 2013 through 2021. All eyes had at least 5 measurements of RNFL thickness on the Cirrus OCT (Carl Zeiss Meditec) with signal strength of 6 or more. METHODS: Rates of RNFL worsening for average RNFL thickness and for the 4 quadrants were measured using linear regression. Simulations were used to estimate the accuracy of detecting worsening-defined as the percentage of patients in whom the true rate of RNFL worsening was at or less than different criterion rates of worsening when the OCT-measured rate was also at or less than these criterion rates-for two different measurement strategies: evenly spaced (equal time intervals between measurements) and clustered (approximately half the measurements at each end point of the period). MAIN OUTCOME MEASURES: The 75th percentile (moderate) and 90th percentile (rapid) rates of RNFL worsening for average RNFL thickness and the accuracy of diagnosing worsening at these moderate and rapid rates. RESULTS: The 75th and 90th percentile rates of worsening for average RNFL thickness were -1.09 µm/year and -2.35 µm/year, respectively. Simulations showed that, for the average measurement frequency in our sample of approximately 3 OCT scans over a 2-year period, moderate and rapid RNFL worsening were diagnosed accurately only 47% and 40% of the time, respectively. Estimates for the number of OCT scans needed to achieve a range of accuracy levels are provided. For example, 60% accuracy requires 7 measurements to detect both moderate and rapid worsening within a 2-year period if the more efficient clustered measurement strategy is used. CONCLUSIONS: To diagnose RNFL worsening more accurately, the number of OCT scans must be increased compared with current clinical practice. A clustered measurement strategy reduces the number of scans required compared with evenly spacing measurements.