Abstract
Purpose :
The visual field (VF) exam, especially the 24-2 or 30-2 version, is commonly used when diagnosing glaucoma. The exam tests patient response to stimuli and summarizes findings in a report. This study evaluates the potential of a machine learning based approach for glaucoma detection using a VF report. This automated approach has the potential to help specialists make better-informed decisions, to optimize screening and to help triage referrals.
Methods :
We collected 144 VF exam reports for this study. There were 28 women and 44 men (average 67 years old), and both eyes were examined. Exams were either 24-2 (58%) or 30-2 (42%). A specialist reviewed the patient charts and labelled the reports as glaucomatous (66%) or not (33%). We separated the reports into two datasets, “training” (70%) and “testing” (30%), each with the same distribution of glaucomatous eyes. For every report, we collected the dB values for each data point on the visual sensitivity map, total deviation, and standard deviation grids. The 24-2 exam has fewer grid points than 30-2, so the 24-2 datasets had null values for the corresponding plot points present in the 30-2 but not 24-2. We also recorded the fixation monitor and losses. We then cleaned the data to prepare it for analysis and applied an xgb machine learning model to the “training” dataset. We selected XGBoost because it can be trained on datasets with null values and still produce high-quality results. We then evaluated the model by running our model on the “testing” dataset and measured precision and area under the curve (AUC). We also used cross-validation to avoid overfitting.
Results :
We achieved 92% precision (in 92% of cases it will detect a glaucomatous patient) and an AUC score of 82% (in 82% of cases it will accurately classify the patient as glaucomatous or not, indicating minimal false negatives). Out of 239 features, 203 were unnecessary for the prediction to achieve the above results.
Conclusions :
We developed a reliable automated glaucoma detection approach that requires only 36 data points from a VF exam. This approach could be used to optimize referrals. Further, combining this model with portable VF exams could improve access to vision screening and hence early detection of glaucoma in rural areas. Further testing on large and diverse samples is needed to validate our findings and to discover additional insights.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.