Predicting Humphrey 10-2 visual field from 24-2 visual field in eyes with advanced glaucoma

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Advanced glaucoma significantly impacts visual field sensitivity, particularly in the central macula.
• Accurate prediction of visual field test results is crucial for monitoring disease progression and treatment efficacy.
• Current methods for predicting detailed visual fields from broader tests have limitations in accuracy.

Purpose of the Study:

• To predict Humphrey Field Analyzer Central 10-2 (HFA 10-2) test results using data from Humphrey Field Analyzer 24-2 (HFA 24-2) tests in eyes with advanced glaucoma.
• To evaluate the accuracy of various machine learning methods, including support vector regression (SVR), for this prediction task.
• To compare prediction accuracy against standard interpolation methods and test-retest variability.

Main Methods:

• Utilized training (175 eyes) and testing (44 eyes) datasets of patients with open advanced glaucoma (mean deviation ≤ -20 dB).
• Employed machine learning techniques, specifically SVR, to predict 68 total deviation (TD) values of HFA 10-2 test points from the innermost 16 HFA 24-2 test points.
• Calculated absolute prediction error (PredError) and compared it with bilinear interpolation (IP) and test-retest variability.

Main Results:

• The test-retest variability for HFA 10-2 results was 2.1±1.0 dB.
• Bilinear interpolation (IP) yielded a prediction error of 5.0±1.7 dB.
• Support vector regression (SVR) demonstrated the lowest prediction error at 4.0±1.5 dB, predicting retinal sensitivity within approximately 25% error.

Conclusions:

• Support vector regression (SVR) effectively predicts Humphrey Field Analyzer 10-2 (HFA 10-2) total deviation values from Humphrey Field Analyzer 24-2 (HFA 24-2) results in advanced glaucoma.
• The prediction error using SVR (approx. 25%) is approximately twice the test-retest variability, indicating a clinically relevant level of accuracy.
• This approach offers a viable method for estimating central visual field status in patients with advanced glaucoma using readily available HFA 24-2 data.