Abstract
Purpose :
To investigate whether mean deviation estimated from optical coherence tomography measurements (OCT-MD) by a machine learning model could be used to replace the visual field mean deviation measurements (VF-MD) and if OCT-MD can detect glaucoma progression with similar predictive ability to VF-MD.
Methods :
A support vector machine (SVM) model was trained to generate OCT-MD estimates from retinal nerve fiber layer thickness, cup volume, disc area, rim area, cup-to-disc ratio, signal strength, and age. The model dataset comprised of 69,509 paired OCT and VF studies obtained from 42,878 eyes with < 7 reliable VF studies, and the training/test split was 80:20.
A separate progression dataset, comprised of 1,488 eyes with ≥ 7 reliable VF studies paired with OCT except for the last 2 studies, was used to evaluate the ability OCT-MD slope (calculated from SVM estimates for each eye) to predict disease worsening. A VF reference standard was created using VF-MD slope (calculated from only VF studies paired with OCT). True progressing eyes (ground truth) were defined as eyes with a VF-MD slope (calculated from all available VF studies) worse than −0.50 dB/year. The area under the curve (AUC) statistics were calculated when 100%, 80%, 60%, 40%, and 20% VF-MDs were randomly replaced with OCT-MD.
Results :
The OCT-MD estimates from the model had mean absolute errors of 2.35 dB, 1.92, 1.96, 5.76, and 15.12 dB for overall, suspect, mild, moderate, and advanced glaucoma, respectively. The AUC (95% CI) of MD slope when 100%, 80%, 60%, 40%, and 20% of VF-MD is randomly replaced with OCT-MD were 0.69 (0.65 to 0.72), 0.73 (0.70 to 0.76), 0.70 (0.67 to 0.73), 0.79 (0.76 to 0.82), and 0.83 (0.80 to 0.85), respectively. Meanwhile, the AUC (95% CI) of the VF reference standard was 0.85 (0.83 to 0.88). The AUC of 20% replacement was not different from the VF reference standard (p = 0.183) but 40%, 60%, and 80% replacement were different (p < 0.05).
Conclusions :
The OCT-MD estimates were close to the real VF-MD measurements in mild and suspect glaucoma but inaccurate in moderate and advanced disease. Up to 20% of VF-MD measurements may be substituted with OCT-MD before the diagnostic ability of MD slope becomes statistically different.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.