advertisement

Editors Selection IGR 11-2

Clinical Examination Methods: Enhancing Detection of Visual Field Progression

Comment by Riccardo Cheloni & Giovanni Montesano & David Crabb on:

112677 Efficacy of Smoothing Algorithms to Enhance Detection of Visual Field Progression in Glaucoma, Mohammadzadeh V; Li L; Fei Z et al., Ophthalmology science, 2024; 4: 100423

Find related abstracts

The idea of post-processing visual fields (VFs) with some sort of spatial filter is 'as old as the hills'.¹ Conventional smoothing algorithms (such as the nearest neighbour smoothing, NN) modify the sensitivity at each VF location based on sensitivities of adjacent locations according to given criteria. Deep-learning algorithms, such as variational autoencoders (VAE), can offer a more sophisticated approach to the same problem. The authors conducted a retrospective longitudinal analysis of 4232 patients (7150 eyes) with glaucoma, comparing VF progression using un-smoothed (original) and data smoothed with either a NN or a VA algorithm. Progression was detected using different methods, including trend analysis of mean total deviation (mTD) and VF index (VFI) and various pointwise progression methods. The improvements in detection for smoothed data were modest, although statistically significant (except for the VFI). Time do detect progression was shorter for smoothed data only for some of the pointwise progression methods. These effects were also statistically significant, but time gains were modest (~0.7 years).

This result is somewhat expected, because noise in global indices, such as VFI or mTD, is mostly affected by visit-to-visit performance fluctuations, which would not be addressed by smoothing methods applied to single tests.^2,3

Unfortunately, the authors did not report the specificity in detecting progression with the smoothed and original data. This takes away important context to determine whether the small increase in detection rate came at the cost of increased false positive detections.

Unfortunately, the authors did not report the specificity in detecting progression with the smoothed and original data. This takes away important context to determine whether the small increase in detection rate came at the cost of increased false positive detections. One conventional approach in the literature has been to estimate false positive rates from permutations of reordered VF series.^4-7

Interestingly, the authors make statements about specificity in their conclusions. Overall, authors concluded that 'smoothing' of VFs data allows earlier detection of glaucoma progression, without reducing specificity. Yet, the accuracy of the increased detection remains unclear. The results, despite being statistically significant, may be less compelling clinically, especially without the context provided by a clear quantification of specificity.

• ← Previous
• Next →