e65
Ophthalmology Volume 125, Number 9, September 2018
because patients’ clinical data were available only from one of the 2. Elze T, Pasquale LR, Shen LQ, et al. Patterns of functional
centers’ dataset, only 40 eyes were diagnosed with glaucoma based vision loss in glaucoma determined with archetypal analysis.
J Royal Soc Interface. 2015;12(103).
on glaucomatous optic disc changes on fundus photographs, cor-
3. Brusini P. OCT Glaucoma Staging System: a new method for
responding VF defects and OCT-retinal nerve fiber layer (RNFL) retinal nerve fiber layer damage classification using spectral-
images; thus, the authors’ conclusions might be not adequately domain OCT. Eye (Lond). 2018;32:113e119.
powered. It would be more useful and interesting to confirm in all
eyes of this study (or at least in a higher proportion of eyes) with
structural imaging whether the probable early glaucoma patients
(lower mean deviation, no glaucoma hemifield test reversal on 3 REPLY: We thank Grassi et al for their comments and for the
consecutive reliable 24-2 VFs þ VF archetypes suggesting early opportunity to clarify the implications of our work. In our
glaucoma) had corresponding abnormalities confirmed by work,1 we statistically predicted the reversal of the glaucoma
OCT-RNFL.3 Also, the authors state that for the eyes from hemifield test (GHT) to “within normal limits” after 2 consecutive
Massachusetts Eye and Ear dataset, an assessment of glaucoma test results outside normal limits. Because it is commonly assumed
status at the time of the third VF test was made based on the that glaucomatous visual field (VF) loss is irreversible, a GHT
consensus of 2 glaucoma specialists by reviewing fundus reversal to normal at the third test can be considered to indicate
photographs for glaucomatous optic disc changes and OCT false-positive results of the 2 preceding outside the normal limits
images for characteristic nerve fiber layer thinning closest to the tests. We combined a set of parameters, including representative
test date of the third VF. However, if this assessment was patterns of glaucomatous VF tests (so-called archetypes determined
performed several years from when the third VF was selected, by statistical learning2) to predict such reversals solely from the
this remains the “closest” to the third VF date but may not be so appearance of the first 2 VF measurements. As we hypothesized,
relevant. We also wonder how long after was the reliable VF that combining patterns of glaucomatous damage to summary metrics
postdated the third test (when structural data were equivocal) was such as the GHT increased our certainty that functional loss was
performed in order to confirm or exclude the diagnosis of likely glaucomatous, and not simply the result of long-term
glaucoma. testeretest variability. This finding was further validated in a
Finally, the authors did not specify what was the interval be- subsample with additional available clinical information.
tween the various VFs, and we wonder whether an earlier glaucoma Grassi et al correctly point out the reliability criteria to include
diagnosis based on the correspondence between optic discs VFs in the present study were stricter than the criteria used to
appearances and VF defects suggesting glaucomatous damage, and statistically learn the archetypes in the earlier study.2 However, we do
confirmed by correspondent changes at the OCT-RNFL would be not believe that this represents a bias in this study. If parameters of a
preferable to the model suggested by the authors (no glaucoma cross-validated model (i.e., a model trained and evaluated on
hemifield test reversal on 3 consecutive reliable 24-2 VFs þ VF different data sets) improve model performance, these parameters are
archetypes suggesting early glaucoma), which can be significantly meaningful. If, as in our case, the parameters were determined based
influenced by the patients’ prior experience with VF testing and on data with a potentially higher noise level, the cross-validation error
might be considered obsolete at the present OCT-RNFL era. may have increased, that is, we might have underestimated (instead
We commend the author’s frank acknowledgement in the dis- of overestimated) the true importance of the archetypes. Because the
cussion of the limitations of the present study but seek clarification inclusion of the archetypes yielded a substantial model evaluation
of the points we have raised. improvement, this demonstrates the archetypes are robust enough to
act as representative patterns of glaucomatous VF loss, even when
PIERGIACOMO GRASSI, MD applied to new datasets.
HENRIETTA HO, FRCOPHTH, FAMS After having evaluated the model performance, we concluded
KIN SHENG LIM, FRCOPHTH, MD our study with an additional analysis that compared the GHT and
Department of Ophthalmology, St Thomas’ Hospital, London, United our model with the diagnosis of 2 glaucoma specialists who
Kingdom reviewed additional optic nerve imaging. Grassi et al express
concern that the 97 subjects in this clinician-diagnosed dataset
Financial Disclosures: The authors have no proprietary or commercial might be too small and that structural properties of all 16 604 eyes
interest in any materials discussed in this article. should have been included. Apart from the fact that structure-based
diagnoses were not possible for most subjects in this large,
Available online: July 3, 2018.
de-identified dataset, we would like to stress that our work entirely
Correspondence:
focuses on the GHT, that is, a purely functional glaucoma impair-
Piergiacomo Grassi, MD, Department of Ophthalmology, St Thomas’ ment measure. Although we acknowledge a clinical glaucoma
Hospital, Westminster Bridge Road, London, London SE1 7EH, United diagnosis should be based on various parameters, some of them
Kingdom. E-mail: pjgrassi@libero.it. related to retinal structure, an evaluation of functional vision
separate from possible structural damage is helpful for clinicians to
interpret possibly confounding structural features and thereby to
References
improve diagnostic accuracy. Therefore, we decided not to add
optic nerve structure to predict GHT false positives. The additional
1. Wang M, Pasquale LR, Shen LQ, et al. Reversal of glaucoma structure-based model evaluation was only included to demonstrate
hemifield test results and visual field features in glaucoma. our model not only optimized the prediction of a GHT reversal, but
Ophthalmology. 2018;125:352e360. performs well even when predicting a structure-based negative
e66