We read with great interest the letter from Soto-Pedre et al. (1), which offered several observations on our article (2). Lin et al. (3) recently concluded that the optimal strategy for diabetic retinopathy (DR) screening remains uncertain, and a number of possible modalities have been considered. Our study, which started in 1998, intended to reaffirm the need of validation of digital technology that has come up in recent years (4,5). Including digital retinal images in health records (6) will improve the management of diabetic patients.
We agree with Soto-Pedre et al. on the suitability of retinal photography for the screening of DR, as we have already mentioned in our article. Our objective was to examine the reliability of the digitally captured retinal image, subsequently compressed and sent to a reference center via internet for analysis. For this we compared the digital image inspection to the direct examination of the eye fundus under mydriasis. This is important at the moment because the increasing use of new technologies makes the need for validating the use of digital images almost totally necessary for DR screening (6–10).
Soto-Pedre et al. cast doubt over some methodological issues such as age, sex, diabetes type, etc., that could invalidate our conclusions. This information, though collected as part of the protocol of study, has not been included in the article. We decided not to because the aim of the study was simply to compare the findings in the digital image with those findings obtained by direct examination of the eye fundus, regardless age, sex, or other parameters mentioned by Soto-Pedre et al. We agree, however, that factors such as those mentioned above may impair image quality and therefore reduce the agreement between direct inspection and digital image examination. We expect that further studies will clarify this and other questions.
The proposed method would function as population screening in all cases in which the eye fundus can be photographed with good quality. As mentioned in our article, we were unable to obtain photographs of the eye fundus in 5% of the initial 140 eyes because of a cataract; in 5% of the remaining 133 eyes, the quality of the digital image was not good enough to make a diagnosis because of pupil narrowing or moderate media opacity. Thus, in 10% of patients, the method used was not adequate, and the patient had to be referred to an ophthalmologist. In another paper published by our group (11), we also found that this problem appeared in 13% of cases, which is in agreement with observations reported by other authors (9). Some studies report figures that range from 3.7 to 22% (12). Recent data suggest that the degree of technical error would be <5% when mydriasis is performed (13).
Since the first formal introductions of κ, more than 40 years ago, numerous papers both using and criticizing its various forms have appeared in the statistical as well as in the medical literature. κ coefficients were designed to measure correlation between nominal, not ordinal, measures (14). In a later study, Cohen and Fleiss (15) have shown that under a squared error weighting system, weighted κ is asymptotically equivalent to the intraclass correlation computed using the category ranks. Although there are several inherent problems in the use of weighted κ statistic for the analysis of ordinal agreement (15), we believe that weighted κ could be a reasonable (though not a unique) choice as a reliability measure. The intraclass correlation has been advocated as an agreement index for both continuous and ordinal data (16), and we have preferred this approach because of the high number of categories for grading diabetic retinopathy.
References
Address correspondence to Francisco Gómez-Ulla or Francisco Gonzalez, Department of Ophthalmology, Hospital Provincial de Conxo, C/Ramón Baltar s/n, E-15706, Santiago de Compostela, Spain. E-mail: [email protected] or [email protected].
