Purchase this article with an account.
Marta Jiménez-García, IKRAM ISSARTI, Elke O. Kreps, Sorcha Ní Dhubhghaill, Carina Koppen, David Varssano, Jos J Rozema; Keratoconus progression forecast by means of a time delay neural network. Invest. Ophthalmol. Vis. Sci. 2021;62(8):513.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Early detection of keratoconus (KC) progression is of utmost importance for the prudent and economical use of crosslinking. If KC progression could be accurately predicted, the timing of the follow-up visits could be customized to the patient’s needs. The aim of this study was to verify whether the progressive trend of the ectasia can be forecasted based on two prior tomographies, and verify the accuracy of the predictive system when labelling eyes as stable or suspect progressive.
The observational multicenter REDCAKE study enrolled 906 KC patients measured at least twice. A time delay neural network was implemented. The network receives 6 features as input, measured in two consecutive examinations; all of them are potentially platform-independent (age, the mean keratometry in a 3mm area around the maximum curvature, the steepest radius and best fit sphere of the front surface, the mean radius of the back surface and LOGIK). Subsequently, the system predicts the values of the 2nd follow-up and determines its classification (stable or suspect progressive) based on the significance of the change from the baseline value (Figure 1).Different configurations and datasets were applied to evaluate robustness to errors and stability of the system. In the first dataset, 3 consecutive examinations of good quality were assembled to create each one of the triplets (N=811); in the second dataset, one examination was allowed to be of lower quality (marked in yellow by the tomographer software, N=1236). Data was divided as follows: 85% for training and 15% for external validation. Each configuration was trained/validated 10 times with random data splits.
The results obtained were modest, with an average sensitivity s=71% and specificity sp=81% when classifying eyes as stable or suspect progressive (Table 1). On average, the positive and negative predictive values were PPV=71% and NPV=80%, respectively. Also including time series with one error did not significantly worsen the results (s=72%, sp=78%, PPV=72%, NPV=78%).
The results obtained seem insufficient to decide on a surgical procedure such as crosslinking, however, they may be used to customize the timing for the next follow-up based on the predicted status. This predictive system constitutes another step towards a personalized management of KC disease.
This is a 2021 ARVO Annual Meeting abstract.
Predictive system for keratoconus progression
This PDF is available to Subscribers Only