Forecasting Progressive Trends in Keratoconus by Means of a Time Delay Neural Network

Abstract

Early and accurate detection of keratoconus progression is particularly important for the prudent, cost-effective use of corneal cross-linking and judicious timing of clinical follow-up visits. The aim of this study was to verify whether a progression could be predicted based on two prior tomography measurements and to verify the accuracy of the system when labelling the eye as stable or suspect progressive. Data from 743 patients measured by Pentacam (Oculus, Wetzlar, Germany) were available, and they were filtered and preprocessed to data quality needs. The time delay neural network received six features as input, measured in two consecutive examinations, predicted the future values, and determined the classification (stable or suspect progressive) based on the significance of the change from the baseline. The system showed a sensitivity of 70.8% and a specificity of 80.6%. On average, the positive and negative predictive values were 71.4% and 80.2%. Including data of less quality (as defined by the software) did not significantly worsen the results. This predictive system constitutes another step towards a personalized management of keratoconus. While the results obtained were modest and perhaps insufficient to decide on a surgical procedure, such as cross-linking, they may be useful to customize the timing for the patient’s next follow-up.

Definitions for Keratoconus Progression and Their Impact on Clinical Practice

“…18 However, none of the five RCTs published afterward followed the recommendation. Moreover, there is no consensus on the cutoffs to demonstrate progression other than the general recommendation that change should be above the repeatability, 18,21,22 although repeatability of some variables highly depends on disease severity. 15,16 To answer the question "How much variation is considered clinically important?"…”

“…New and interesting methods have been proposed recently to stage KC and predict KC progression using artificial intelligence. 21,38 However, it may take a while until these approaches become available in the daily clinic. In the meantime, the knowledge obtained from this study should serve as guidance in future endeavors to develop easy-to-use tools for everyday practice.…”

“…19 Unsurprisingly, the ABCD progression display, following the consensus recommendations, is gaining popularity in the clinic. 14,20 Other than the general recommendation that change should be above the repeatability, 18,21,22 there is no consensus whatsoever on how to define KC progression. In a step toward more consistent criteria for KC progression, this study assesses how three sets of parameters (habitual criteria, ABCD progression display, and combinations of more consistent and monotonic variables 17 ) classify progression in a cohort of untreated KC cases.…”

Assessing the visual image quality provided by refractive corrections during keratoconus progression

“…As illustrated by the large error bars in Figures 2 and 3, one cannot forecast the progression of keratoconus in an individual patient as it is impossible to know the precise stage of their first presentation. Some degree of short-term forecasting may be possible using machine learning, 25 but this requires tomographical information beyond that provided by the present model. Another limitation is that the model does not consider ocular changes due to physiological ageing (e.g., changes in pupil size and exact pupil location, crystalline lens thickness, etc.).…”

Keratoconus Severity Classification Using Features Selection and Machine Learning Algorithms

“…Trained on a total of 3390 color-coded map images representing 4 eyes classes, the proposed system achieved an accuracy of 78.5% in keratoconus classification. Authors of [ 10 ] proposed an intelligent system based on time delay neural network (TDNN) to verify both the progression predictability using two prior tomography measurements and the system accuracy when labelling the eye as stable or suspect progressive. Obtained results showed a sensitivity of 70.8% and a specificity of 80.6% using data of 743 patients captured by Pentacam.…”