Prediction Of Progression In Keratoconus Patients Defined By Adjusted Kmax Values Using Machine Learning

Keratoconus is an often asymmetric occurring disease which is characterised by progressive thinning and steepening of the central cornea. In order to classify the progression of the disease, at least two topographic measurements of the cornea with a time interval in between are required to detect progression. The change in the K_max value is often assessed for this purpose. The aim of this study is to use machine learning to predict the progression of K_max values based on a single measurement.

All study-related measures as well as data collection were performed at the Vienna Institute for Research in Ocular Surgery (VIROS), Department of Ophthalmology, Hanusch Hospital, Vienna, Austria and the Department of Ophthalmology, Kepler University Clinics, Linz, Austria.

Keratoconus patients that had two measurements performed with a Placido tomograph (MS-39, CSO) with a time interval of at least four months were included in this retrospective study. The eyes were classified as ‘progressive’ or ‘non-progressive’ based on the repeatability-adjusted difference in K_max values adjusted to the tomograph. Using a wrapper algorithm, suitable predictors were selected, and a random forest algorithm was trained. The algorithm was validated using an external data set (25% of the data).

A total of 158 eyes from 95 patients were included. The mean age was 31.8±11.7 years. Especially various posterior K values and the geometric centres of the keratometric measurements proved to be promising predictors for keratoconus progression. During internal validation, the algorithm achieved an area under the receiver operating characteristic (AUROC) curve of 0.87, a sensitivity of 63.5% and a specificity of 87.4%. In the course of the external validation, an accuracy of 83.9% was observed. In addition, the sensitivity and specificity in the external test set were 75.0% and 87.0%, respectively.

This study shows that the progression of the K_max value in keratoconus patients can be predicted with high accuracy using a single tomographic measurement. In particular, radii of curvature of the posterior corneal surface seem to play an important role in the prediction of keratoconus progression. Due to the low number of cases, further analyses using larger data sets are necessary to be able to implement the algorithm in clinical practice.