Machine learning for keratoconus patients

Poster Details

First Author: C.Peris Martinez SPAIN

Co Author(s):    M. A. Valdes   J. D. Martin   F. Pastor Pascual    M. J. Ruperez      

Abstract Details

Purpose:

The main goal of this work is to profile patients undergoing keratoconus who have been treated by means of intracorneal rings. The models have been obtained using Machine Learning approaches. The outcome of the models enables ophthalmologists to predict the vision gain depending on the position of the rings, thus becoming a decision support system to plan the most appropriate therapeutic strategy for future patients.

Setting:

Cornea and Refractive Surgery Unit. Fundación Oftalmológica del Mediterrįneo (FOM), Valencia, Spain.

Methods:

A cohort of 287 patients undergoing keratoconus was included in the study. Patients were treated with intracorneal rings from 2008 to 2012. The variables included in the study were the following: Avcc, sphere, cylinder, K1, astigmatism, IOPcc, IOPg, CRF, CH and surgically variables. After a first preliminary data analysis, the data set used for ulterior analysis with Machine Learning techniques was reduced to 74 patients, being the variables that appeared to be relevant those related to biomechanical properties. After a preliminary data analysis and a clustering approach to find profiles of typical patients, the efforts were devoted to find models able to predict the vision gain after an intervention with intracorneal rings. Two basic kinds of approaches were taken into account: • Linear models, such as multiple regression and robust regression. Those approaches obtain a linear relationship between the input variables mentioned previously and the output variable (in our case, the vision gain). • Nonlinear models (universal function approximators), such as decision trees (DTs) and artificial neural networks (ANNs). A DT is a nonparametric statistical method that is used both for classification and regression; it is based on minimizing the entropy of splitting data in similar groups.

Results:

Nonlinear models outperformed linear models, being the ANN the most accurate approach; in particular, Mean Absolute Error (MAE) in ANN was 42% (1.66D) while in linear models was 55% (2.17D). However, linear approaches also obtained relatively accurate models with the additional advantage of being models that can be easily interpreted by non-experts in Machine Learning. Summarizing, models are able to predict with an error of just 1.6 D the vision gain after surgery; it can be considered quite a good since many patients usually have an astigmatism of 15D.

Conclusions:

Models based on ANNs seem to be a suitable approach for the tackled problem, as it is shown in the achieved results shown in the previous section. The obtained models can be used as a decision support system to help the surgeon in making a decision on the number and position of intracorneal rings for patients undergoing keratoconus. Furthermore, the model also shows up the most relevant variables in the description of the problem; in particular, biomechanical variables turned out to be especially important for this kind of operation. Future work will be devoted to producing a biomechanical model of the cornea to explain the vision gain that has been modeled using different approaches in the present work. FINANCIAL DISCLOSURE?: No