Research Article

Using neural networks to predict the outcome of refractive surgery for myopia

¹ Ophthalmica Institute of Ophthalmology and Microsurgery, V. Olgas 196, Thessaloniki 546 55, Greece
² Faculty of Medicine, Aristotle University of Thessaloniki, Thessaloniki 54124, Greece
³ Laboratory of Information Technologies, Faculty of Information Science and Informatics, Ionian University, Corfu 49100, Greece
⁴ Department of Ophthalmology, Cornea, Cataract and Refractive Surgery, University Eye Hospital Basel USB, Mittlere Strasse 91, 4031 Basel, Switzerland
⁵ Association for Training in Biomedical Technology, 6 Aristogeitonos Street, Thessaloniki 54628, Greece

^* Corresponding author: anogian@hotmail.com

Received: 21 May 2019
Accepted: 24 September 2019

Abstract

Introduction: Refractive Surgery (RS), has advanced immensely in the last decades, utilizing methods and techniques that fulfill stringent criteria for safety, efficacy, cost-effectiveness, and predictability of the refractive outcome. Still, a non-negligible percentage of RS require corrective retreatment. In addition, surgeons should be able to advise their patients, beforehand, as to the probability that corrective RS will be necessary. The present article addresses these issues with regard to myopia and explores the use of Neural Networks as a solution to the problem of the prediction of the RS outcome.

Methods: We used a computerized query to select patients who underwent RS with any of the available surgical techniques (PRK, LASEK, Epi-LASIK, LASIK) between January 2010 and July 2017 and we investigated 13 factors which are related to RS. The data were normalized by forcing the weights used in the forward and backward propagations to be binary; each integer was represented by a 12-bit serial code, so that following this preprocessing stage, the vector of the data values of all 13 parameters was encoded in a binary vector of 1 × (13 × 12) = 1 × 156 size. Following the preprocessing stage, eight independent Learning Vector Quantization (LVQ) networks were created in random way using the function Ivqnet of Matlab, each one of them responding to one query with (0 retreat class) or (1 correct class). The results of the eight LVQs were then averaged to permit a best estimate of the network’s performance while a voting procedure by the neural nets was used to arrive at the outcome

Results: Our algorithm was able to predict in a statistically significant way (as evidenced by Cohen’s Kappa test result of 0.7595) the need for retreatment after initial RS with good sensitivity (0.8756) and specificity (0.9286).

Conclusion: The results permit us to be optimistic about the future of using neural networks for the prediction of the outcome and, eventually, the planning of RS.