Employing deep learning for prediction of the refractive error after cataract surgery

Aims/Purpose: In addition to good visual acuity, the refractive outcome after cataract surgery is an important indicator of the success of the operation. Many preoperative factors can increase the risk of a high refractive error after cataract surgery. Neural networks are predestined to extract meaningful information from such complex data sets. This study therefore aimed to establish a neural network to evaluate the risk of high refractive error individually before cataract surgery based on various factors.Methods: This is a retrospective, monocentric study that was approved by the responsible ethics committee. Following strict quality control, a total of 1357 eyes with cataract surgery were included. Preoperative biometry was performed on all eyes using the IOL Master 700 (ZEISS). The following factors were considered for the study: K1, age, K1 of the fellow eye, lens thickness, axial length, corneal thickness and corneal thickness of the fellow eye, anterior chamber depth, chord alpha and main diagnosis (e.g. cataract, epiretinal membrane, amotio retinae). The eyes were then annotated; a postoperative refractive error of more than 0.75 diopters deviation was classified as a high refractive error. After annotation, the dataset was divided into a training dataset, a validation dataset and a test dataset.Results: The neural network was trained using the training and validation data sets. The accuracy of the correct classification was then evaluated using the unknown test data set. Using the neural network, an accuracy of 83% was achieved for the classification into a high or low postoperative refractive error.Conclusions: Using the developed neural network, we were able to show that a prediction of a high postoperative refractive error after cataract surgery was only possible with the help of values from biometry in more than 8 out of 10 patients, which could improve preoperative counseling.