Machine learning technology for predicting intraocular lens power. Diagnostics data generalization

Abstract

Backgraund. Implantation of recent IOLs allows ophthalmologists to effectively solve the problems of surgical rehabilitation of patients with cataracts. The degree of improving the patient's visual functions is directly dependent on the accuracy of the preoperative calculation of the optical power of IOLs. The most famous formulas to calculate this indicator are SRK II, SRK/T, Hoffer-Q, Holladay II, Haigis, Barrett. All of them work well for an "average patient", however, they are not adequate enough at the boundaries of input variables ranges.
 Aims. To study the possibility of using mathematical models obtained as a result of deep learning of artificial neural networks (ANN models) to generalize data and predict the optical power of modern intraocular lenses.
 Materials and methods. ANN models were trained on large-scale samples, including depersonalized data for patients of ophthalmology clinic. Data provided in 2021 by ophthalmologist K.K. Syrykh, and reflect the results of both preoperative and postoperative observation of patients. The source file used to build the ANN model includes 455 records - 26 columns of input factors and one column for the output factor - calculating IOL (diopters). To conveniently build ANN models, we used a simulator program, previously developed by the authors.
 Results. The resulting models, in contrast to the traditionally used formulas, reflect the regional specificity of patients to a much greater extent. They also make it possible to retrain and optimize the structure based on newly received data, which allows taking into account the non-stationarity of the object. A distinctive feature of such ANN models in comparison with the well-known formulas SRK II, SRK/T, Hoffer-Q, Holladay II, Haigis, Barrett, widely used in the surgical cataract treatment in ophthalmology, is their ability to take into account a significant number of recorded input quantities, which makes it possible to reduce the mean relative error in calculating the optical power of IOL from 10–12% to 3.5%.
 Conclusions. This work shows the fundamental possibility of generalizing a significant amount of empirical data on calculating the optical power of the IOL using training ANN models that have a significantly larger number of input variables than when using traditional formulas and methods. The results obtained allow, in principle, to build an intelligent expert system with a continuous flow of new data from a source and a step-by-step retraining of the ANN model.