Abstract
Purpose :
For patients that undergo cataract surgery, Spectacle Independence (SI) is one of the desired clinical outcomes, especially when implanted with presbyopia correcting intraocular lenses (IOLs). Although Through Focus Visual Acuity (TFVA) can be accurately predicted based on preclinical simulations and measurements [1], there is a need for a robust model to allow predictions of SI ahead of clinical trials.
Methods :
It is expected that the TFVA curve measured on an implanted subject is related to the SI outcome. However, establishing a direct correlation between TFVA and SI is challenging. In this work, we build on previous work and further characterize the link between observed TFVA curves and the SI outcome. We use a Machine Learning (ML) approach to train several different algorithms to classify TFVA curves into SI and non-SI. These include LR, LDA/QDA, KNN, NB, SVM. The training data used for these classification algorithms are binocular TFVA curves (from -3D to 0D, 0.5D steps) and answers to questionnaires related to the subjects SI compiled from multiple clinical studies with a variety of implanted lens models. We use a subset of this full clinical set as a validation check on the different algorithms and compare the results to those obtained with the model developed previously. Additionally, we compare the performance impact of also including each subject’s manifest refraction in the training data.
Results :
Results show an improvement with respect to the accuracy (at least 10%) of our previous model when classifying curves from the validation dataset. We also find that the strongest predictors of SI in this context are the VA scores in the near range (-3 to -2D) consistent with previous findings. We also compare the predictions of SI for average clinical and preclinical curves with the observed SI in the clinical datasets showing good correlation between them. Including manifest refraction in the training data did not noticeably improve prediction accuracy.
Conclusions :
We conclude that the tested ML models, TFVA can be used to predict the SI outcome with a reasonable degree of accuracy and can already give directional information when used with preclinical data. They give improved accuracy as compared to the previous model. However, the attained accuracy (above 70%) may indicate that other factors could also play a role in the SI and could be used to improve predictions further.
This is a 2021 ARVO Annual Meeting abstract.