Purchase this article with an account.
Christopher Anderson Clark; Artificial intelligence to predict longitudinal changes in refractive error. Invest. Ophthalmol. Vis. Sci. 2020;61(7):569.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Refraction for spectacles, contact lenses and refractive surgery requires significant time. Even with refraction, up to 20% of patients are unsatisfied with their final spectacle prescriptions. Artificial Intelligence (AI) offers a potential ability to improve patient’s vision while reducing testing time and errors. AI has already been used to predict refractive error from fundus images. The purpose of this study is to determine the feasibility of predicting refractive error changes using AI.
Two data sets were used for this study. The first study (N=309, ages 8-15, training set = 200, test set = 109) was a longitudinal study with changes in refractive error over 1 year with near and distance visual acuities (VA) but no retinal images. The second longitudinal study (N = 1036, ages 8-40, training set = 700, test set = 336) had changes in refractive error over 1 year without Vas but had retinal SD OCT images. Both data sets had baseline refractive error, pupil size, age, gender, keratometry, corneal asphericity, angle alpha, and lag of accommodation. Axial length and anterior chamber depth were not used in the training of any models. AI models were run on both data sets using custom Matlab software (IBM, Armonk, NY). Training sets were used on approximately two thirds of the data and test sets used the remaining data, all randomly selected. All refractive measurements were converted to spherical equivalent (SE), J0 and J45 using power vectors to simplify the processing time. All R2 values are reported from the test data.
Visual acuity data improved the AI’s ability to predict longitudinal refractive error regardless of the presence or absence of retinal image data (SE R2 = 0.92, J0 R2 = 0.46, and J45 R2 = 0.36.) The largest contributor to predicting longitudinal refractive error changes was baseline refractive error (full model R2 = 0.923, baseline refractive error alone R2 = 0.80.) Using SD OCT images alone, AI was capable of predicting SE (R2 = 0.30, P = 0.001) but unable to predict J0 or J45 astigmatism (R2 = 0.02, R2 = 0.03.) Using Bland-Altman, the average difference between measured refractive error and SD OCT predicted error was 0.26D, which is approximately half the error predicted by color fundus images.
AI can predict refractive error from a number of sources including baseline refraction, visual acuity and retinal images. This may reduce examination time and reduce errors.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only