Abstract
Purpose :
To test machine learning models to predict subjective ocular refraction from patients’ demographics and ophthalmological data, and to compare the model’s performance with an automatic refractometer.
Methods :
Dataset comprised ophthalmic examination data of 17,039 eyes from the TeleOftalmo, a teleophthalmology project in the Brazilian public health system. We collected the following variables to be tested as attributes in the predictive model: age, gender, race, symptoms, uncorrected visual acuity, best corrected visual acuity, pinhole visual acuity, intraocular pressure (Visuplan, Zeiss, Germany), current spectacles’ dioptric power, keratometry measurements and automatic refraction (Visuref, Zeiss, Germany). Same-day subjective refraction performed by an ophthalmologist was defined as the target attribute. Subjective refraction was converted into power vectors (M, J0 and J45). We used the Orange Data Mining toolbox in Python to run the tests. Performances of Random Forest, Linear Regression, and Neural Network (Multi-Layer Perceptron) algorithms in predicting subjective refraction were assessed in terms of mean absolute error (MAE) and root mean square error (RMSE). For comparison purposes, we determined the automatic refraction error, defined as the M component’s difference between automatic refraction and subjective refraction without a predictive model.
Results :
Machine learning model using Random Forest algorithm had the best performance. Table 1 presents MAE and RMSE for M, J0 and J45 components with each algorithm. Main predictors among all attributes were: automatic refraction’s M, J0, J45, spherical power, and cylindrical power; lensmeter’s spherical power; uncorrected visual acuity; and keratometric astigmatism. Compared to automatic refraction error (0.71 ± 1.35 diopters), predictive model had a lower mean error (0.31 ± 0.61 diopters, p<0.01).
Conclusions :
Artificial intelligence can predict subjective refraction from clinical data. Automatic refraction, uncorrected visual acuity, keratometry and current spectacle power provided the main predictors to the model. Automatic refraction alone had a significantly higher error than the predicted subjective refraction using machine learning algorithms.
This is a 2020 ARVO Annual Meeting abstract.