Abstract
Purpose :
To assess the performance of Machine Learning (ML) regression models in the prediction of the spherical component (M) of Subjective Refraction (SR) by including information about the accommodative response.
Methods :
The Accommodative Response (AR) data of 190 eyes was measured during a sweep of lenses with gradual decreasing powers +2.00, +1.50, +1.00, +0.75, +0.50, +0.25, 0.00, -0.25, -0.50, -0.75, -1.00, -1.50 and -2.00 D relative to the SR using a custom-developed Hartmann-Shack aberrometer coupled to a phoropter. Previously, objective refraction (OR) was measured with a commercial autorefractor, and SR was performed with the conventional procedure. Three ML models (Normal Equation (NE), Gradient Descent (GD) and Extreme Gradient Boosting (XGB)) were trained and tested for the M component of SR. The models were trained on a dataset with 142 eyes and tested with 48 test samples. Input variables were the measured M values with each lens of the sweep, the coefficients of a fourth-degree polynomial adjustment to the AR, OR and age. The performance and agreement between real and predicted values of SR was done with the Bland-Altmann analysis, Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and percentage of agreement for thresholds of 0.25 and 0.50 D. Moreover, features weight was analyzed.
Results :
Subjects ages ranged from 20 to 73 years and spherical equivalent error from -6.88 D to +5.00 D. The MAE and RMSE for the three ML models showed a significantly better performance compared to OR (Table 1). NE and GD performed similarly in terms of MAE, RMSE, agreement in the thresholds of 0.25 and 0.50 D and in the Bland Altman analysis. The XGB performed better than OR, however results were poorer than those obtained with the linear regression models. Regarding the analysis of the feature weights, the input variables with higher values, in descent order, were the measured M with lens of -0.25 D, the measured M with 0 D, OR, the measured M with +0.75 D, and the third-degree term of the polynomial coefficients.
Conclusions :
ML determination of spherical refraction based on the accommodative response provided promising results under the tested conditions, with LoA of ±0.25 D for the best model. This proof-of-concept of the models should be ratified in a larger data set.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.