Abstract
Purpose :
This study aims to determine the patient's refractive error from the foveal pit's morphology using tridimensional modeling of the foveal pit and machine-learning algorithms.
Methods :
Eighty-five subjects (47 males and 38 females) with ages ranging from 18 to 27 were evaluated. We divided the study into three stages: (1) clinical, (2) extraction of the foveal pit morphology, and (3) machine-learning models. (1) The axial length (AL) and spherical equivalent (SE) of each participant were obtained with an IOL Master 700 and an autorefractometer. The spherical equivalent was also verified through wet retinoscopy. We classified the groups in terms of their mean axial length and spherical equivalent. (2) A 3D foveal pit model of each participant was obtained from images obtained with a Spectralis OCT (Heidelberg Engineering). We obtained three areas along with the pit: the rim, the mid, and the flat. The volume of the pit, the macula, and the retina were also calculated. (3) Then, we trained seven machine learning models: Artificial Neural Network (ANN), Classification and Regression Trees (CART), k-Nearest Neighbors (kNN), Linear Discriminant Analysis (LDA), Ordinal Logistic Regression (OLR), Support Vector Machines (SVM), and Random Forest (RF).
Results :
(1) The short-eye group has an average axial length of 22.5 mm (21.82/22.89 mm) and a spherical equivalent of +0.5 D (+0.25/+1.5 D), the emmetropic group has an average axial length of 23.62 mm (23.08/24.46 mm), and an average spherical equivalent of -0.25 D (-2.0/+1.25 D), and the long-eye group has 25.48 mm (24.7/26.60 mm) of axial length and -3.75 D (- 6.0/+0.5 D) of spherical equivalent. (2) We found that the rim's area presents the highest correlation with the refractive error of the patient. Mid and the flat areas are also good predictors of the patient's refractive status. Pit volumes showed a strong association with participants' refractive status. (3) The random Forest model shows the best performance with a predictive power of 77%.
Conclusions :
The random Forest model shows the best performance in predicting axial refractive errors. This model was able to properly classify 77% of the subjects in one of the three groups: short-eyes (21 mm axial length), normal-eyes (23 mm axial length), and long-eyes (25 mm axial length).
This is a 2021 ARVO Annual Meeting abstract.