Abstract
Purpose :
Using Artificial Intelligence techniques to develop an accurate prediction algorithm for subjective refraction from wavefront aberrometry. Data were analyzed with a new aberration series in which higher order polynomials are devoid of linear and quadratic terms to better fit the low and higher order components of the wavefront.
Methods :
A database study of 17,900 ocular wavefront exams captured by an aberrometer (OPD-Scan III ®, Nidek, Gamagori, Japan) and corresponding non-cycloplegic subjective refractions conducted on the same day by an experienced optometrist was analyzed. Refraction datasets were vectorized using power vector analysis. The aberrometer was specifically programmed to use the new aberration series expansion to decompose the data on a 4.5 mm pupil. The coefficients g(n,m) of the new polynomial modes G(n,m) were computed for each sampled ocular wavefront up to the 6th radial order. A training set constituting 80% of the data was randomly selected. Three machine learning models were separately trained to predict each component of the power vector analysis (M, J0, J45) from the weighted coefficients of the new aberration modes. The accuracy of the model was evaluated on the test set consisting of the remaining 20% of the data, never seen by the model, to avoid overfitting.
Results :
Our model had a Root Mean Square Error (RMSE) of 0.48 when predicting the M component of the power vector analysis. RMSE was 0.22 for the J0 component and 0.30 for the J45 component. The most influential Higher Order modes coefficients found on primary analysis were: M: g(2,0), g(4,0), g(0,0), g(3,-1); J0: g(2,2), g(1,1), g(4,0), g(3,-1); J45: g(2,-2), g(6,2), g(2,2), g(4,2).
Conclusions :
Subjective refraction can be predicted with acceptable clinical precision using machine learning and new aberration series wavefront data analysis. Higher order modes devoid of low order terms influence the outcome of subjective refraction.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.