Kateryna Burlak1, Colin Sheppard4 ,2, Marc Sarossy4 ,3
Purpose: To investigate the application of machine learning methods with additional input parameters in calculating the ideal IOL.
Methods: A retrospective review of 70 patients (107 eyes; M:F ratio of 19:51; average age of 72) who underwent cataract surgery with the same surgeon (CS), from July 2011 to March 2016. The ideal IOL (mean of 20.6D; range 9.9-29.6D) was calculated from the implanted Bausch + Lomb LI61 SofPort aspheric IOL and the post-operative refraction. Machine learning methods were used to explore the relationship between input parameters (age, gender, eye, axial length, corneal refractive power, pre-operative spherical equivalent, anterior chamber depth) against the ideal IOL. The following machine learning models were fitted using R after centering each of the input features in the input matrix (transformation to zero mean and unit variance): Linear model (LM), Multiple adaptive regression splines (MARS) model, and Kernel regression (KR). These models were compared to SRK-T, Haigis, Holladay 1 and Barrett Universal II formulas, using the residual sum of squares (RSS) to estimate their IOL prediction accuracy.
Results: RSS using all predictors: KR (3.41), MARS (15.23), LM (19.27), Barrett Universal II (24.21), SRK-T (35.09), Holladay 1 (35.71), and Haigis (49.28). The respective root-mean squared error: 0.18, 0.38, 0.42, 0.48, 0.57, 0.58 and 0.68.
Conclusion: KR showed the most promise, likely due to allowing for non-linearities and interactions. Additional input features allowed better predictive accuracy for all of the new methods compared with the existing formulas. Further study will generalize the technique to more lenses and surgeons.