Training data size and predication errors in the use of machine-learning assisted intraocular lens power calculation

Abstract

AbstractThis retrospective study examined the effect of the size of training data on the accuracy of machine learning-assisted SRK/T power calculation. Clinical records of 4800 eyes of 4800 Japanese patients with intraocular lenses (IOLs) were reviewed. A support vector regressor (SVR) was used for refining the SRK/T formula, and dataset sizes for training and evaluation were reduced from full to 1/64. The prediction errors from the postoperative refractions were calculated, and the proportion within ± 0.25 D, ± 0.50 D, and ± 1.00 D of errors were compared with those using full data. The influence of the difference in A-constant was also evaluated. Prediction errors within ± 0.50 D in the use of full data were obtained with the dataset of ≥ 150 eyes (P = 0.016), whereas the datasets of ≥ 300 eyes were required for the error within ± 0.25 D (P < 0.030). The prediction errors did not alter with the A-constant values among IOLs with open-loop haptics, except for IOLs with plated haptics. In conclusion, the accuracy of SVR-assisted SRK/T could be achieved with the training dataset of ≥ 150 eyes for the Japanese population, and the calculation was versatile for any open-looped IOLs.