Accuracy of Artificial Intelligence Formulas and Axial Length Adjustments for Highly Myopic Eyes.

Abstract

To compare the accuracy of artificial intelligence formulas (Kane formula and Radial Basis Function [RBF] 2.0) and other formulas, including the original and modified Wang-Koch (MWK) adjustment formulas for Holladay 1 (H1-MWK) and SRK/T (SRK/T-WK and SRK/T-MWK), the Barrett Universal II (BUII), the emmetropia-verifying optical (EVO), and the Haigis equation in highly myopic eyes. Retrospective consecutive case-series study. A total of 370 eyes with an axial length (AL) ≥26.0 mm of 370 patients were enrolled, and subgroup analyses was performed based on ALs. The median absolute error (MedAE), the percentages of eyes with hyperopic outcome and within ±0.25 diopters (D), ±0.50 D, and ±1.00 D of prediction error were determined. Overall, the Kane equation had the lowest MedAE (0.26 D), followed by H1-WK (0.27 D) and H1-MWK (0.28 D). There were no significant differences in MedAE among the Kane equation, the RBF 2.0, the BUII, the H1-MWK, and the H1-WK, whereas the Kane equation had a significantly lower MedAE than EVO (P < .001), SRK/T-MWK (P = .001), SRK/T-WK (P = .006), and Haigis (P < .001). In extremely myopic eyes with an AL ≥30.0 mm (n = 115), the Kane equation had a significantly lower MedAE than the RBF 2.0 (P = .001), the EVO (P = .019), the BUII (P = .013), and the Haigis method (P = .005), whereas no significant differences were found among the Kane, H1-MWK, and H1-WK equations. The Kane equation was comparable to RBF 2.0, BUII, H1-MWK, and H1-WK in highly myopic eyes and was better than RBF 2.0 and BUII in extremely myopic eyes. The Kane, H1-MWK, and H1-WK methods were equally accurate in eyes with high to extreme myopia.