Accuracy assessment of artificial intelligence IOL calculation formulae: utilizing the heteroscedastic statistics and the Eyetemis Analysis Tool

Objectives: To analyse the accuracy of artificial intelligence (AI)-driven intraocular (IOL) calculation formulae, together with established formulae using the heteroscedastic methodology and the Eyetemis Analysis Tool. Methods: Data from 404 eyes who underwent uneventful phacoemulsification with implantation of the SN60WF IOL were retrospectively reviewed. IOL power calculations were performed using the Barrett Universal II (BUII), EVO 2.0, Hoffer QST, K6, Ladas Super Formula (LSF), Nallasamy, PEARL-DGS and RBF 3.0 formulae. The SD of the prediction error (PE), served as the primary metric for accuracy. The mean absolute deviation (MAD) and the predictability rates within intervals from ±0.25 D to ±1.50 D were also evaluated. The Eyetemis Analysis Tool was used for further validation. Results: The SD ranged from 0.468 (Nallasamy) to 0.510 (LSF). The Nallasamy formula had a significantly lower SD than the BUII (0.505, p = 0.025) and K6 (0.489, p = 0.022) formulae. The Nallasamy formula also exhibited the lowest MAD (0.358) with a significant difference compared with the Hoffer QST formula (0.384, p < 0.001). Finally, a significantly higher percentage of eyes achieving ± 0.50 D of the target refraction was seen using the Nallasamy formula (77.19%) compared with the Hoffer QST (71.04%, p = 0.019) and Ladas Super Formula (70.79%, p = 0.030) formulae. Conclusions: The Nallasamy formula, incorporating AI technology, demonstrated superior accuracy according to the analysis guidelines for PE statistics for non-gaussian datasets recommended by Holladay et al. and the online Eyetemis Analysis Tool.