Demystifying AI in IOL Power Calculations

Does AI disguise more than it explains? Dermot McGrath reports from the 39th Congress of the ESCRS in Amsterdam.

Default banner image for Demystifying AI in IOL Power Calculations
Photo of Dermot McGrath

Does AI disguise more than it explains? Dermot McGrath reports from the 39th Congress of the ESCRS in Amsterdam. Although there is considerable hype around the potential of artificial intelligence (AI) as a means to improve IOL power calculation accuracy, the reality is a big data approach is less accurate than a physical method such as raytracing, according to Paul-Rolf Preussner MD, PhD. “A big data approach can be more accurate than classical IOL formulas but not more accurate than a physical approach such as raytracing—particularly so in complex eyes that fall outside the ‘normal’ range of measurements,” Dr Preussner told the conference. Big data also typically need adjustment for different IOL models and are not appropriate for post-LASIK eyes. “You cannot take the results from one IOL model and try to apply it to another lens model without losing accuracy. A big data approach is also not applicable to post-LASIK eyes where a full corneal tomography is needed,” he said. Creator of the OKULIX raytracing software, Preussner said he considers the term “artificial intelligence” a bit of a misnomer when it comes to discussing algorithms for IOL power calculation. “To be frank, I do not really like the term ‘artificial intelligence’ for software algorithms because I think it disguises more than it explains. Things mostly appear much simpler once you have cracked the mystifying shell.” In IOL power calculation, the key is to understand the mathematical structure of the impact of the input variables (corneal and axial data) on the output (IOL power and refraction), he explained. The predicted refraction differences can be represented graphically using pseudo-colours for axial eye length and corneal radius function as input valuables. “These functions are smooth, steady, and monotonic, which means the problem is actually mathematically simple,” he said. Dr Preussner conducted a study comparing the predicted refraction differences between the SRK/T, Hoffer Q, Holladay I, and Haigis formulas, as well as the Hill RBF algorithm (the “big data” approach) and OKULIX raytracing software in 6,004 eyes implanted with an Alcon SN60WF lens. The Hill RBF algorithm and raytracing method were also compared to a simple two-dimension linear interpolation algorithm (LINEAR) using the same approach. In terms of the results, Dr Preussner said that comparing the accuracy of the different methods is a “bit like comparing apples with pears” because of the different number of used input variables. In the study of 6,004 eyes, the Hill RBF using three variables performed better than OKULIX with two input variables, which was better than LINEAR. However, in a subset of 908 eyes and using four input variables for OKULIX (axial length, corneal radius, anterior chamber depth, and lens thickness), OKULIX performed better than Hill RBF and LINEAR. “All of these methods—Hill RBF, OKULIX, and LINEAR— were better than standard formulas. The differences were statistically significant but are not very relevant clinically,” Dr Preussner said. Paul-Rolf Preussner MD, PhD, University Eye Hospital, Mainz, Germany. pr.preussner@uni-mainz.de