AI in IOL Power Selection

Substituting a machine learning (ML)-derived prediction of postoperative anterior chamber depth (ACD) for effective lens position (ELP) estimates can improve the accuracy of many existing intraocular lens (IOL) power formulas.

However, with power formulas that are primarily ML-derived, there is a trade-off between generalisability and customisation. Ideally, they should be optimised for specific patient populations and lenses. New measures are also needed to evaluate their performance, said Nambi Nallasamy MD.

ML and existing formulas

ACD is a measure of the true IOL position after surgery that can be predicted based on preoperative biometry using ML-derived algorithms, Dr Nallasamy said. By contrast, ELP is the distance between the anterior surface of the cornea and the plane of the IOL, if it were infinitely thin. As such, ELP is a theoretical estimate that is back-calculated based on postoperative refractive error, or a “fudge factor,” Dr Nallasamy explained. “[ELP estimates] are really not accurate in existing formulas.”

To test whether using ML-derived ACD predictions could improve IOL power prediction accuracy, Dr Nallasamy directly measured the ACD in 1,205 eyes of 678 patients implanted with AcrySof IQ SN60WF (Alcon) monofocal IOLs. Measured by mean absolute error (MAE), the resulting base model and the base plus a model that included IOL data both predicted the final IOL position significantly more accurately than the Haigis, Hoffer Q, Holladay 1, Olsen, and SRK/T formulas, as well as a linear regression model, Dr Nallasamy reported. A model not considering keratometry for use in ectatic and post-refractive surgery patients also outperformed the existing formulas.¹

In a related study, Dr Nallasamy and colleagues substituted their ML-derived ACD predictions for ELP in four existing vergence formulas. The ML-modified ELP significantly improved the performance of the Haigis, Hoffer Q, Holladay, and SRK/T as measured by MAE. The study was based on a training set of 5,761 eyes (80%) and a test set of 961 eyes (20%) of 4,806 patients.² Similarly, the ML-powered ACD prediction improved the performance of the OKULIX raytracing formula.³

ML-derived IOL formulas

Several currently available IOL power formulas are based primarily on ML-derived algorithms, including the Pearl-DGS, Hill-RBF, Kane, and Nallasamy formulas. However, Dr Nallasamy stressed the trade-off between accuracy and generalisability must be addressed for them to be reliable in the real world.

Though ML-derived formulas can be very accurate within their training sets and similar populations, they also tend to take on the biases of these sets. For example, medium axial lengths and 0.0 D to -3.0 D refractive targets are generally over-represented, resulting in formulas that are less accurate for patients with shorter or longer eyes or unusual refractive targets. “If a machine sees that data and only that data, it’s going to start thinking that predicting zero all the time would make a pretty good formula, and that’s really a problem,” Dr Nallasamy said.

Dr Nallasamy reviewed several methods to address this ‘overfitting.’ One method uses an ensemble approach to derive the formula, in which the first step produces multiple models based on raw biometric and postoperative refractive data, and the second step produces a model based on the predictions of the first-step models plus postoperative refraction. This resulting second-step model tends to even out the biases the first-stage models may pick up from the raw data. Cross-validation further ensures a less biased model.⁴

Even so, ML models may need further optimisation for different lenses and patient populations, Dr Nallasamy said. IOL constants and transfer learning, in which the ML model is applied to a data set from the targeted population, can markedly improve performance. For example, using transfer learning to optimise the Nallasamy formula helped it outperform several existing formulas using IOL constant optimisation when applied to data from the Aravind eye system in India, as measured by MAE, he reported. Incorporating detailed modelling of the properties of the IOL and the eye can also improve ML formula performance without further optimisation.

However, MAE and median absolute error alone—which measure refraction errors—are not reliable for evaluating the generalisability of ML-based formulas, Dr Nallasamy said. Alongside them, he proposes using mean absolute error of the prediction of the IOL (MAEPI), which measures IOL power prediction errors rather than refractive error, and the correct IOL rate (CIR), which measures the proportion of IOL power prediction errors within 0.5 D of target. A high MAEPI and a low CIR suggest an overfit model.⁵ “We want to be very careful with our imbalanced data sets,” he said.

Dr Nallasamy presented in the JCRS Symposium at the 2025 ASCRS annual meeting in Los Angeles.

Nambi Nallasamy MD is assistant professor of ophthalmology and visual science and of computational medicine and bioinformatics at the University of Michigan Kellogg Eye Center, Ann Arbor, Michigan, US. nnallasa@med.umich.edu

1. Li T, Yang K, Stein JD, et al. Transl Vis Sci Technol, 2020 Dec 21; 9(13): 38.

2. Li T, Stein J, Nallasamy N. Br J Ophthalmol, 2021 Apr 9; 106(9): 1222–1226.

3. Li T, Reddy A, Stein JD, et al. Br J Ophthalmol, 2023 Apr; 107(4): 483–487.

4. Li T, Stein J, Nallasamy N. Br J Ophthalmol, 2022 Apr 4; 107(8): 1066–1071.

5. Li T, Stein JD, Nallasamy N. Transl Vis Sci Technol, 2023 Mar 28; 12(3): 29.