Targeting More Accurate IOL Power Prediction

While artificial intelligence (AI) has been playing an increasingly important role in the development of intraocular lens (IOL) power calculation formulas, refractive outcomes using pure AI or hybrid AI formulas are only incrementally better than those achievable with non-AI-enhanced tools. Furthermore, achieving high levels of refractive predictability in extremely long and extremely short eyes remains a particular challenge, said Mitchell P Weikert MD.

“The AI-based formulas are moving the bar a little bit in a positive direction, but certainly there is not a big sea change in the results they are producing compared to geometric optics and raytracing formulas,” he commented.

“One limiting factor in the performance of the AI-enhanced formulas is that the supervised machine learning process used for their development and refinement is dependent on accurate postoperative refraction data.”

Speaking at the 2026 ASCRS meeting in Washington, DC, Dr Weikert reviewed findings from published studies assessing the performance of different AI-enhanced formulas. A paper reporting on the development of the Nallasamy formula, a ‘pure’ AI formula, included data from eyes with axial lengths (ALs) ranging from 20 mm to 32 mm and found that the Nallasamy and EVO formulas predicted refractive outcomes within ±0.5 D of target in 80% of eyes.¹ These performances were slightly better than comparators that included one hybrid AI formula and five other geometric optics formulas.

However, Dr Weikert cautioned that this is a good example of how ophthalmologists need to dig a little deeper when assessing formula performance, as the Nallasamy formula was only trained for a single IOL (SN60WF, Alcon)—so it is unknown whether its results can be generalised to other IOL models.

In addition, although these formulas performed relatively well across a broad range of ALs, their accuracy, as with other formulas, decreased in very long and very short eyes.

A study focusing on eyes with an AL greater than 30 mm also highlighted the limitations of AI-enhanced formulas in very long eyes.²

“Fortunately, these are not commonly encountered eyes,” Dr Weikert said. “And even though formulas are not highly accurate in this group, based on what we know about their performance, we can at least set appropriate patient expectations.”

In another study, the same investigators evaluated the same set of formulas and found they were more accurate in medium-long eyes (AL 24.5–26.0 mm).³ The predicted outcome for all formulas in this investigation was ±0.5 D of target in 82% to 87% of eyes.

The disappointing performance of newer formulas was seen in a study including eyes with an AL less than 22 mm.⁴ In this study, conducted by Dr Weikert and colleagues, the best-performing formula (Zeiss AI) predicted a refractive outcome ±0.5 D of target in only 73% of cases.

“Mean absolute error with the Zeiss AI formula was 0.4 D, which is pretty good but still not where we would like it to be,” Dr Weikert said.

He recommended using IOL power calculations with more than one formula (e.g., a geometrics optic and an AI-based), particularly in cases of ‘atypical’ eyes. Dr Weikert mentioned the ESCRS online calculator as a useful resource for this task, as it provides power calculations using seven modern formulas and requires users to input data only once.

Mitchell P Weikert MD, MS is Professor of Ophthalmology, Baylor College of Medicine, Houston, Texas, US. mweikert@bcm.edu

1. Li T, et al. Br J Ophthalmol, 2023; 107(8): 1066–1071.

2. Stopyra W, et al. Am J Ophthalmol, 2025; 271: 337–346.

3. Stopyra W, et al. Life (Basel), 2025; 15(1): 45.

4. Kenny PI, et al. J Cataract Refract Surg, 2023; 49(7): 697–703.