Comparing the accuracy of intraocular lens power calculation formulas using artificial intelligence and traditional formulas in highly myopic patients: a meta-analysis

Author: Hao Y, Fu J, Huang J, Chen D.

Geographical coverage: China, Japan, Portugal and Germany

Sector: Artificial intelligence

Sub-sector: Intraocular lens

Equity focus: Not reported

Study population: Patients with high myopia (axial length > 26 mm)

Review type: Effectiveness review

Quantitative synthesis method: Meta-analysis

Qualitative synthesis method: Not applicable

Background:

High myopia—defined as a refractive error of –6.00 dioptres or greater—is a major cause of vision loss, especially in East and South-East Asia. It is frequently complicated by early cataract, yet anatomical variations such as posterior staphyloma make intra-ocular lens (IOL) power estimation particularly difficult. Traditional IOL-power formulae have evolved through five ‘generations’, gradually incorporating additional biometric variables. More recently, artificial-intelligence (AI) methods (e.g. Kane, Hill-RBF) have been introduced to improve refractive predictability, but results have been inconsistent.

Objective:

To compare the accuracy of AI-based and conventional IOL-power formulae in highly myopic eyes undergoing cataract surgery.

Main findings:

Nine studies (1 291 eyes) published between 2019 and 2023 were included: six from China and one each from Japan, Portugal and Germany. Six formulae were examined—Kane, Hill-RBF, Barrett Universal II, EVO (AI-based), SRK/T and Haigis (traditional). Review findings include:

• Mean absolute error (MAE): Kane produced the lowest MAE and was significantly more accurate than SRK/T, Haigis and Barrett Universal II.
• Refractive predictability: Kane achieved the highest proportion of eyes within ± 0.25 D, ± 0.50 D and ± 1.00 D of the target refraction. Hill-RBF out-performed SRK/T and Haigis, but did not differ significantly from Kane or Barrett Universal II.

Overall authors concluded that AI-based formulae—particularly Kane—consistently out-performed traditional approaches in highly myopic eyes.

Methodology:

PubMed, Embase and the Cochrane Library were searched (to March 2023) for studies of cataract patients with axial length > 26 mm that compared at least two of the target formulae and reported MAE or refractive prediction error ≥ 1 month post-operatively. Reference lists were scanned manually. Two reviewers independently screened records, extracted data and assessed quality using a QUADAS-2-based checklist; disagreements were resolved by discussion or third-reviewer arbitration. Random-effects meta-analysis synthesised outcomes; heterogeneity was evaluated with the I² statistic and explored through sensitivity analyses.

Applicability / external validity:

Most studies originated from China, reflecting the regional burden of myopia, which may limit generalisability elsewhere. Variation in surgical technique, biometry devices and formula versions could influence accuracy, and follow-up was usually short term (≥ 1 month). Larger, multicentre studies across diverse populations are needed.

Geographic focus:

The included studies were conducted in China, Japan, Portugal and Germany; no geographical limits were imposed.

Summary of quality assessment:

There is medium confidence in the review’s conclusion. Searches were comprehensive, inclusion criteria explicit and dual-reviewer methods employed. Study characteristics and statistical analyses were adequately reported, and heterogeneity was assessed. However, the review lacked an excluded-studies list and did not stratify results by risk-of-bias level.