The accuracy of intraocular lens power calculation formulas based on artificial intelligence in highly myopic eyes: a systematic review and network meta-analysis

Authors: Zhou Y, Dai M, Sun L, Tang X, Zhou L, Tang Z, Jiang J, Xia X.

Geographical coverage: China, Europe and Australia

Sector: Cataract treatment

Sub-sector: Intraocular lens power calculation

Equity focus: Not reported

Study population: Adults with high myopia

Review type: Effectiveness review

Quantitative synthesis method: Meta-analysis

Qualitative synthesis method: Not applicable

Background:

Myopia is a widespread refractive error, with high myopia (defined as axial length >26.0 mm or refractive error worse than −6.00 D) increasingly prevalent worldwide. It is associated with structural changes in the eye and is a risk factor for cataract development. In cataract surgery, accurate intraocular lens (IOL) power calculation is crucial, particularly in highly myopic eyes, which present additional challenges due to anatomical variations. Traditional vergence-based IOL formulas often result in refractive prediction errors in these eyes. Artificial intelligence (AI)-based formulas have recently been developed to improve prediction accuracy, but a comprehensive comparison of their performance is lacking.

Objective:

To systematically compare and rank the accuracy of AI-based intraocular lens (IOL) power calculation formulas versus traditional formulas in highly myopic eyes using network meta-analysis methods.

Main findings:

This systematic review and network meta-analysis included 12 retrospective studies involving a total of 2,430 highly myopic eyes (axial length >26.0 mm) that underwent uncomplicated cataract surgery with in-the-bag mono-focal IOL implantation. Twenty-one IOL power calculation formulas were compared, including both traditional and AI-driven methods.

The AI-based formulas—specifically XGBoost, Hill-RBF, and Kane—were consistently the most accurate across multiple metrics: lower mean absolute error (MAE), lower median absolute error (MedAE), and higher percentages of eyes achieving refractive prediction errors within ±0.25, ±0.50, and ±1.00 diopters. These significantly outperformed traditional formulas such as SRK/T, Haigis, Hoffer Q, and Holladay 1 and 2. Although Barrett Universal II and Olsen (newer-generation traditional formulas) also performed well, their advantages over AI methods were not statistically significant.

Subgroup analyses for extremely myopic eyes (≥30.0 mm axial length) reaffirmed the superior accuracy of AI-based formulas, particularly XGBoost and Hill-RBF v3.0. All included studies were considered high quality. However, geographical representation was limited, with 83% of studies from China.

Methodology: 

A search was conducted in PubMed, Web of Science, Embase, and the Cochrane Library from inception to 5 April 2023. Studies were included if they involved adults with high myopia undergoing uncomplicated cataract surgery and compared at least two IOL power formulas, including one AI-based formula. Exclusions included studies on non-AI methods only, patients under 18, coexisting ocular conditions, non-standard surgeries, or lack of relevant data.

Primary outcomes included MAE, MedAE, and prediction accuracy within ±0.25, ±0.50, and ±1.00 diopters. Network meta-analyses used random-effects models (R and STATA). Formula ranking was based on surface under the cumulative ranking curve (SUCRA). Subgroup analyses by axial length, sensitivity analyses, and publication bias assessment via funnel plots were performed.

Applicability/external validity:

External validity is limited due to the dominance of Chinese studies (10 of 12). Results may not generalise to populations in North America, Africa, South America, or other parts of Asia. Findings are based solely on adults with uncomplicated cataract surgery using mono-focal IOLs, limiting broader applicability.

Geographic focus:

Primarily China (10 of 12 studies), with limited data from Europe and Australia.

Summary of quality assessment:

There is low confidence in the review conclusions due to several issues: incomplete search strategy (no grey literature or language restrictions mentioned), no list of excluded studies, and no author contact for missing data. Although a valid risk of bias tool was used and results were clearly presented, limitations in comprehensiveness and transparency reduce confidence.