Iol-Agent: An Ai-Driven Multi-Agent System For Intelligent Iol Formula Selection — Conceptual Framework And Preliminary Retrospective Validation

Intraocular lens (IOL) power calculation is a critical determinant of refractive outcomes after cataract surgery. No single formula consistently demonstrates superior accuracy across all biometric profiles. We developed IOL-Agent, a multi-agent AI system built on large language model (LLM) architecture that autonomously selects the optimal IOL formula per eye. Specialized agents handle biometric classification, literature-based formula ranking, and final arbitration, each equipped with domain-specific tools and evidence-based decision logic derived from peer-reviewed literature. This study presents the system architecture and preliminary retrospective proof-of-concept results.

Augenklinik Ahaus - Fachkrankenhaus, Ahaus, North Rhine-Westphalia, Germany.

Consecutive eyes undergoing uncomplicated phacoemulsification with monofocal IOL implantation were included. Postoperative refraction was recorded at 4 weeks. Seven formulae (Barrett Universal II, PEARL-DGS, Kane, Hill-RBF 3.0, Cooke K6, EVO 2.0, Hoffer QST) were computed via the ESCRS IOL Calculator. IOL-Agent was implemented using a multi-agent LLM architecture with specialized agents for biometric classification, formula ranking, and arbitration. Only non-personal de-identified biometric parameters were processed. The orchestrating agent stratified eyes by axial length, keratometry, and anterior segment depth. Outcomes included SD of PE, MAE, MedAE, and percentage within ±0.25, ±0.50, and ±1.00 D.

A total of 146 eyes of 94 patients (55.3% male; mean age 67.76 ± 6.86 years) were analyzed. The AI-Agent selection achieved a SD of PE of ±0.387 D, MAE of 0.29 D, and MedAE of 0.24 D, with 54.1% of eyes within ±0.25 D, 84.9% within ±0.50 D, and 97.9% within ±1.00 D. Among individual formulae, the best-performing were Barrett Universal II (79.5% within ±0.50 D, SD ±0.426 D), Kane (78.8% within ±0.50 D, SD ±0.423 D), and Hill-RBF 3.0 (78.1% within ±0.50 D, SD ±0.456 D). The AI-Agent improved the percentage within ±0.50 D by 5.4 percentage points over Barrett, by 6.1 points over Kane and by 6.8 points over Hill-RBF, while achieving the lowest MAE (0.29 D) and MedAE (0.24 D) across all evaluated methods.

In this retrospective proof-of-concept study of 146 eyes, an LLM-based multi-agent AI system for individualized formula selection demonstrated superior refractive predictability compared with any single IOL power calculation formula. These findings support the feasibility and potential clinical utility of autonomous AI-driven formula selection in cataract surgery. Prospective multi-center validation with a larger cohort, predefined selection algorithms is warranted.