Retrospective Assessment of LLM Decision-Making for IOL Power Selection in Cataract Surgery in Comparison with Surgeon Performance.

Objectives

To evaluate a general-purpose large language model (LLM)'s ability to perform surgeon-level decision-making in choosing an intraocular lens (IOL) power when provided the same preoperative data.

Methods

This study included 293 eyes (174 patients) that underwent monofocal or toric cataract surgery performed by a single surgeon. GPT-5.2 analyzed the same biometric printouts, topographic displays, and preoperative data used in surgical planning. Postoperative refraction for GPT-5.2-selected IOLs was simulated from the measured postoperative refraction using a fixed conversion factor. The mean refractive deviation from target, mean absolute error (MAE), and number of eyes within ±0.25 D and ±0.50 D of target refraction were calculated. A subgroup analysis evaluated toric IOL cylinder magnitude and axis.

Results

GPT-5.2-selected IOL power differed significantly from that of the surgeon (19.24 ± 4.01 vs. 19.64 ± 4.07 D [mean ± SD], p < 0.001). Using simulated refractive outcomes for GPT-5.2, MAE was higher than for surgeon-selected IOLs (0.45 ± 0.03 vs. 0.35 ± 0.02 D [mean ± SE], p = 0.002), and a greater percentage of eyes fell within ±0.25 D of target refraction for the surgeon (51.9% vs. 44.7%, p < 0.01), with no difference within ±0.50 D (p = 0.06). In toric eyes, the cylinder magnitude did not differ, but simulated residual astigmatism was higher for GPT-5.2 (0.78 ± 0.55 vs. 0.58 ± 0.47 D, p < 0.001), with an axis flip in 47.1% versus 39.2% of eyes.

Conclusions

GPT-5.2 synthesized clinical data and approximated surgeon IOL selection, but its simulated refractive outcomes were less accurate than the surgeon's measured outcomes. These findings suggest that a domain-specific LLM may serve as a support tool in IOL selection and cataract surgery planning.