ESCRS - POREF101 - Ai–Assisted Versus Expert Surgeon Refractive Planning In Smooth Incision Lenticule Keratomileusis (Silk): Comparative Analysis And Visual Outcomes

Ai–Assisted Versus Expert Surgeon Refractive Planning In Smooth Incision Lenticule Keratomileusis (Silk): Comparative Analysis And Visual Outcomes

Published 2026 - 44th Congress of the ESCRS

Reference: POREF101 | Type: Free paper | DOI: 10.82333/hyzr-tg43

Authors: Dr Pier Luigi Surico* 1 , A/Prof. David HK Ma 2 , A/Prof. Paolo Lanzetta 3 , A/Prof. Chi Chin Sun 4

1Ophthalmology,University of Udine,Udine,Italy;Ophthalmology,Chang Gung Memorial Hospital,Linkou,Taiwan;Ophthalmology,Chang Gung Memorial Hospital,Keelung,Taiwan;Sense Organs,Sapienza University,Rome,Italy, 2Ophthalmology,Chang Gung Memorial Hospital,Linkou,Taiwan, 3Ophthalmology,University of Udine,Udine,Italy, 4Ophthalmology,Chang Gung Memorial Hospital,Keelung,Taiwan

Purpose

To evaluate the performance of a large language model (GPT-5) in refractive surgery planning by comparing AI-generated refractive targets with expert surgeon planning in Smooth Incision Lenticule Keratomileusis (SILK), and to report visual and anatomical outcomes achieved with surgeon-led treatment.

Setting

Retrospective study conducted at a tertiary ophthalmology center in Taiwan.

Methods

Retrospective analysis of patients undergoing SILK with the ELITA femtosecond laser was conducted. GPT-5 generated refractive plans using preoperative clinical and tomographic data, which were compared with surgeon plans for sphere, cylinder, and axis. Postoperative outcomes at 1 and 3 months included UDVA, CDVA, keratometry, corneal thickness, and astigmatic axis concordance (Δ Axis).

Results

A total of 134 eyes (67 patients; mean age 32.4 ± 6.7 years) were analyzed. GPT-5 planned significantly less myopic spherical correction than surgeons (−4.29 ± 2.42 D vs −4.70 ± 1.98 D; p = 0.0025), while cylindrical power was comparable (−0.91 ± 0.69 D vs −0.92 ± 0.80 D; p = 0.92). Mean absolute astigmatic axis deviation (Δ Axis) was 61.9° ± 50.1°, with only 18% of eyes within 15°. Surgeon-planned SILK yielded excellent outcomes: UDVA improved from 0.74 ± 0.36 to −0.06 ± 0.07 logMAR at 3 months (p < 0.0001), non-inferior to preoperative CDVA. Mean keratometry flattened from 43.05 ± 1.16 D to 39.30 ± 1.35 D, and corneal thickness stabilized after early postoperative reduction.

Conclusions

GPT-5 demonstrated partial agreement with expert surgeons in refractive magnitude but showed clinically relevant limitations in spherical precision and astigmatic axis alignment. Surgeon-planned SILK resulted in excellent visual acuity, refractive predictability, and early structural stability, reinforcing the safety and efficacy of keratorefractive lenticule extraction in real-world clinical practice. While LLMs may offer future decision-support potential, dedicated training and validation are required before clinical implementation in refractive surgery.