Purpose

To evaluate the predictive accuracy of ChatGPT-generated upper blepharoplasty simulations and investigate the effect of prompt specificity on quantitative periocular morphometric accuracy.

---

Methods

Thirty patients (60 eyes) who underwent isolated upper blepharoplasty were included. Standardized preoperative and 6-month postoperative photographs were analyzed. Preoperative photographs were processed using ChatGPT Pro to generate two simulation models: (1) general AI-generated prediction and (2) anatomically constrained prediction preserving margin reflex distance (MRD)1, MRD2, and palpebral fissure height (PFH). Quantitative eyelid measurements were performed using ImageJ software, including MRD1, MRD2, PFH, pretarsal show (PTS), and mid-pupillary lid crease distance (MPLCD), defined as the distance between the corneal light reflex and upper eyelid crease measured at the 15° to165° meridians.Quantitative results obtained from both AI-generated prediction models were compared with real postoperative photographs.

---

Results

Preoperative MRD1, MRD2, PFH, and PTS values were 2.71±0.7 mm, 5.28±0.6 mm, 8.02±0.8 mm, and 1.95±1.3 mm, compared with postoperative values of 2.74±0.7 mm, 5.31±0.9 mm, 8.03±1.2 mm, and 3.24±1.5 mm. Significant postoperative changes were observed only in pretarsal show and MPLCD measurements (p≤0.05).
Compared with real postoperative photographs, the general AI prediction group demonstrated significant differences in MRD1 (3.40±0.6 mm, p≤0.001), MRD2 (5.99±0.4 mm, p≤0.001), PFH (9.39±0.7 mm, p≤0.001), and all MPLCD measurements (p≤0.001), whereas PTS (2.97±1.2 mm, p=0.11) remained comparable. In contrast, the anatomically constrained prediction group showed quantitatively comparable results with real postoperative photographs (p≥0.05).

---

Conclusion

General AI-generated simulations tended to standardize periocular anatomy, whereas anatomically constrained prompts produced realistic postoperative predictions. Prompt engineering appears critical for accurate AI-based surgical visualization and avoidance of unrealistic patient expectations.

---

Additional Authors

First name	Last name	Base Hospital / Institution
NİSA	AKIN	Zonguldak Bülent Ecevit University
SERDAR	BİLİCİ	Zonguldak Bülent Ecevit University
SUAT HAYRİ	UĞURBAŞ	Zonguldak Bülent Ecevit University

↑ Back to top