Background The use of multimodal artificial intelligence (AI) in plastic surgery is steadily increasing. Whether a general-purpose multimodal AI tool can, from photographs alone, assess facial aging and facelift candidacy at a level comparable to board-certified specialist plastic surgeons remains unknown. Objectives To determine if ChatGPT-5 (OpenAI, San Francisco, CA, USA) can identify facial aging features, stratify severity, and judge facelift candidacy from photographs alone, compared with board-certified plastic surgeons. Methods Two-center observational pilot. Twenty-two volunteers (mean age 42.0 ± 16.8 years; median 34 years; range 24–80) provided standardized four-view facial composite photographs. Five board-certified plastic surgeons independently completed an eight-item questionnaire per case. ChatGPT-5 assessed the same images with identical wording. Assessments were image-only and blinded (no demographics/history). Surgeon consensus was defined by plurality. Primary outcomes were agreement and Cohen’s κ; for ordinal items, weighted κ, Spearman’s ρ, and mean absolute error (MAE) were reported. McNemar’s test assessed discordance for binary items. Results For facelift candidacy, agreement was 95.5% (21/22; Cohen’s κ = 0.91; McNemar P = 1.00). For binary aging features, agreement ranged from 81.8 to 90.9% (κ ≈ 0.61 to 0.81). For ordinal severity (lower face and midface), exact agreement was 77.3%, disagreements were adjacent only, weighted κ = 0.74 to 0.86, Spearman’s ρ = 0.84 (P < .001). Inter-surgeon agreement on ordinal items was moderate to fair. For the adjunct-procedure recommendation, Top-1 accuracy was 70.6% (12/17; κ = 0.58) and Top-2 agreement was 77.3% (17/22). Conclusions In a blinded, standardized-photograph setting, ChatGPT-5 matched surgeons on binary facelift candidacy assessment and closely tracked severity grading with small, one-level differences at most. These findings may support use as a decision-support tool (triage, patient education) while surgeons retain hands-on examination and personalized planning. Larger, multicenter studies with more diverse image datasets are warranted to confirm generalizability and define deployment standards. Level of Evidence IV This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors <a href="http://www.springer.com/00266">www.springer.com/00266</a>.

ChatGPT-5 Matches Surgeon-Level Assessment of Facelift Candidacy: A Pilot Proof-of-Concept Study

期刊论文
发表日期：

2026年2月11日

查看全文

Ayman T. Saeed,
R. William F. Breakey,
Daniel B. Saleh,
Kemal Tunç Tiryaki,
Bryan J. Mayou,
Tariq M. Saeed

摘要

Background

The use of multimodal artificial intelligence (AI) in plastic surgery is steadily increasing. Whether a general-purpose multimodal AI tool can, from photographs alone, assess facial aging and facelift candidacy at a level comparable to board-certified specialist plastic surgeons remains unknown.

Objectives

To determine if ChatGPT-5 (OpenAI, San Francisco, CA, USA) can identify facial aging features, stratify severity, and judge facelift candidacy from photographs alone, compared with board-certified plastic surgeons.

Methods

Two-center observational pilot. Twenty-two volunteers (mean age 42.0 ± 16.8 years; median 34 years; range 24–80) provided standardized four-view facial composite photographs. Five board-certified plastic surgeons independently completed an eight-item questionnaire per case. ChatGPT-5 assessed the same images with identical wording. Assessments were image-only and blinded (no demographics/history). Surgeon consensus was defined by plurality. Primary outcomes were agreement and Cohen’s κ; for ordinal items, weighted κ, Spearman’s ρ, and mean absolute error (MAE) were reported. McNemar’s test assessed discordance for binary items.

Results

For facelift candidacy, agreement was 95.5% (21/22; Cohen’s κ = 0.91; McNemar P = 1.00). For binary aging features, agreement ranged from 81.8 to 90.9% (κ ≈ 0.61 to 0.81). For ordinal severity (lower face and midface), exact agreement was 77.3%, disagreements were adjacent only, weighted κ = 0.74 to 0.86, Spearman’s ρ = 0.84 (P < .001). Inter-surgeon agreement on ordinal items was moderate to fair. For the adjunct-procedure recommendation, Top-1 accuracy was 70.6% (12/17; κ = 0.58) and Top-2 agreement was 77.3% (17/22).

Conclusions

In a blinded, standardized-photograph setting, ChatGPT-5 matched surgeons on binary facelift candidacy assessment and closely tracked severity grading with small, one-level differences at most. These findings may support use as a decision-support tool (triage, patient education) while surgeons retain hands-on examination and personalized planning. Larger, multicenter studies with more diverse image datasets are warranted to confirm generalizability and define deployment standards.

Level of Evidence IV

This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266.