Evaluation of Aesthetic Outcomes Following Botulinum Toxin Treatment Using Multimodal Large Language Models: A Paired...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Aesthetics
• Computer Vision

Background:

• Multimodal large language models (MLLMs) are increasingly used for visual assessment.
• Their efficacy in evaluating aesthetic treatment outcomes is not well-established.

Purpose of the Study:

• To evaluate if current MLLMs can identify treatment status and regional aesthetic improvements after botulinum neurotoxin injections.
• To assess MLLM performance against established benchmarks.

Main Methods:

• Observational study analyzing 23 paired facial image cases (46 images).
• Four leading MLLMs (GPT-4.5 Pro, Grok 4.1, Gemini 3.1 Pro, Claude Opus 4.6) performed multiple inference runs.
• Models assessed post-treatment status and regional improvement (forehead, glabella, periorbital).
• Performance metrics included accuracy, sensitivity, specificity, and Fleiss' kappa.

Main Results:

• All MLLMs accurately identified post-treatment images (100% accuracy).
• Region-specific improvement detection often failed to surpass majority baselines (65.2%-91.3%).
• Gemini 3.1 Pro demonstrated superior forehead and glabella performance; no model met the periorbital baseline.
• Models systematically overestimated improvement, with false positives outweighing false negatives.

Conclusions:

• MLLMs can recognize aesthetic changes but lack clinical nuance.
• Domain-specific fine-tuning, expert benchmarking, and structured outcome frameworks are crucial for clinical deployment.
• Governance of training data and safeguards are essential before MLLM use in clinical or research settings.