Evaluation of Aesthetic Outcomes Following Botulinum Toxin Treatment Using Multimodal Large Language Models: A Paired Before-and-After Analysis

Background Multimodal large language models (MLLMs) are increasingly applied to visual assessment tasks, yet their ability to evaluate aesthetic treatment outcomes remains unclear. Objectives To assess whether contemporary MLLMs can identify treatment state and detect region-specific aesthetic improvement following botulinum toxin (BoNT) treatment. Methods In this observational study, 23 paired facial image cases (46 images; 460 model evaluations) were analyzed. Four MLLMs (GPT-5.4 Pro, Grok 4.1, Gemini 3.1 Pro, Claude Opus 4.6) performed five independent inference runs per case. Models identified the post-treatment image and assessed regional improvement (forehead, glabella, periorbital). Accuracy, sensitivity, specificity, balanced accuracy, Matthews correlation coefficient (MCC), and Fleiss’ κ were calculated descriptively. Performance was compared to majority-class baselines. Exploratory outputs included aesthetic scores and apparent age estimates. Results All models identified the post-treatment image (100% accuracy). Region-specific improvement detection frequently failed to exceed majority-class baselines (65.2–91.3%). Gemini 3.1 Pro showed the highest performance for forehead (74.8%) and glabella (63.5%), while no model reached the periorbital baseline. Inter-run reliability varied widely (κ −0.113 to 0.719). High reliability did not imply correctness. All models systematically overestimated improvement (62.6–94.8% of predictions exceeded ground truth). False positives exceeded false negatives in all 12 model–task combinations. Exploratory outputs indicated perceived rejuvenation. Conclusions MLLMs recognize the format of aesthetic change but not its clinical nuance. Bridging this gap requires more than improved accuracy: a coordinated agenda of domain-specific fine-tuning, expert-rater benchmarking, and structured outcome frameworks, supported by governance of training-data provenance and clear safeguards before any clinical or research deployment.