Large language models for patient-facing consultation in elective aesthetic and appearance-related plastic surgery: a systematic review and exploratory meta-analysis.
Abstract
Background
Large language models (LLMs) are increasingly used to provide patient-facing health information, yet evidence regarding their performance in elective aesthetic and appearance-related plastic surgery consultation remains heterogeneous. This review evaluated the scope, quality, usability, safety, and technical reliability of LLM-generated responses in this setting.
Methods
PubMed, Embase, Web of Science Core Collection, IEEE Xplore, and Scopus were searched from inception to September 30, 2026. Original studies that generated and evaluated LLM responses to patient-facing consultation, education, or perioperative information tasks in elective aesthetic or appearance-related plastic surgery were included. Outcomes were mapped to an 11-metric framework encompassing content quality, patient usability and communication, clinical safety, and technical reliability. Methodological quality was assessed using the revised JBI analytical cross-sectional appraisal tool with prespecified adaptations for LLM studies. Multilevel random-effects meta-analyses were conducted for Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease (FRE) when sufficiently comparable data were available.
Results
Thirty-eight studies were included; 20 compared multiple LLMs or AI systems, and three directly compared LLM responses with human experts. Of 418 study-metric combinations, 107 (25.6 %) received formal quantitative evaluation and 34 (8.1 %) formal non-quantitative evaluation. Medical correctness and completeness were assessed most frequently, whereas actionability and response generation were underexamined; no study formally evaluated model stability. Performance varied by task and evaluation dimension, with limitations in individualized recommendations, postoperative triage, and citation reliability. Twelve distinct studies contributed readability data, with 11 studies included in each outcome analysis. Exploratory pooling showed very high heterogeneity (FKGL: I2 = 89.6 %; FRE: I2 = 92.7 %). Pooled FKGL was 11.41 (95 % CI, 10.43-12.39; 95 % prediction interval, 7.45-15.36), and pooled FRE was 39.81 (95 % CI, 34.09-45.52; 95 % prediction interval, 19.15-60.47). The wide prediction intervals limit generalization of these averages to individual models or consultation settings.
Conclusions
Current evidence primarily supports the potential use of LLMs as clinician-supervised adjuncts for standardized patient education and communication in elective aesthetic and appearance-related plastic surgery.Although LLMs can provide generally useful information for common questions, Clinical effectiveness remains unestablished, and evidence is insufficient to support autonomous individualized surgical advice, postoperative triage, or management decisions. Future studies should prioritize patient-level validation of understanding, treatment expectations, and clinically meaningful safety outcomes.
