Artificial Intelligence-Assisted Reconstruction Planning of Cheek Soft Tissue Defects: A Comparative Study Against Surgical Expert Consensus.

BACKGROUND

Reconstructive planning for cheek soft tissue defects requires expert assessment of defect size and location to preserve aesthetic harmony. Artificial intelligence (AI), particularly large language models (LLMs), has emerged as a potential decision-support tool in plastic and reconstructive surgery. This study evaluates the accuracy and reliability of ChatGPT-4o in proposing reconstructive approaches for post-excisional cheek defects.

METHODS

A comparative observational image-based study was performed using 46 standardized oblique facial photographs of healthy volunteers, on which cheek soft tissue defects were manually delineated to simulate post-excisional tissue loss. Each annotated photograph was submitted to ChatGPT-4o alongside a standardized prompt incorporating patient-specific clinical data. Two experts independently rated AI-generated responses across five domains using a five-point Likert scale (1-5). Inter-rater agreement was assessed using Cohen's kappa and intraclass correlation coefficient (ICC).

RESULTS

ChatGPT-4o achieved high ratings across all domains: clinical accuracy 4.88/5, safety 4.59/5, aesthetic result 3.96/5, first reconstruction selection 4.05/5, and relevance of alternatives 3.65/5, with an average of 4.22/5. Strong inter-rater agreement was observed between evaluators (ICC = 0.75, 95% CI 0.57-0.86, p < 0.001), with high consistency for safety (κ = 0.82) and first reconstruction selection (κ = 0.76). However, ChatGPT-4o failed to generate illustrative depictions of the proposed reconstructive approaches.

CONCLUSION

ChatGPT-4o demonstrated strong concordance with expert evaluators and shows promise as a decision-support adjunct in cheek reconstructive planning. However, its inability to generate visual depictions of proposed flaps remains a notable limitation. Expert supervision remains indispensable, and prospective clinical validation is warranted before routine implementation.

LEVEL OF EVIDENCE III: 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 .