Artificial intelligence-based decision support in simulated free flap Re-exploration for head and neck...

Area of Science:

• Microsurgery
• Artificial Intelligence
• Head and Neck Reconstruction

Background:

• Free flap compromise is a critical complication in head and neck reconstruction.
• Prompt recognition and management are essential for successful outcomes.
• The role of large language models (LLMs) in supporting microsurgical decision-making for flap compromise is not well-defined.

Purpose of the Study:

• To evaluate the performance of three LLMs (ChatGPT, Gemini, Copilot) in simulated head and neck free flap compromise scenarios.
• To compare LLM accuracy, management recommendations, and potential for harm against expert microsurgeon ratings.
• To assess LLM utility in both intraoperative and postoperative flap compromise situations.

Main Methods:

• A simulated case study using 30 standardized head and neck free flap compromise scenarios (15 intraoperative, 15 postoperative).
• Three LLMs (ChatGPT, Gemini, Copilot) were prompted for diagnosis, mechanism, and management.
• Four consultant microsurgeons rated LLM responses on clinical accuracy, management, explanation, confidence, and potential for harm.

Main Results:

• All LLMs demonstrated high clinical accuracy in recognizing flap compromise.
• Gemini outperformed ChatGPT and Copilot in management recommendations and explanation clarity for postoperative cases.
• Potentially harmful recommendations were infrequent but varied between models, particularly in postoperative scenarios.

Conclusions:

• LLMs show significant alignment with expert recognition of flap compromise.
• Differences in management advice and explanatory clarity, especially in postoperative settings, warrant careful consideration.
• Current LLMs should be used cautiously and do not currently support independent clinical use in microsurgical care.