AI-based Prediction of Neck Metastasis in Early-stage Lip Squamous Cell Carcinoma Using Preoperative Clinical–morphologic Features

The objective of this study is to evaluate the feasibility of artificial intelligence (AI)-based large language models (LLMs) for predicting cervical lymph node metastasis (pN+) in early-stage lip squamous cell carcinoma (SCC) using preoperative clinical photographs and basic clinical parameters. This retrospective, single-center study included 14 patients (12 males, 2 females; mean age 59.4) with histopathologically confirmed lip SCC who underwent primary tumor excision and selective neck dissection. Preoperative standardized JPEG photographs were analyzed using 2 LLMs-ChatGPT (GPT-5, OpenAI) and Gemini (Google)-through a zero-shot approach. Model predictions were compared with the pathologic findings. The mean depth of invasion (DOI) was 5.8 mm. ChatGPT predicted an average metastasis risk of 13.1% and Gemini of 22.4%, both of which overestimated the actual risk (0%). ChatGPT showed better calibration (Brier score 0.023) and weaker correlation with DOI (r=0.16) than Gemini (Brier 0.064, r=0.56). Both models tended to assign a higher risk to the lower-lip and commissural lesions. AI-based large language models can extract clinically relevant risk cues from preoperative photographs. While oversensitivity and reduced specificity remain limitations, this study demonstrates a novel and accessible approach for AI-assisted preoperative risk assessment of lip SCC.