TITLE:
Evaluation of Vision-Language Models for Traditional Chinese Medicine Tongue Diagnosis: A Comparative Study of AI Performance against Human Expert Assessment
AUTHORS:
Ryan Kern, Brian Z. Ring, Huijun Z. Ring
KEYWORDS:
Traditional Chinese Medicine, Tongue Diagnosis, Vision-Language Models, Artificial Intelligence, Medical Image Analysis, Gemini, Qwen-VL, Llava
JOURNAL NAME:
Health,
Vol.18 No.11,
November
17,
2026
ABSTRACT: Background: Traditional Chinese Medicine (TCM) tongue diagnosis is a complex diagnostic practice requiring years of training. Recent advances in vision-language models (VLMs) offer potential for automated assistance, though their capability for nuanced TCM assessment remains unexplored. This preliminary study investigates whether modern VLMs can perform TCM tongue feature assessment with agreement approaching that observed between human evaluators. Methods: A small dataset of standardized tongue images, contributed voluntarily with informed consent, was independently scored by two human evaluators and three VLMs using a consistent, structured prompt. Nine key diagnostic features were evaluated on a 0 - 100 scale across four tongue regions (base, center, sides, tip): tongue body color, shape, swelling, coating color, coating thickness, coating distribution, cracks or fissures, spots, and tooth marks. Agreement was measured using mean absolute deviation (MAD), root mean squared error (RMSE), bias, and categorical agreement of binned scores. Results: Inter-rater agreement between human evaluators established a benchmark (MAD: 14.4, agreement of binned scores: 66.5%). All three VLMs demonstrated measurable capability for TCM tongue assessment, with varying performance levels. The best-performing model achieved agreement approaching the human inter-rater benchmark (MAD: 18.0, agreement: 62.8%), while other models showed MAD values ranging from 18.8 to 37.9. A single standardized prompt was used across all models without model-specific optimization, which likely limited individual model performance. Performance patterns remained consistent when restricted to features with positive Lin CCC between human evaluators, suggesting similar performance across features of differing reliability. Conclusions: This small preliminary study demonstrates that vision-language models possess fundamental capabilities for Traditional Chinese Medicine tongue diagnosis, with agreement approaching that observed between human evaluators. The observed differences between models should be interpreted cautiously, as model-specific prompt optimization was not performed and could substantially alter relative performance. These findings establish proof-of-concept that VLMs can perform complex, nuanced diagnostic assessments requiring cultural medical knowledge and subtle visual discrimination, suggesting promising applications for clinical decision support, education, and diagnostic standardization in TCM practice.