Large language models (LLMs) show promise in specialized domains, but their capabilities in Traditional Chinese Medicine(TCM)-a field with complex theoretical foundations and clinical practices-remain inadequately assessed. This study aims to develop and apply a comprehensive, multidimensional benchmark to systematically evaluate LLM proficiency in this culturally-grounded medical discipline. We constructed the TCM-3CEval benchmark to assess models across three core dimensions: mastery of fundamental knowledge, comprehension of classical texts, and clinical decision-making. The dataset comprises 450 expert-validated single-choice questions. We evaluated a diverse set of models, including general-purpose international models, Chinese general models, and medical domain-specific models. Model performance was measured by accuracy under a rigorous permutation-based consistency test to control for positional bias. We show that while top-tier models approach human-level performance in aggregate scores, distinct deficits remain in clinical reliability. Models trained with Chinese linguistic and cultural priors demonstrate superior capability in interpreting classical texts compared to international counterparts. However, despite achieving comparable overall accuracy to human students (p > 0.05), models exhibit pronounced deficiencies in critical subdomains such as TCM Diagnostics and Syndrome Differentiation, where human reasoning remains significantly more robust. The benchmark effectively discriminates between rote memorization and the holistic inference required for clinical practice. This study establishes a standardized, triaxial evaluation paradigm for assessing AI in Traditional Chinese Medicine. The findings underscore the importance of cultural-contextual alignment and domain-specific adaptation for developing capable medical LLMs. This benchmark provides a foundation for optimizing models in culturally-grounded medical domains and supports the responsible integration of AI into traditional medical education and practice.

The emergence of Large Language Models (LLMs) within the Traditional Chinese Medicine (TCM) domain presents an urgent need to assess their clinical application capabilities. However, such evaluations are challenged by the individualized, holistic, and diverse nature of TCM’s “Syndrome Differentiation and Treatment” (SDT). Existing benchmarks are confined to knowledge-based question-answering or the accuracy of syndrome differentiation, often neglecting assessment of treatment decision-making. Here, we propose a comprehensive benchmark derived from clinical cases, classical case records, and authoritative examination questions. Designated TCM-BEST4SDT, the dataset comprises 600 questions covering four tasks: TCM Basic Knowledge, Medical Ethics, LLM Content Safety, and SDT. Data annotation adheres to a rigorous three-stage pipeline consisting of expert annotation, mutual cross-validation, and independent third-party review. The evaluation framework integrates three mechanisms, namely selected-response evaluation, judge model evaluation, and a specialized reward model employed to quantify prescription-syndrome congruence. Crucially, a controlled evaluation on the Qwen3 series demonstrated a strictly monotonic correlation between performance and model scale, verifying the benchmark’s sensitivity in discriminating model capabilities. This study establishes a standardized evaluation framework for intelligent TCM research, objectifies the personalized diagnostic logic of complementary medicine, and facilitates the optimization and application of large language models in digital medicine.

Objective: To address the critical gap that existing medical large language model evaluation systems are predominantly based on western medical paradigms and lack specialized assessment standards for the field of traditional Chinese medicine (TCM), this study constructs a comprehensive evaluation framework specifically designed for TCM clinical large language models, providing scientific assessment tools for their development, validation, and clinical application. Methods: This framework addresses the critical issue of missing TCM evaluation standards by incorporating TCM-specific elements such as pattern differentiation and treatment determination, as well as formula and herb recommendations, while emphasizing patient safety and data privacy. Results: The framework encompasses 3 core dimensions: (1) clinical scenario coverage and authenticity, including renowned physician case mining, pattern identification assistance, and therapy recommendations; (2) core diagnostic and therapeutic capability support, including symptom terminology recognition, pattern identification reasoning, and medical record generation; and (3) clinical application maturity and safety assurance, including data diversity, output reliability, and privacy protection. Conclusions: This evaluation system provides essential tools for the development and clinical deployment of TCM large language models, facilitating the responsible integration of AI technology into traditional medicine practice while preserving the theoretical integrity of TCM.

Background: Autoimmune diseases such as primary Sjögren’s syndrome and type 1 diabetes are frequently complicated by hepatic injury, yet therapies that simultaneously target inflammation and parenchymal damage remain limited. Mitochondrial dysfunction with excessive reactive oxygen species (ROS) production drives a self-amplifying pathogenic loop by activating the cGAS-STING innate immune pathway. We previously observed that a Chinese herbal formula preserved mitochondrial ultrastructure in autoimmune NOD mice, and computational screening identified trifolirhizin—a natural pterocarpan flavonoid—as the candidate active constituent mediating this protection. Here, we investigated the hepatoprotective effects and underlying mechanisms of trifolirhizin in autoimmune-associated liver injury. Methods: Female NOD mice received trifolirhizin (5, 10, or 20 mg/kg/day) for four weeks, with C57BL/6J mice as healthy controls. Hepatic histopathology, inflammatory cytokines, mitochondrial ultrastructure (TEM), mitochondrial membrane potential (ΔΨm), and ROS levels were evaluated. Integrated transcriptomic and metabolomic profiling was performed to unbiasedly characterize protective mechanisms. In vitro, H₂O₂-induced oxidative stress was established in HepG2 cells. Cells were treated with trifolirhizin (15–25 µM) and assessed for antioxidant enzyme activities, ΔΨm, ROS production, glycolytic and mitochondrial respiration (Seahorse analysis), and cGAS-STING pathway protein expression. Pharmacological rescue experiments using the cGAS agonist cGAMP were conducted to test pathway dependency. Results: Trifolirhizin dose-dependently alleviated hepatic pathological damage and reduced pro-inflammatory cytokine levels in NOD mice. Multi-omics profiling revealed that oxidative stress responses, the mitochondrial electron transport chain, and glutathione metabolism were the most significantly restored pathways. Trifolirhizin preserved mitochondrial ultrastructure, restored ΔΨm, and attenuated ROS accumulation both in vivo and in vitro. Functionally, Seahorse analysis demonstrated that trifolirhizin rescued overall cellular bioenergetics, restoring both glycolytic capacity and mitochondrial respiratory parameters (basal respiration, ATP production, maximal respiration, and spare respiratory capacity). Mechanistically, trifolirhizin suppressed the cGAS-STING-TBK1-IRF3 axis, as evidenced by reduced expression of cGAS, p-STING, ZBP1, p-TBK1, and p-IRF3. Importantly, the cGAS agonist cGAMP abrogated the protective effects of trifolirhizin, confirming that the cGAS-STING pathway is functionally required for its action downstream of mitochondrial protection. Conclusion: Trifolirhizin attenuates liver injury in the nod mouse by preserving mitochondrial integrity, maintaining cellular energy metabolism, and thereby suppressing the ROS/cGAS-STING inflammatory cascade. These findings position trifolirhizin as a promising mitochondria-targeted therapeutic candidate for pSS-related hepatic complications and provide a mechanistic framework for discovering active compounds from mitochondrially active herbal formulations.

Objective: Tongue image segmentation is fundamental for automated tongue diagnosis in Traditional Chinese Medicine (TCM). Existing deep learning approaches suffer from heavy reliance on manual annotation and limited adaptability to annotation‐scarce settings, resulting in suboptimal segmentation performance and poor generalization across diverse clinical environments. This study proposes a novel minimally supervised framework. In this approach, the coarse segmentation stage employs lightweight supervised training with limited annotations, while the fine segmentation stage achieves superior accuracy through fully unsupervised, zero‐shot inference leveraging SAM2. Methods: We developed the Centroid‐SAM2 model, a minimally supervised tongue image segmentation framework. The first stage employs a coarse segmentation model fine‐tuned on 120 annotated images from an independent public dataset, while the second stage leverages the zero‐shot capabilities of SAM2 with no labeled data for training or inference; annotations on the target datasets are used solely for evaluation. The approach integrates DeepLabV3_ResNet50 for initial coarse segmentation, followed by a novel centroid coordinate algorithm based on image moments to automatically extract precise guidance points. These centroid coordinates serve as adaptive prompts for the SAM2 model, enabling effective attention guidance in annotation‐free segmentation tasks. Results: Comprehensive experiments across three distinct tongue image datasets demonstrated that Centroid‐SAM2 significantly outperformed traditional supervised methods and other minimally supervised approaches. The model achieved exceptional performance with 98.21% mIoU, 98.92% accuracy, 99.73% recall, and 99.32% F1‐score on Dataset 1, with consistently high performance maintained across Dataset 2 and Dataset 3. Comparative analysis of different SAM2 model scales revealed that sam2.1‐hiera_small provided optimal performance with minimal edge artifacts while maintaining computational efficiency. Conclusion: The Centroid‐SAM2 model represents a significant advancement in minimally supervised medical image segmentation, substantially reducing dependency on manual effort by confining labeling requirements to a small independent dataset, while achieving superior accuracy in tongue segmentation tasks. This innovative approach addresses the critical challenge of annotation scarcity in medical imaging, offering significant practical value for both TCM modernization and the broader deployment of automated tongue diagnosis systems in complex clinical environments.

The emergence of Large Language Models (LLMs) within the Traditional Chinese Medicine (TCM) domain presents an urgent need to assess their clinical application capabilities. However, such evaluations are challenged by the individualized, holistic, and diverse nature of TCM's "Syndrome Differentiation and Treatment" (SDT). Existing benchmarks are confined to knowledge-based question-answering or the accuracy of syndrome differentiation, often neglecting assessment of treatment decision-making. Here, we propose a comprehensive, clinical case-based benchmark spearheaded by TCM experts, and a specialized reward model employed to quantify prescription-syndrome congruence. Data annotation follows a rigorous pipeline. This benchmark, designated TCM-BEST4SDT, encompasses four tasks, including TCM Basic Knowledge, Medical Ethics, LLM Content Safety, and SDT. The evaluation framework integrates three mechanisms, namely selected-response evaluation, judge model evaluation, and reward model evaluation. The effectiveness of TCM-BEST4SDT was corroborated through experiments on 15 mainstream LLMs, spanning both general and TCM domains. To foster the development of intelligent TCM research, TCM-BEST4SDT is now publicly available.