Objective: To address the challenges of systematically identifying herb pairs in traditional Chinese medicine (TCM), we proposed HerbGL, a framework for predicting potential herb pairs that integrates network propagation and graph regularization. Methods: Based on the assumption that herbal actions induce subtle perturbations in biological systems, a framework named HerbGL was proposed. Random walk with restart (RWR) was first applied to the protein-protein interaction (PPI) network to reconstruct herb-specific perturbation effects and generate weighted subnetworks. Then, to quantify affinity between herb pairs, two network-proximity metrics, Closeness and PageRank, were computed from the weighted subnetworks to construct herb-pair affinity matrices. Finally, these matrices, together with known herb pairs (derived from co-occurrence analysis of TCM formulas with a threshold determined from the inflection point of the frequency distribution), were incorporated into a graph regularization model to predict potential herb pairs. Model performance was assessed through baseline comparison, ablation and robustness experiment under different ratios of positive and negative samples, using the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), accuracy, and precision as evaluation metrics. Furthermore, the predicted herb pairs were validated through both literature evidence and Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. Results: The weighted subnetworks constructed by RWR provided a refined simulation of herb-specific perturbation effects, which formed the basis for subsequent affinity modeling and prediction. Analysis of herb pair co-occurrence frequencies revealed a marked change around 150, which was selected as the threshold to distinguish herb pairs from non-herb pairs. HerbGL exhibited superior predictive performance compared with baseline models (AUROC = 0.970 5, AUPRC = 0.955 5, accuracy = 0.726 6, precision = 0.970 6). Ablation results showed that removing the Closeness and PageRank metrics substantially degraded performance (AUROC = 0.819 1, AUPRC = 0.876 8), confirming their necessity. Robustness evaluation under an imbalanced positive-to-negative sample ratio of 1 : 5 yielded AUROC = 0.969 6 and AUPRC = 0.840 4, indicating stable predictive ability. Moreover, multiple case studies further validated the rationality of the predicted herb pairs, such as Fangfeng (Saposhnikoviae Radix) and Qingpi (Citri Reticulatae Pericarpium Viride) which are recorded in Liangpeng Huiji (《良朋汇集》, Collection of Excellent Recipes) Vol. 3: Fangfeng Shengma Tang (防风升麻汤). Additionally, pathway enrichment analysis of the Renshen (Ginseng Radix et Rhizoma) and Lianqiao (Forsythiae Fructus) pair further supported the biological plausibility of their compatibility. Conclusion HerbGL offers an effective and biologically informed framework for identifying herb pairs in TCM. Beyond improving herb pair prediction, the framework also provides data support for research on herb compounds and mechanisms, thereby supporting data-driven exploration of TCM compatibility.

Objective: This study proposes a clustering framework for Chinese materia medica (CMM) based on a large language model (LLM), aiming to explore potential compatibility patterns among CMMs from the semantic perspective of CMM property theory. Methods: First, a CMM property knowledge base was constructed based on Chinese Materia Medica, including 567 commonly used CMMs characterized by four properties, five flavors, and meridian tropism. Then, 49 CMMs derived from 10 prescriptions for Zangdu (脏毒, pathogenic toxins) recorded in Waike Zhengzong (《外科正宗》, Orthodox Manual of External Medicine) and Yangke Xinde Ji (《疡科心得集》, Collected Insights on Ulcer Medicine) were selected as the experimental dataset. Five semantic representation methods—One-Hot, Word2Vec, Bidirectional Encoder Representations from Transformers (BERT), Beijing Academy of Artificial Intelligence General Embedding (BGE), and Qwen—were applied to encode CMM property information into vector representations. Subsequently, t-distributed Stochastic Neighbor Embedding (t-SNE) was used for nonlinear dimensionality reduction on high-dimensional semantic vectors, followed by k-means clustering (k = 7). Clustering performance was evaluated using the Silhouette Score (SS), Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI). Results: The Qwen-based clustering method, CMM-EmbedCluster, achieved the highest SS (0.607 4) and CHI (158.057 2), as well as the lowest DBI (0.499 5), indicating improved cluster separation and compactness compared with other methods. Visualization of CMM clustering results showed that the clusters were well separated in the low-dimensional space, with strong inter-cluster discrimination and high intra-cluster functional consistency. Further interpretability analysis of CMM clustering results revealed stable structural differences among clusters in terms of four properties, five flavors, and meridian tropism, forming functional partitions consistent with CMM property theory. Conclusion CMM-EmbedCluster utilizes an LLM to achieve semantic-level representation and clustering of CMMs within the framework of CMM property theory, providing support for exploring potential compatibility patterns among CMMs from the perspective of CMM property semantics.

Objective: To develop a facial image generation method based on a facial color-preserving generative adversarial network (FCP-GAN) that effectively decouples identity features from diagnostic facial complexion characteristics in traditional Chinese medicine (TCM) inspection, thereby addressing the critical challenge of privacy preservation in medical image analysis. Methods: A facial image dataset was constructed from participants at Nanjing University of Chinese Medicine between April 23 and June 10, 2023, using a TCM full-body inspection data acquisition equipment under controlled illumination. The proposed FCP-GAN model was designed to achieve the dual objectives of removing identity features and preserving colors through three key components: (i) a multi-space combination module that comprehensively extracts color attributes from red, green, blue (RGB), hue, saturation, value (HSV), and Lab spaces; (ii) a generator incorporating efficient channel attention (ECA) mechanism to enhance the representation of diagnostically critical color channels; and (iii) a dual-loss function that combines adversarial loss for de-identification with a dedicated color preservation loss. The model was trained and evaluated using a stratified 5-fold cross-validation strategy and evaluated against four baseline generative models: conditional GAN (CGAN), deep convolutional GAN (DCGAN), dual discriminator CGAN (DDCGAN), and medical GAN (MedGAN). Performance was assessed in terms of image quality [peak signal-to-noise ratio (PSNR) and structural similarity (SSIM)], distribution similarity [Fréchet inception distance (FID)], privacy protection (face recognition accuracy), and diagnostic consistency [mean squared error (MSE) and Pearson correlation coefficient (PCC)]. Results: The final analysis included facial images from 216 participants. Compared with baseline models, FCP-GAN achieved superior performance, with PSNR = 31.02 dB and SSIM = 0.908, representing an improvement of 1.21 dB and 0.034 in SSIM over the strongest baseline (MedGAN). The FID value (23.45) was also the lowest among all models, indicating superior distributional similarity to real images. The multi-space feature fusion and the ECA mechanism contributed significantly to these performance gains, as evidenced by ablation studies. The stratified 5-fold cross-validation confirmed the model’s robustness, with results reported as mean ± standard deviation (SD) across all folds. The model effectively protected privacy by reducing face recognition accuracy from 95.2% (original images) to 60.1% (generated images). Critically, it maintained high diagnostic fidelity, as evidenced by a low MSE (< 0.051) and a high PCC (> 0.98) for key TCM facial features between original and generated images. Conclusion The FCP-GAN model provides an effective technical solution for ensuring privacy in TCM diagnostic imaging, successfully having removed identity features while preserving clinically vital facial color features. This study offers significant value for developing intelligent and secure TCM telemedicine systems.

OBJECTIVE To address the issue of recommending traditional Chinese medicine(TCM)prescriptions,and to utilize the clinical records of lung cancer from TCM experts to automatically generate prescriptions,providing reference for the study of medi-cation rules and TCM clinical decision-making assistance.METHODS This algorithm transformed clinical manifestations,standard-ized tongue diagnosis,and pulse diagnosis into TCM prescriptions through a large model,thereby converting the task of TCM prescrip-tion recommendation into a text generation task.The CHATGLM3 model,based on the GLM structure,was used to enhance the under-standing of lung cancer cases and learn the intrinsic experiential knowledge of TCM experts in treating lung cancer,thereby improving the prescription generation effectiveness of the model.This was compared with traditional generative models.RESULTS The study demonstrated that integrating TCM knowledge from lung cancer cases into large language models effectively improved the model's pre-scription generation capabilities.Particularly in generating commonly used core medications by TCM experts,the model showed a high tendency and provided rich and valuable reference information.The lung cancer TCM prescription recommendation model achieved 64.62%in BLEU,55.78%in ROUGE,and 47.39%in METEOR scores.It also achieved accuracies of 67.79%,63.66%,56.76%,and 51.93%in the top 5,10,15,and 20 TCM prescriptions,respectively,outperforming the baseline model.CONCLUSION The lung cancer TCM prescription recommendation model presented in this paper achieves better prescription generation results com-pared to traditional generative models.It demonstrates the model's ability to learn knowledge about lung cancer diagnosis and treatment from cases,thereby generating TCM prescriptions that align with TCM treatment principles.This also provides a potential direction for future assistance in clinical decision-making.

Objective To address the challenges of constructing large language models for traditional Chinese medicine(TCM)classics,which are complex and expensive to fine-tune,this study explores a lightweight fine-tuning method for such models,aiming to develop a question-answering model centered on TCM classics,particularly various editions of Shang Han Lun through the ages.Methods Dataset construction involved designing prompts to guide GPT-4 in generating Q&A pairs based on Shang Han Lun and integrating them with the ShenNong_TCM_Dataset and cMedQA2 datasets.Five general-purpose large models were selected for Lora fine-tuning.The best model was chosen through evaluation,and the performance of multiple quantized versions was validated.Results After fine-tuning,the BLEU,ROUGE-1,ROUGE-2,and ROUGE-L metrics for the Qwen-7B-Chat model improved by 17.61,19.63,14.3,and 21.4,respectively,compared to the base model.Conclusion The selected model in this study is capable of effectively understanding and utilizing professional terms and concepts from TCM classics,such as Shang Han Lun,to provide accurate answers to user queries.Compared to similar models,it requires lower fine-tuning costs and computational power,contributing to the dissemination of TCM knowledge and the development of intelligent systems.

Objective To construct a traditional Chinese medicine(TCM)knowledge question-answering model with strong reasoning capabilities and reliable results,TCM Q&A datasets and TCM literature were fully utilized.Methods Large-scale TCM corpus and Q&A data were collected and organized,with ChatGLM3 serving as the base model.The PissA method was used for supervised fine-tuning,combined with retrieval-augmented generation(RAG)techniques,to build a TCM knowledge Q&A model that integrates supervised fine-tuning and retrieval-augmented generation.The model was compared with ChatGLM3,SFT,and RAG,with evaluations based on classic metrics such as BLEU,ROUGE1,and F-scores.Results The model in this paper achieved BLEU and ROUGE1 scores of 14.5830 and 34.6730,respectively.After incorporating retrieval-augmented generation,the model attained an F score of 0.6398 in the inference results on a TCM dataset,outperforming the ChatGLM3 baseline model's 0.2654.Conclusion The construction method of a large model in the TCM domain that integrates supervised fine-tuning and retrieval augmentation can effectively enhance the model's reasoning performance and reliability in TCM.

OBJECTIVE To provide help for further exploring the mechanism of action of traditional Chinese herb by constructing a complex network of traditional Chinese herb-gene-protein,optimizing the mining method of potential associated genes of traditional Chinese herb and improving the mining efficiency of traditional Chinese herb system biology information.METHODS A graph neural network model HERBGAT with an attention mechanism was proposed.A small amount of traditional Chinese herb-related gene data in the public data platform was used as input,and deep mining was performed in the traditional Chinese herb-gene-protein complex net-work to output potential traditional Chinese herb-related genes.The prediction results were analyzed by disease association analysis and KEGG signaling pathway analysis on the bioinformatics platform to clarify their mechanism of action,and the prediction results were verified by the literature retrieval platform.RESULTS The training results showed that the average prediction accuracy of the HERB-GAT model could reach 94%.Compared with the other two advanced complex network mining methods,HERBGAT showed better per-formance in the three indicators of ACC,AUC and AUPR.In the literature verification stage,the model prediction results were verified by TCM clinical literature and modern pharmacology literature,showing the good effect of HERBGAT in practical application.At the end of this paper,taking the HERBGAT model and the improved EMOGI model to explore the mechanism of action of Pinellia ternata in treating lung cancer as an example,199 potential associated genes of Pinellia ternata in treating lung cancer were found,and these potential associated genes were preliminarily analyzed and discussed with the help of bioinformatics methods.CONCLUSION The HERBGAT model can effectively mine potential traditional Chinese herb-associated genes,improve the mining efficiency of traditional Chinese herb-gene-protein complex networks,provide new ideas and references for the optimization of traditional Chinese herb system biology information mining methods,and provide data basis and experimental direction for exploring the mechanism of action of tradi-tional Chinese herb.

OBJECTIVE To establish an accurate and timely intelligent auxiliary diagnosis method for facial paralysis in order to enable patients and doctors to diagnose the disease faster and achieve the purpose of early detection,early diagnosis and early treatment.METHODS A method integrating SOLOv2-Vision Transformer was proposed.The collected facial paralysis data was seg-mented by the SOLOv2 model with a replaced backbone network,the interference part in the image was removed,and then inputted in-to the Vision Transformer model for classification training.By adopting the principle of segmentation first and then classification,the classification effect of facial paralysis images was improved.RESULTS The accuracy rate of the experimental method on the MEEI facial paralysis dataset was 0.982,the recall rate was 0.982,and the F1-score was 0.981,which were respectively increased by 2%,4%,and 4%compared with the basic model.CONCLUSION The facial paralysis classification model integrated with SOLOv2-Vi-sion Transformer can achieve higher recognition accuracy than the unsegmented method,and provides a new method for the diagnosis of facial paralysis.

OBJECTIVE To address the issue of recommending traditional Chinese medicine(TCM)prescriptions,and to utilize the clinical records of lung cancer from TCM experts to automatically generate prescriptions,providing reference for the study of medi-cation rules and TCM clinical decision-making assistance.METHODS This algorithm transformed clinical manifestations,standard-ized tongue diagnosis,and pulse diagnosis into TCM prescriptions through a large model,thereby converting the task of TCM prescrip-tion recommendation into a text generation task.The CHATGLM3 model,based on the GLM structure,was used to enhance the under-standing of lung cancer cases and learn the intrinsic experiential knowledge of TCM experts in treating lung cancer,thereby improving the prescription generation effectiveness of the model.This was compared with traditional generative models.RESULTS The study demonstrated that integrating TCM knowledge from lung cancer cases into large language models effectively improved the model's pre-scription generation capabilities.Particularly in generating commonly used core medications by TCM experts,the model showed a high tendency and provided rich and valuable reference information.The lung cancer TCM prescription recommendation model achieved 64.62%in BLEU,55.78%in ROUGE,and 47.39%in METEOR scores.It also achieved accuracies of 67.79%,63.66%,56.76%,and 51.93%in the top 5,10,15,and 20 TCM prescriptions,respectively,outperforming the baseline model.CONCLUSION The lung cancer TCM prescription recommendation model presented in this paper achieves better prescription generation results com-pared to traditional generative models.It demonstrates the model's ability to learn knowledge about lung cancer diagnosis and treatment from cases,thereby generating TCM prescriptions that align with TCM treatment principles.This also provides a potential direction for future assistance in clinical decision-making.

Objective To construct a traditional Chinese medicine(TCM)knowledge question-answering model with strong reasoning capabilities and reliable results,TCM Q&A datasets and TCM literature were fully utilized.Methods Large-scale TCM corpus and Q&A data were collected and organized,with ChatGLM3 serving as the base model.The PissA method was used for supervised fine-tuning,combined with retrieval-augmented generation(RAG)techniques,to build a TCM knowledge Q&A model that integrates supervised fine-tuning and retrieval-augmented generation.The model was compared with ChatGLM3,SFT,and RAG,with evaluations based on classic metrics such as BLEU,ROUGE1,and F-scores.Results The model in this paper achieved BLEU and ROUGE1 scores of 14.5830 and 34.6730,respectively.After incorporating retrieval-augmented generation,the model attained an F score of 0.6398 in the inference results on a TCM dataset,outperforming the ChatGLM3 baseline model's 0.2654.Conclusion The construction method of a large model in the TCM domain that integrates supervised fine-tuning and retrieval augmentation can effectively enhance the model's reasoning performance and reliability in TCM.