A deep neural network model for optimizing traditional Chinese medicine prescriptions with data augmentation

BACKGROUND AND PURPOSE: Traditional Chinese Medicine (TCM) has become a prominent and challenging area within the field of clinical decision support. However, most existing approaches fail to adequately address two critical issues: the inherent noise in TCM prescription data and the potential risks associated with excessive prescription modifications. EXPERIMENTAL APPROACH: To address these challenges, we propose a deep learning framework named DA-TCMPO (Data Augmentation for TCM Prescription Optimization), which aims to optimize TCM prescriptions using tailored data augmentation techniques. The model incorporates two data augmentation modules: a Diffusion Model Based on Double Attention (DAD), which enhances sample diversity, and a Variable Noise Embedding Module (VNE), which focuses on denoising and refining the augmented data. KEY RESULTS: To address the limitations of existing datasets, we constructed the Chinese Herbal Prescriptions for Diseases (CH) dataset, specifically designed for training models in TCM prescription optimization. Comprehensive experimental results on the CH dataset demonstrate that the proposed DA-TCMPO achieves performance significantly superior to the best-performing baseline, with precision, accuracy, recall, and F1-score representing relative improvements of 67.8%, 83.3%, 83.3%, and 83.6%. Furthermore, in vivo validation using Ulcerative colitis (UC) mouse models confirmed that DA-TCMPO-optimized prescriptions, such as CYKKL-2, produced statistically significant improvements in body weight, Disease activity index (DAI) score, and colon length compared with the model group, demonstrating the practical efficacy of the optimized prescriptions. CONCLUSION AND IMPLICATIONS: These findings indicate that DA-TCMPO holds promise for supporting diagnostic and therapeutic decision-making in practical TCM clinical scenarios.