FMASH: Advancing Traditional Chinese Medicine Formula Recommendation with Efficient Fusion of Multiscale Associations of Symptoms and Herbs

TL;DR

FMASH tackles the challenge of recommending Traditional Chinese Medicine formulas by integrating multiscale symptom–herb associations with molecular-level herb features in a unified heterogeneous graph. It introduces HGRE to encode local and global macro-relations, MLFIE to fuse molecular-scale and macroscopic herb properties (with Latent Mol and VAE-based imputation for missing data), and two task-specific models FMASH_RS and FMASH_Seq guided by GELRAM and Transformer architectures. The framework achieves state-of-the-art performance on the TCM-PD dataset, with notable gains in $Precision@5$, $Recall@5$, $F1@5$, and Best Matched Precision for sequences, and demonstrates robust ablation-supported benefits of each component. Case studies illustrate practical advantages, showing FMASH_Seq’s ability to generate coherent, clinically plausible prescriptions while FMASH_RS can trisect across multiple prescriptions but risks incompatibilities, underscoring the value of the sequential generation approach for safe clinical deployment.

Abstract

Traditional Chinese medicine (TCM) exhibits remarkable therapeutic efficacy in disease treatment and healthcare through patienti-specific formulas. However, current AI-based TCM formula recommendation models and methods mainly focus on data-based textual associations between symptoms and herbs, and have not fully utilized their features and relations at different scales, especially at the molecular scale. To address these limitations, we propose the Fusion of Multiscale Associations of Symptoms and Herbs (FMASH), an novel framework that effectively combines molecular-scale features and macroscopic properties of herbs with clinical symptoms, and provides the refined representation of their multiscale associations, enhancing the effectiveness of TCM formula recommendation. This framework can integrate molecular-scale chemical features and macroscopic properties of herbs, and capture complex local and global relations in the heterogeneous graph of symptoms and herbs, providing the effective embedding representation of their multiscale features and associations in a unified semantic space. Based on the refined feature representation, the framework is not only compatible with both traditional unordered formula recommendation task and the ordered herb sequence generation task, but also improves model's performance in both tasks. Comprehensive evaluations demonstrate FMASH's superior performance on the TCM formula recommendation over the state-of-the-art (SOTA) baseline, achieving relative improvements of 9.45\% in Precision@5, 12.11% in Recall@5, and 11.01% in F1@5 compared to the SOTA model on benchmark datasets. This work facilitates the practical application of AI-based TCM formula recommendation system.

Figures (3)

• Figure 1: The framework of FMASH that mainly consists of two phases, where Phase 1 involves the integration and embedding of multiscale symptom-herb associations and features, and Phase 2 involves the training of FMASH$\_$RS and FMASH$\_$Seq, respectively. The HGRE provides unified graph embeddings for macroscopic hierachical relations of symptoms and herbs, the MLFIE helps to characterize the functions and features of herbs at the molecular scale. The GELRAM enhances the integration and utilization the global features of symptoms and herbs. Feature Refinement (FR) improves the multiscale features embeddings of symptoms and herbs.
• Figure 2: The Framework of Molecular-Level Feature Integration and Enhancement method, which enhances the refined representation of multiscale herb features through effectively integrating molecular-scale chemical characteristics and macroscopic properties of herbs.
• Figure 3: Ablation study: comparison of FMASH_RS performance with and without FR, demonstrating its importance.