CFHR-TCM: a co-occurrence feature-enhanced deep learning framework for personalized herb recommendation in traditional Chinese medicine

Background Traditional Chinese medicine (TCM) is widely recognized for its personalized treatment strategies and holistic diagnostic philosophy. Driven by advancements in computer technology, herbal recommendation systems have emerged as a key direction in the modernization of TCM. Nevertheless, existing methods are often constrained by the sparsity of data regarding interactions between symptoms and herbal medicines, insufficient hierarchical modeling of diagnostic patterns, and limited semantic alignment between symptoms and prescribed herbs, which restrict their recommendation accuracy. Methods To overcome these challenges, we propose a co-occurrence feature-enhanced deep learning framework for personalized herb recommendation in TCM (CFHR-TCM). This framework employs pointwise mutual information (PMI) to capture global co-occurrence patterns and augment highly sparse clinical input data; it then maps these enhanced features into a probabilistic latent space via a variational autoencoder (VAE) to computationally approximate the underlying patterns of TCM treatment principles. Finally, by integrating a cross-attention mechanism with contrastive learning, the framework dynamically matches the inferred latent treatment principles with a patient’s specific symptoms while simultaneously avoiding semantic drift. Results Experiments conducted on two real-world TCM prescription datasets demonstrate that CFHR-TCM outperforms all baseline methods. Specifically, the framework achieved improvements of 9.43%, 8.20%, and 10.75% across the F1-Score@5, F1-Score@10, and F1-Score@20 metrics, respectively. Conclusion The experimental results demonstrate that CFHR-TCM exhibits promising performance in personalized herbal recommendation, which can assist clinical TCM tasks by facilitating more reliable diagnoses and herbal treatments.