Prescription Recommendation Based on Multiview Learning with Latent Dirichlet Allocation

Tongue diagnosis, as a key process in Traditional Chinese Medicine (TCM), enables clinicians to assess bodily conditions by visually inspecting the tongue, based on which appropriate prescriptions are formulated with corresponding TCM theory. However, conventional tongue diagnosis relies heavily on the subjective expertise of clinicians, which is labor-intensive, prompting the development of deep learning-based approaches to automate the process. To further explore the application of deep learning in prescription recommendation from tongue images, we introduce a Prescription Recommendation method based on Latent Dirichlet Allocation (PR-LDA). The proposed method begins by extracting tongue image features using a multi-path convolutional neural network enhanced with dense connections, which effectively captures rich and multi-scale visual information. These features are then processed through multiple fully-connected layers to predict relevant medications and their dosages, leveraging similarity measures between the input tongue features and those associated with existing prescriptions. Additionally, Latent Dirichlet Allocation (LDA) is incorporated to uncover latent topic structures within prescriptions, thereby improving accuracy by integrating image features with textual prescription data. Experimental evaluations on our tongue-prescription dataset demonstrate that PR-LDA achieves Precision, Recall, and F1-score values of 0.3212, 0.2447, and 0.2777, respectively.