A Machine Learning Approach for Recommending Herbal Formulae with Enhanced Interpretability and Applicability

Abstract

Herbal formulae (HFs) are representative interventions in Korean medicine (KM) for the prevention and treatment of various diseases. Here, we proposed a machine learning-based approach for HF recommendation with enhanced interpretability and applicability. A dataset consisting of clinical symptoms, Sasang constitution (SC) types, and prescribed HFs was derived from a multicenter study. Case studies published over 10 years were collected and curated by experts. Various classifiers, oversampling methods, and data imputation techniques were comprehensively considered. The local interpretable model-agnostic explanation (LIME) technique was applied to identify the clinical symptoms that led to the recommendation of specific HFs. We found that the cascaded deep forest (CDF) model with data imputation and oversampling yielded the best performance on the training set and holdout test set. Our model also achieved top-1 and top-3 accuracies of 0.35 and 0.89, respectively, on case study datasets in which clinical symptoms were only partially recorded. We performed an expert evaluation on the reliability of interpretation results using case studies and achieved a score close to normal. Taken together, our model will contribute to the modernization of KM and the identification of an HF selection process through the development of a practically useful HF recommendation model.

Advancing precision rheumatology: applications of machine learning for rheumatoid arthritis management

“…Owing to the ‘black box’ nature of deep learning models, the decision-making processes are opaque and difficult to comprehend, which may affect both physician and patient trust and understanding of model predictions ( 83 ). To address this limitation, some well-known methods can be utilized: The Class Activation Mapping (CAM) technique helps in understanding the regions of interest within images as attended by the model ( 84 ); Shapley Additive exPlanations (SHAP) elucidate the global impact of each feature on the model ( 85 ); and Local Interpretable Model-agnostic Explanations (LIME) explicate the local prediction process for individual samples ( 86 ). Collectively, these methods provide interpretability tools that enhance comprehension of the model’s decision-making process and improve its interpretability.…”