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

Area of Science:

• Integrative Medicine
• Computational Biology
• Pharmacology

Background:

• Herbal formulae (HFs) are central to Korean medicine (KM) for disease prevention and treatment.
• Current HF selection relies on expert knowledge, presenting challenges in standardization and scalability.
• A data-driven approach is needed to enhance the interpretability and applicability of HF recommendations.

Purpose of the Study:

• To develop and validate a machine learning-based approach for recommending herbal formulae (HFs) in Korean medicine (KM).
• To enhance the interpretability and practical applicability of the HF recommendation system.
• To leverage clinical data and expert knowledge for improved traditional medicine interventions.

Main Methods:

• A multicenter dataset of clinical symptoms, Sasang constitution (SC) types, and prescribed HFs was utilized.
• Machine learning classifiers, oversampling techniques, and data imputation were evaluated.
• The Local Interpretable Model-agnostic Explanation (LIME) technique was employed for model interpretability.
• A cascaded deep forest (CDF) model was selected for its superior performance.

Main Results:

• The CDF model, combined with data imputation and oversampling, achieved the best performance on training and test datasets.
• The model demonstrated high top-1 (0.35) and top-3 (0.89) accuracies, even with partially recorded clinical data.
• Expert evaluation confirmed the reliability of the model's interpretation of clinical symptom-HF relationships.

Conclusions:

• The developed machine learning model offers a practical and interpretable solution for HF recommendation in Korean medicine.
• This approach contributes to the modernization of KM by standardizing and enhancing the HF selection process.
• The model has the potential to improve clinical decision-making and patient outcomes in traditional medicine settings.