A hierarchical and interpretable machine learning model for acupoint determination.

Area of Science:

• Integrative Medicine
• Artificial Intelligence in Healthcare
• Traditional Chinese Medicine

Background:

• Acupuncture and moxibustion therapy (AMT) effectiveness can be enhanced with personalized treatment strategies.
• Developing data-driven models for acupoint prescription is crucial for advancing AMT.

Purpose of the Study:

• To develop a machine learning model for personalized acupoint prescriptions based on patient symptoms.
• To improve the efficiency and effectiveness of acupuncture and moxibustion therapy through intelligent systems.

Main Methods:

• A hierarchical attention-based recurrent neural network (HARNN) was employed for symptom-based acupoint prediction.
• Data preprocessing and augmentation were utilized to build a robust machine learning database.
• Local interpretable model-agnostic explanation (LIME) was applied for model interpretability and clinical validation.

Main Results:

• The HARNN model achieved high performance, with an intersection over union (IoU) of 0.954 after data augmentation.
• The model demonstrated strong predictive accuracy on both cross-validation and test datasets.
• LIME provided intuitive visualizations, enhancing the model's clinical reliability and understanding.

Conclusions:

• A hierarchical and interpretable machine learning model was successfully developed for predicting acupoint prescriptions.
• The integration of HARNN and LIME offers a novel technical approach for the intellectualization of AMT.
• This study provides a robust methodology for personalized and data-driven acupuncture treatments.