An Ensemble Learning Based Framework for Traditional Chinese Medicine Data Analysis with ICD-10 Labels

Area of Science:

• Computational intelligence in medicine
• Traditional Chinese Medicine (TCM) informatics
• Ensemble learning for healthcare

Background:

• Clinical experience of veteran doctors in Traditional Chinese Medicine (TCM) holds valuable diagnostic and treatment knowledge.
• Analyzing complex, historical TCM clinical records presents a significant challenge.
• Existing methods may not fully capture the nuanced experience embedded in TCM data.

Purpose of the Study:

• To develop an advanced computational model for analyzing TCM veteran doctors' clinical experience.
• To leverage ensemble learning for effective diagnosis and acupoint recommendation using ICD-10 labeled clinical records.
• To establish a framework that models implied knowledge from historical TCM data.

Main Methods:

• An ensemble learning framework was designed, integrating base learners like decision trees (DT) and support vector machines (SVM).
• Base learners were trained using bootstrap aggregation of the training dataset.
• A deep ensemble strategy combined learners, optimized for accuracy and diversity via the nondominated sort (NDS) algorithm.

Main Results:

• The proposed ensemble model achieved 88.2% ± 2.8% accuracy (zero-one loss) and 79.6% ± 3.6% accuracy (Hamming loss) in clinical diagnosis and acupoint recommendation tasks.
• Performance was evaluated against two established methods on a manually labeled ICD-10 dataset.
• The ensemble approach significantly outperformed existing methods in both ICD-10 label annotation and acupoint recommendation.

Conclusions:

• The developed ensemble model effectively captures and models the implicit knowledge and experience within historical TCM clinical data.
• The computational cost associated with training the ensemble of base learners is demonstrably low.
• This approach offers a promising avenue for enhancing TCM clinical decision support systems.