CED: A case-level explainable paramedical diagnosis via AdaGBDT.

Area of Science:

• Machine Learning
• Medical Informatics
• Data Science

Background:

• The increasing volume of medical data necessitates advanced machine learning for diagnostics.
• Current machine learning models often lack crucial local or case-specific explainability.
• This limitation hinders the adoption of AI in critical medical decision-making.

Purpose of the Study:

• To introduce an explainable Gradient Boosting Decision Tree (GBDT)-based model for case-specific medical diagnosis.
• To enhance the interpretability of machine learning models in healthcare.
• To address the challenge of local explainability in medical diagnostics.

Main Methods:

• Proposed an adaptive Gradient Boosting Decision Tree (AdaGBDT) for efficient path mining.
• Utilized bi-side mutual information for case-specific feature importance embedding.
• Integrated AdaGBDT with case-based reasoning (CBR) for collaborative decision-making and identification of difficult cases.

Main Results:

• The AdaGBDT model achieved superior performance on the Wisconsin diagnostic breast cancer and UCI heart disease datasets.
• Achieved F1-scores of 0.9647 and 0.8405 on the respective datasets.
• Experimental analyses confirmed the effectiveness of the feature importance embedding for interpretability.

Conclusions:

• The proposed AdaGBDT model offers high predictive accuracy for case-specific medical diagnosis.
• The model provides both case-level and global explainability, enhancing trust and utility.
• This approach advances the application of explainable AI in paramedical diagnostics.