Explaining Black-Box Models for Biomedical Text Classification.

Area of Science:

• Biomedical Informatics
• Machine Learning Explainability
• Natural Language Processing

Background:

• Black-box machine learning models are increasingly used in biomedical text classification.
• Post-hoc explanation methods are crucial for understanding model decisions in critical applications.
• Existing methods like perturbation-based and decision sets have limitations in accuracy and interpretability.

Purpose of the Study:

• To propose a novel post-hoc explanation method, Biomedical Confident Itemsets Explanation (BioCIE), for black-box biomedical text classification models.
• To enhance the fidelity, interpretability, and coverage of explanations.
• To demonstrate BioCIE's superiority over existing explanation techniques.

Main Methods:

• BioCIE utilizes domain knowledge and confident itemset mining to discretize the model's decision space.
• It extracts semantic relationships between input text and class labels within these subspaces.
• The method approximates black-box behavior for individual predictions using extracted itemsets.

Main Results:

• BioCIE significantly improved explanation fidelity (11.6% instance-wise, 7.5% class-wise) and interpretability (8%).
• Evaluations on diverse biomedical text classification tasks showed BioCIE outperforms perturbation-based and decision set methods.
• The method produces concise, accurate, and interpretable class-wise explanations representing decision boundaries.

Conclusions:

• BioCIE effectively explains how black-box models semantically link input texts to class labels in biomedical contexts.
• The proposed method offers a robust approach to enhance trust and understanding of AI in healthcare.
• BioCIE provides a valuable tool for researchers and practitioners in biomedical natural language processing.