eXplainable artificial intelligence-Eval: A framework for comparative evaluation of explanation methods in...

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning Explainability
• Clinical Decision Support Systems

Background:

• Machine learning (ML) models are vital in healthcare, but their adoption requires explainability for trust and compliance.
• Evaluating the effectiveness of ML explanation methods in real-world clinical settings is crucial.

Purpose of the Study:

• To quantitatively compare local and global explainability methods in healthcare.
• To assess the suitability of various explanation techniques for clinical applications.

Main Methods:

• Introduced a structured methodology for evaluating explainability techniques.
• Assessed five local methods (LIME, contextual importance, RuleFit, RuleMatrix, Anchor) and four global methods (LIME, Anchor, RuleFit, RuleMatrix).
• Conducted experiments on diverse healthcare datasets and tasks using multiple explainability criteria.

Main Results:

• RuleFit and RuleMatrix demonstrated consistent, robust global explanations across healthcare tasks.
• Local explanation methods exhibited variable performance based on dataset and evaluation criteria.
• Identified trade-offs between fidelity, stability, and complexity in explanation methods.

Conclusions:

• Developed a practical framework for assessing ML explanation methods in healthcare.
• Provided guidance for selecting trustworthy and transparent ML techniques for clinical deployment.
• Emphasized the importance of systematic evaluation for safe ML implementation in sensitive environments.