Ada-WHIPS: explaining AdaBoost classification with applications in the health sciences

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Computer-Aided Diagnostics (CAD) systems aid medical practitioners but often use "black box" models, hindering trust and interpretability.
• Explainable Artificial Intelligence (XAI) is crucial for addressing these interpretability and trust concerns in clinical practice.

Purpose of the Study:

• To develop a novel algorithm for explaining AdaBoost classification models commonly used in CAD.
• To improve the interpretability of CAD systems, fostering greater trust among medical practitioners.

Main Methods:

• Introduction of Adaptive-Weighted High Importance Path Snippets (Ada-WHIPS), a novel algorithm specifically for explaining AdaBoost models.
• Ada-WHIPS utilizes AdaBoost's adaptive classifier weights, redistributing them among decision nodes to identify dominant rules.
• Evaluation of explanations using precision, coverage, and a novel stability measure in an experimental study.

Main Results:

• Ada-WHIPS explanations demonstrated superior generalization (15%-68% mean coverage) compared to state-of-the-art methods across 9 CAD datasets.
• The algorithm maintained competitive specificity (80%-99% mean precision), with a minor trade-off to prevent overfitting.
• Experimental results confirm the effectiveness of Ada-WHIPS for explaining AdaBoost classifiers in CAD.

Conclusions:

• The novel Ada-WHIPS algorithm offers a significant advancement in explaining CAD AdaBoost classifiers.
• This AdaBoost-specific approach outperforms model-agnostic methods, providing a valuable XAI solution for practitioners.
• Ada-WHIPS enhances trust and utility of CAD systems by making their reasoning transparent.