Ranking XAI Methods for Head and Neck Cancer Outcome Prediction

XAI methods vary significantly in quality: The study evaluated 13 different explainability techniques across 24 metrics and found "large variations across evaluation aspects among different XAI methods," indicating that not all methods for explaining AI predictions are equally reliable or useful in clinical settings.

Integrated Gradients and DeepLIFT performed best: Two specific XAI methods—Integrated Gradients (IG) and DeepLIFT (DL)—"consistently obtained high rankings for faithfulness, complexity and plausibility," suggesting these approaches may be more trustworthy for explaining cancer outcome predictions to clinicians.

Comprehensive evaluation is essential for clinical adoption: The authors emphasize "the importance of comprehensive XAI method evaluation," noting that previous HNC studies "empirically selected" explanation techniques without rigorous comparison. This suggests that responsible AI deployment in healthcare requires systematic validation rather than ad-hoc method selection.

Multiple evaluation dimensions matter: The study assessed XAI methods across four distinct quality dimensions (faithfulness, robustness, complexity, and plausibility), implying that a single metric is insufficient—responsible AI evaluation in medicine requires examining whether explanations are accurate, reliable, understandable, and clinically sensible.

Findings could improve AI adoption in cancer care: By identifying which explanation methods work best for head and neck cancer prediction, this research could help clinicians better understand and trust AI-assisted treatment recommendations, potentially accelerating responsible AI integration into personalized cancer treatment planning.