A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI

Analysis

Why This Paper Matters

This survey addresses a critical bottleneck in AI adoption: the lack of interpretability in deep learning models, especially in high-stakes fields like medicine. As AI systems achieve remarkable performance in tasks such as image diagnosis and treatment planning, their black-box nature undermines trust among clinicians and regulators. The paper systematically organizes the fragmented landscape of explainable AI (XAI) methods, offering a taxonomy that helps practitioners choose appropriate techniques. By focusing on medical XAI, it highlights unique requirements such as causal reasoning, domain-specific validation, and regulatory compliance, which are often overlooked in general XAI research.

The timing of this survey (2020) is significant, as it captures the state of XAI before the explosion of large language models and foundation models, yet its categories remain relevant. It serves as a foundational reference for subsequent work in medical AI interpretability, evidenced by over 2200 citations.

Technical Contributions

• Categorization of interpretability methods: The authors propose a multi-dimensional taxonomy, distinguishing between inherently interpretable models (e.g., linear models, decision trees) and post-hoc explanations (e.g., LIME, SHAP, saliency maps). They further break down post-hoc methods into global vs. local, model-specific vs. model-agnostic, and textual vs. visual explanations.
• Medical XAI focus: The paper maps general XAI categories to medical use cases, such as explaining radiology image classifications or electronic health record predictions. It discusses challenges like data heterogeneity, small sample sizes, and the need for causal explanations in clinical decision support.
• Educational call to action: The authors advocate for integrating interpretability into medical curricula, emphasizing that future clinicians must understand AI's mathematical and technical foundations to critically evaluate model outputs.

Results

As a survey, the paper does not present experimental results. Instead, it synthesizes findings from dozens of prior works, noting that most XAI methods are evaluated on benchmark datasets (e.g., MNIST, ImageNet) rather than real medical data. The authors observe that medical XAI studies often lack rigorous validation, with few comparing explanation quality against clinician judgments. They report that inherently interpretable models (e.g., logistic regression) remain common in medical literature, while deep learning explanations are still nascent.

Significance

This survey has become a standard reference for researchers entering XAI, particularly those working on medical applications. Its taxonomy provides a common vocabulary that facilitates cross-disciplinary collaboration between AI researchers and clinicians. By highlighting the gap between technical XAI advances and practical medical needs, it has spurred follow-up work on domain-specific explanation methods, such as concept-based explanations and counterfactual reasoning. The paper's emphasis on education also aligns with broader trends in AI ethics and responsible AI, making it a cornerstone for discussions on trustworthy AI in healthcare.