A manifesto on explainability for artificial intelligence in medicine.

Area of Science:

• Biomedical informatics
• Artificial intelligence
• Computer science

Background:

• Growing use of artificial intelligence (AI) in various applications.
• Concerns regarding the "black box" nature of AI and its lack of transparency.
• Heightened importance of AI explainability in biomedical fields due to patient safety implications.

Purpose of the Study:

• To explore the concept of explainable AI (XAI) in depth.
• To provide a functional definition and conceptual framework for XAI.
• To identify desiderata for achieving AI explainability in biomedicine.

Main Methods:

• A position paper synthesizing perspectives from seven researchers with diverse roles in AI and biomedicine.
• Conceptual analysis of explainable AI (XAI).
• Identification of key domains within biomedicine relevant to AI explainability.

Main Results:

• A functional definition and conceptual framework for XAI are proposed.
• A series of desiderata for achieving explainability in AI within biomedical contexts are presented.
• The paper highlights the critical need for understandable AI outputs in healthcare.

Conclusions:

• Explainable AI (XAI) is essential for safe and effective AI implementation in biomedicine.
• The proposed framework and desiderata offer guidance for developing and deploying explainable AI systems.
• Addressing AI explainability is vital for user trust and adoption in medical applications.