Causability and explainability of artificial intelligence in medicine

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Data Science

Background:

• Explainable AI (XAI) is crucial for medical applications, addressing the opacity of modern machine learning, especially deep learning (DL).
• Classic AI offered transparency but struggled with real-world uncertainties, leading to probabilistic models that are powerful yet opaque.
• The increasing success of AI in medicine necessitates methods for understanding its decision-making processes.

Purpose of the Study:

• To differentiate between explainability and causability in the context of AI in medicine.
• To introduce the concept of causability as a necessary step beyond explainability for achieving truly "explainable medicine."
• To present a use-case of DL interpretation and human explanation in histopathology.

Main Methods:

• Conceptual analysis to define and differentiate explainability and causability.
• Review of classic AI, probabilistic learning, and deep learning (DL) approaches to AI explainability.
• Case study application in histopathology involving DL interpretation and human explanation.

Main Results:

• Explainability is a property of the AI system, focusing on transparency and traceability.
• Causability is proposed as a property of the human user, measuring the quality of explanation and understanding.
• The study highlights the need for causability to achieve "explainable medicine," going beyond current explainable AI (XAI) methods.

Conclusions:

• Current explainable AI (XAI) methods focus on system transparency but may not guarantee genuine understanding.
• Causability, defined as the quality of human comprehension derived from explanations, is essential for advancing AI in medicine.
• Future research should focus on developing and measuring causability to enhance trust and utility of AI in clinical practice.