Care to Explain? AI Explanation Types Differentially Impact Chest Radiograph Diagnostic Performance and Physician...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• The impact of artificial intelligence (AI) explanations on physician diagnostic performance in radiology remains unclear.
• Understanding how AI explanations influence decision-making is crucial for effective AI integration in healthcare.

Purpose of the Study:

• To evaluate how AI explanation types (local vs. global) and AI advice characteristics (correctness, confidence) affect physician diagnostic performance, perceived usefulness, and trust.
• To investigate these effects in the context of chest radiograph diagnosis.

Main Methods:

• A multicenter, prospective randomized study involving 220 physicians (radiologists and non-experts).
• Participants reviewed chest radiographs with simulated AI advice, with AI explanation type as a between-participant factor.
• Generalized linear mixed-effects models analyzed diagnostic accuracy, efficiency, and trust, controlling for expertise and AI knowledge.

Main Results:

• Local AI explanations significantly improved diagnostic accuracy when AI advice was correct and increased overall diagnostic efficiency.
• Physician task expertise interacted with explanation type and AI confidence, influencing diagnostic accuracy.
• Local explanations fostered greater "simple trust" (alignment with AI advice) compared to global explanations, irrespective of AI advice correctness.

Conclusions:

• The type of AI explanation critically impacts physician diagnostic performance and trust in AI systems.
• Local, feature-based AI explanations appear more beneficial than global, prototype-based explanations in radiological decision-making.
• These effects on performance and trust can occur even without conscious physician awareness.