2-Factor Retrieval for Improved Human-AI Decision Making in Radiology

2FR improves clinician accuracy by showing similar past cases: The study found that a new explainability method called "2-factor retrieval" (2FR)—which shows doctors previously diagnosed cases similar to the current patient—led to increases in clinician accuracy on chest X-ray diagnoses, particularly benefiting radiologists with low confidence in their initial assessment.

Current AI explanations may not help doctors verify results: Existing explainability techniques like saliency maps and Shapley values don't allow physicians to independently verify whether the AI's reasoning is sound, highlighting a gap between technical explainability and clinical usability.

The "2-factor security blanket" requires AI to work correctly twice: 2FR creates two verification points—the AI must retrieve genuinely similar cases, and the clinician must recognize those cases as relevant to the current diagnosis—reducing the risk that doctors blindly trust incorrect AI suggestions.

How AI explains itself matters as much as accuracy: The research emphasizes that the interface design and presentation method of AI assistance significantly impacts whether clinicians actually benefit from it, suggesting that responsible AI deployment requires careful attention to how explanations are communicated, not just whether they're technically sound.

Low-confidence decisions benefit most from AI assistance: The findings suggest that AI decision support is most valuable when clinicians are uncertain, rather than as a universal second opinion, which has implications for how and when to deploy these systems in clinical practice.