Optimal presentation of decision explanations: investigating the effects of explainability and reliability on medical diagnostic performance with conversational AI.

INTRODUCTION

The effective and efficient presentation of decision-making rationales within artificial intelligence black boxes remains a critical factor limiting the application of conversational AI in medical diagnosis. This study investigated the effects of decision explanation formats and system reliability on physicians' diagnostic behavior.

METHODS

A simulated diagnostic experiment involving 22 clinicians identifying complex pulmonary conditions was conducted. Four explanation formats and two levels of system reliability were manipulated. Diagnostic accuracy, task completion time, subjective cognitive load, trust ratings, and user preferences were measured.

RESULTS

Explanation format did not significantly affect diagnostic accuracy but significantly affected task completion time, subjective cognitive load, and trust ratings. A significant interaction between explanation format and system reliability was observed for task completion time. Decision tree explanations yielded the lowest cognitive load, whereas flowcharts resulted in the longest task completion times across reliability conditions. Highlighted text received the highest trust ratings and subjective preference scores. System reliability and explanation format played complementary roles in human-AI collaborative diagnosis, with the efficiency of different explanation formats varying across reliability conditions.

DISCUSSION: These findings highlight the joint roles of decision explanation format and system reliability in conversational AI and provide empirical evidence for the design and optimization of medical AI interfaces.