Automation-bias in CXR reporting - does AI influence reader decision
Research type
Research Study
Full title
Automation bias in CXR reporting, how does AI influence reader decision-making and confidence– an experimental study
IRAS ID
353910
Contact name
Jenna Tugwell-Allsup
Contact email
Sponsor organisation
BCUHB
Duration of Study in the UK
0 years, 7 months, 30 days
Research summary
AI has been shown to aid clinician's performance in reporting chest x-rays to reduce discrepancies and aid early detection of lung nodules. Nevertheless, evidence provided within studies often lack the use of reporting radiographers as observers, fail to use clinically relevant real-world images (of varying quality), and focus merely on diagnostic performance and accuracy. The impact of AI on patient outcomes such as earlier diagnosis has been the primary focus with its effects on diagnostic thinking efficacy, hence impact on radiology workforce and workflow, often overlooked. How AI affects diagnostic decision making for radiologists and reporting radiographer when interpreting CXRs needs further consideration. A phenomenon known as ‘automation bias’ has been recognised with the introduction of AI-systems, it refers to the tendency of observers to rely excessively on AI software, resulting in underutilisation of their own judgment or critical-thinking skills. Limited studies have been conducted to help understand this phenomenon better especially surrounding the effects of incorrect AI output on human decision-making
This study will explore the impact of commercially available AI assisted CXR interpretation software on radiologists and reporting radiographers’ decision-making and confidence levels for nodule detection
This will be a multi-reader multiple case (MRMC) experimental study using a case-by-case sequential design. The experiment involves multiple readers from various specialties (radiologists and reporting radiographers) of varying experience interpreting a pre-selected set of CXRs with and without AI. This paired case-by-case design allows comparison of the performance of AI-unassisted and AI-assisted interpretation, where each observer serves as their own control, to ensure perfect comparability This is because all participants will be required to interpret the same set of CXRs, with and without AI output. Reader recruitment and reading sessions will be carried out across Health Boards in Wales.REC name
N/A
REC reference
N/A