Reporting radiographers’ interaction with Artificial Intelligence—How do different forms of AI feedback impact trust and decision switching?

• DOI: 10.1371/journal.pdig.0000560
• Publication Date: 2024-08-07T17:24:29Z
• Authors: Clare Rainey; Raymond Bond; Jonathan McConnell; Ciara Hughes; Devinder Kumar; Sonyia McFadden
• Keywords: Computer applications to medicine. Medical informatics; R858-859.7; Research Article

Artificial Intelligence (AI) has been increasingly integrated into healthcare settings, including the radiology department to aid radiographic image interpretation, reporting by radiographers. Trust cited as a barrier effective clinical implementation of AI. Appropriating trust will be important in future with AI ensure ethical use these systems for benefit patient, clinician and health services. Means explainable AI, such heatmaps have proposed increase transparency elucidating which parts ‘focussed on’ when making its decision. The aim this novel study was quantify impact different forms feedback on expert clinicians’ trust. Whilst conducted UK, it potential international application interface design, either globally or countries similar cultural and/or economic status UK. A convolutional neural network built study; trained, validated tested publicly available dataset MU sculoskeletal RA diographs (MURA), binary diagnoses Gradient Class Activation Maps (GradCAM) outputs. Reporting radiographers (n = 12) were recruited from all four regions Qualtrics used present each participant total 18 complete examinations MURA test (each examination contained more than one image). Participants presented images first, next finally an diagnosis sequential order. Perception obtained following presentation heatmap feedback. participants asked indicate whether they would change their mind (or decision switch) response disagreed abnormal 45.8% time agreed 86.7% (26/30 presentations).’Only two indicated that switch (GradCAM binary) (0.7%, n 2) across datasets. 22.2% 32) localisation pathology heatmap. level agreement GradCAM found correlated (GradCAM:—.515;—.584, significant large negative correlation at 0.01 (p < .01 and—.309;—.369, medium .01) respectively). This shows extent both is study, where greater form associated particular Forms should developed cognisance need precision accuracy promote appropriate end users.