Appropriate trust in artificial intelligence for the optical diagnosis of colorectal polyps: The role of human/artificial intelligence interaction

Abstract

Background and AimsComputer-aided diagnosis (CADx) for optical diagnosis of colorectal polyps is thoroughly investigated. However, studies on human-artificial intelligence (AI) interaction are lacking. Aim was to investigate endoscopists’ trust in CADx by evaluating whether communicating a calibrated algorithm confidence improved trust. MethodsEndoscopists optically diagnosed 60 colorectal polyps. Initially, endoscopists diagnosed the polyps without CADx assistance (initial diagnosis). Immediately afterwards, the same polyp was again shown with CADx prediction; either only a prediction (benign or pre-malignant) or a prediction accompanied by a calibrated confidence score (0-100). A confidence score of 0 indicated a benign prediction, 100 a (pre-)malignant prediction. In half of the polyps CADx was mandatory, for the other half CADx was optional. After reviewing the CADx prediction, endoscopists made a final diagnosis. Histopathology was used as gold standard. Endoscopists’ trust in CADx was measured as CADx prediction utilization; the willingness to follow CADx predictions when the endoscopists initially disagreed with the CADx prediction. ResultsTwenty-three endoscopists participated. Presenting CADx predictions increased the endoscopists’ diagnostic accuracy (69.3% initial vs 76.6% final diagnosis, p<0.001). The CADx prediction was utilized in 36.5% (n=183/501) disagreements. Adding a confidence score led to a lower CADx prediction utilization, except when the confidence score surpassed 60. A mandatory CADx decreased CADx prediction utilization compared to an optional CADx. Appropriate trust, utilizing correct or disregarding incorrect CADx predictions was 48.7% (n=244/501). ConclusionsAppropriate trust was common and CADx prediction utilization was highest for the optional CADx without confidence scores. These results express the importance of a better understanding of human-AI interaction.