Abstract
Purpose :
One possible barrier to AI deployment in healthcare is over-reliance from users. This, known as automation bias (AB), has not been assessed in ophthalmology but may impact clinical decisions. We tested whether AB was present when optometrists used AI diagnostic support for retinal disease using an online experiment.
Methods :
Thirty hospital optometrists (15 more experienced and 15 less experienced) assessed 30 cases. Ten consisted of an OCT scan, basic clinical information, and a fundus image (manual). Ten also displayed AI diagnoses suggestions (AI). Ten additionally displayed an AI-produced OCT segmentation map (segmentation). Participants chose the most probable diagnosis per case and gave their diagnostic confidence. Level of trust in the AI outputs was also reported. Cases were chosen to be matched across conditions and to give 70% accuracy on the AI diagnoses.
Results :
Compared to the gold standard clinical diagnoses, 670/900 (74%) responses were correct. There were significantly fewer correct responses for segmentation (204/300, p<0.001) and AI (224/300, p=0.049) than manual (242/300) and for segmentation compared to AI (p=0.010). Agreement with correct AI diagnosis decreased when segmentations were displayed (174/210 vs 199/210, p<0.001)(Table 1), with examples suggesting that this may be due to participants paying more attention to segmentation maps over AI diagnoses. There was no significant effect of experience on the number of correct responses across the three conditions (p=0.24). More experienced participants were more confident in their diagnoses (p=0.012) and trusted the AI less (p=0.038). Participants trusted the AI more when segmentation was displayed (p=0.029) but AI did not affect diagnostic confidence (p=0.461).
Conclusions :
Using an imperfect AI system has a significant negative effect on correct diagnoses irrespective of experience. Displaying segmentation maps may increase the likelihood of acceptance if used in practice due to an increased level of trust. If used together in practice, the synchronization of algorithms for segmentation maps and diagnostic suggestions must be improved.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.