Diagnostic decisions of specialist optometrists exposed to ambiguous deep-learning outputs

Abstract

Abstract Artificial intelligence (AI) has great potential in ophthalmology; however, there has been limited clinical integration. Our study investigated how ambiguous outputs from an AI diagnostic support system (AI-DSS) affected diagnostic responses from optometrists when assessing cases of suspected retinal disease. Thirty optometrists at Moorfields Eye Hospital (15 more experienced, 15 less) assessed 30 clinical cases in counterbalanced order. For ten cases, participants saw an optical coherence tomography (OCT) scan, basic clinical information and a retinal photograph (‘no AI’). For another ten, they were also given the AI-generated OCT-based probabilistic diagnosis (‘AI diagnosis’); and for ten, both AI-diagnosis and an AI-generated OCT segmentation (‘AI diagnosis + segmentation’) were provided. Cases were matched across the three types of presentation and were purposely selected to include 40% ambiguous and 20% incorrect AI outputs. Optometrist diagnostic agreement with the predefined reference standard was lowest for the ‘AI diagnosis + segmentation’ presentation (204/300, 68%) compared to both ‘AI diagnosis’ (224/300, 75% p = 0·010), and ‘no Al’ (242/300, 81%, p = < 0·001). Agreement in the ‘AI diagnosis’ presentation was lower (p = 0·049) than in the ‘no AI’. Agreement with AI diagnosis consistent with the reference standard decreased (174/210 vs 199/210, p = 0·003), but participants trusted the AI more (p = 0·029) when segmentations were displayed. There was no significant effect of practitioner experience on diagnostic responses (p = 0·24). More experienced participants were more confident (p = 0·012) and trusted the AI less (p = 0·038). Our findings also highlighted issues around reference standard definition.