Comprehensive Summary
This study conducted by Wysocki et al. explored the attitudes of physicians to different explainability methods provided by a sample future AI algorithm that identifies abnormalities in optical coherence tomography (OCT) scans. A questionnaire with 5 sample cases of patients with differing levels of accuracy in the AI model’s analysis was given to a total of 27 participants, which included groups of ophthalmic specialists, ophthalmic practitioners, and non-ophthalmic specialists. For each case, participants were asked if they thought the algorithm’s analysis of the patient’s condition was correct, their confidence in the findings, and whether or not the generated heatmap on the OCT image increased their trust in AI models. Responses were gathered in a Likert scale format ranging from strongly disagree (-3) to strongly agree (3), which were then categorized by participant group and presented in a graph. In cases 1 and 2 where the algorithm correctly analyzed the OCT images, there was a high level of trust in the results. However, in cases 3-5 where the algorithm was incorrect or generated a faulty explanation, all groups indicated low levels of confidence, with ophthalmic specialists having the lowest level of trust.
Outcomes and Implications
This study reflects the current perception of clinicians from both ophthalmic and non-ophthalmic practices of AI explanation models. Current AI models may be difficult for clinicians to trust because they cannot easily access the algorithm’s rationale for predictions. Wysocki et al. collected realistic data as to how helpful and accurate various healthcare professionals thought explanations provided by the AI algorithm were. Varying opinions on how useful explanations are across the 3 participant groups suggests that different explainability models may be effective across different disciplines, and overall low understanding of AI models among participants suggest that AI education and preparation for healthcare applications should be more widely implemented.