Background and aim:
To elicit clinician preferences for AI use in skin cancer diagnosis, a pilot discrete-choice experiment (DCE) survey was conducted. The pilot aimed to (i) provide parameter estimates to inform a Bayesian statistically efficient main survey design, and (ii) check face validity with respondents.
Study design and methods
Choice sets consisted of 2 alternatives: skin cancer diagnosis with and without AI assistance. Attributes for clinicians included the test performance (sensitivity and specificity); supporting evidence for the AI program; time savings; and the cost to the clinical practice for AI diagnostic technology, with three levels for each attribute. The DCE design created using Ngene software consisted of six blocks, 12 choice sets each, with participants randomly assigned to one block. Participants were recruited via the ACEMID platform and the Australasian College of Dermatology newsletter. Choice data were analysed using a multinomial logit (MNL) model with Nlogit software.
Results
A total of 24 clinicians participated in the pilot: 33% were aged 45-54 years, 71% were females. Clinicians showed no clear overall preference for AI-assisted diagnosis compared with no AI-assistance (β =0.03; p =0.93). False positives (β = -0.09, p <0.001) and false negatives (β = -0.27, p <0.001), cost to clinical practice (β = -0.28 × 10⁻⁴, p <0.001) and having to spend longer with patients for AI-assisted diagnosis (β = -0.58, p = 0.04) were negatively associated with preferences.
Conclusion
While there was no overall preference for skin cancer diagnosis with versus without AI-assistance, preferences were strongly associated with test accuracy, additional costs to the practice, and time required to confirm a diagnosis. These attributes are valid for use in the larger definitive study.