Withdrawn from program - did not register Australasian Melanoma Conference 2026

Patient preferences for Artificial Intelligence-assisted skin cancer diagnosis: A pilot discrete choice experiment (#333)

Anh D Tran 1 , Monika Janda 2 , Peter Soyer 3 , Rachael L Morton 1 , Alison Button-Sloan 4
  1. NHMRC Clinical Trials Centre, University of Sydney, Sydney, NSW, Australia
  2. Centre for Health Services Research, The University of Queensland, Brisbane, Queensland, Australia
  3. Frazer Institute, Dermatology Research Centre,, The University of Queensland, Brisbane, Queensland, Australia
  4. Australian Melanoma and Skin Cancer Alliance, Brisbane, QLD, Australia

Background and aim:

To elicit patient preferences for AI use in skin cancer diagnosis, a discrete-choice experiment (DCE) survey was conducted. The survey design was pre-tested with a pilot sample. The pilot aimed to (i) provide beta priors to inform a statistically efficient main survey design, and (ii) check face validity with participants.

Study design and methods

Choice sets consisted of 2 alternatives: skin cancer diagnosis with or without AI assistance, described using four attributes - diagnostic test performance (false positives, false negatives); waiting time to book an appointment; and out-of-pocket costs, three levels for each attribute. The DCE design consisted of six blocks, with participants randomly assigned to one block. The pilot using a D-efficient design in Ngene software consisted of six blocks, 10 choice sets each, with participants randomly assigned to one block. Participants were recruited via the ACEMID platform and newsletter. Choice data were analysed using a multinomial logit (MNL) model using Nlogit software, and trade-offs (i.e. willingness to pay to improve test accuracy or reduce wait times) were calculated. Feedback from participants on survey content and presentation was also collected to inform the main survey design.

Results

34 patients were recruited: 47% aged 55-64 years, 47% females, 56% reported a previous diagnosis of melanoma. Patients showed an overall preference for AI-assisted diagnosis (β =0.73, p =0.02). Both false positives (β = -0.01; p =0.03) and false negatives (β = -0.10; p< 0.001) were significantly associated with choice of diagnostic service. Compared with waiting < one month, waiting > six months significantly reduced preference (β = -0.90; p< 0.001). Out-of-pocket costs were significantly associated with utility (β = -0.00281; p< 0.001). Patients were willing to pay $36 for a 1% reduction in false negatives, and $321 to reduce waiting time from > 6 months to < one month.

Conclusion

The pilot survey was understandable for patients, with results suggesting a preference for AI-assisted skin cancer diagnosis over non-AI-assisted diagnosis. Test accuracy, waiting times and out-of-pocket costs were all associated with preferences, validating these attributes for the main survey design.