Melanoma, Skin Cancer
Conditions
Keywords
melanoma, artificial intelligence, AI augmentation, deep learning, teledermatoscopy, dermatoscopy, skin cancer
Brief summary
In this study an artificial intelligence (AI) tool for skin cancer diagnosis is implemented in a teleldermatoscopy platform. The aim is to study the effects on clinician diagnostic accuracy, management decisions, and confidence. Furthermore, this prospective randomized study investigates the role of human factors in determining clinician reliance on AI tools and the consequent accuracy in a real-world setting.
Detailed description
Deep-learning algorithms can potentially benefit many areas in healthcare, including the diagnosis of skin cancer using teledermatoscopy. However, there is a dearth of clinical, prospective research on human-AI interaction in diagnostic tasks that take human factors into account. In this study we will examine the impact of such factors in a real-world setting where we integrate an algorithm in an existing teledermatoscopy platform that is used clinically at a tertiary hospital in Sweden. We will investigate what impact various implementations of AI tool output in relation to human factors have on diagnostic accuracy and management decisions. Study subjects are recruited at the Department of Dermatology at Karolinska University Hospital and will be asked to rate prospective teledermatoscopic consults with and without AI-support. Each consult will be randomized into one of three workflows with or without one pre-defined implementation of the AI tool. Study subjects are also asked to complete two surveys with demographic information and questions relating to various human factors. Patients participating in the study will be diagnosed outside the study prior to inclusion without any involvement of an AI tool, notably by two experienced dermatologists who do not participate as study subjects.
Interventions
OTHERAI assistance
Participants will be informed of the diagnostic probabilities for each of ten differential diagnoses according to the AI tool
Sponsors
Karolinska Institutet
Medical University of Vienna
Stockholm School of Economics
Karolinska University Hospital
Study design
Allocation
RANDOMIZED
Intervention model
CROSSOVER
Primary purpose
DIAGNOSTIC
Masking
NONE
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
Yes
Inclusion criteria
* Licensed physician * Working at a dermatology clinic * Sufficient knowledge in Swedish * Written consent to participate
Exclusion criteria
* No experience of using dermatoscopy * Does not wish to participate * Incomplete answers * Physicians that are involved in the patients' clinical care relating to the teledermoscopical consult
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Diagnostic accuracy | 1 year | Determine sensitivity, specificity, accuracy and AUROC in terms of diagnostic accuracy for dermatologists with vs without AI advice. Further, to investigate the role of the different workflows (diagnosis with or without AI with varying sequencing) and the influence of demographics and human factors (e.g. level of experience) on diagnostic accuracy |
| Accuracy of management decisions | 1 year | Determine sensitivity, specificity, accuracy and AUROC in terms of accuracy for management decisions for dermatologists with vs without AI and investigate the role of the different workflows (with or without AI with varying sequencing) and the influence of demographics and human factors (e.g. level of experience) on management decisions (biopsy/surgery, no intervention, or follow-up) |
| Tendency to change initial diagnosis or management decision | 1 year | Evaluate which factors affect the likelihood of a physician changing their evaluation after receiving algorithmic input |
| Self-reported confidence in diagnosis and management decisions | 1 year | Investigate whether AI or other factors affect the physician's confidence in their diagnosis and management decisions |
Countries
Sweden
Outcome results
None listed