New guideline outlines practical and ethical considerations for adoption of the technology in dermatology practice.
The Australasian College of Dermatologists has released a position statement on the use of artificial intelligence in dermatology in Australia.
The statement was adapted from a paper published in the Australasian Journal of Dermatology and includes a full set of recommendations for dermatologists.
The statement aims to “provide a high-level overview of the Australasian College of Dermatologists (ACD) key policy recommendations to guide the development, regulation and adoption of Al in dermatology in Australia”.
“Dermatologists use AI for various research projects and the combination of AI, research and clinical judgement will have benefits for patient care,” ACD Fellow Professor Peter Soyer told Dermatology Republic.
Broadly speaking, dermatology lends itself well to AI based on visual pattern recognition from image data – it’s also utilised for image classification in ophthalmology, radiology and cardiology. Thus far, however, most dermatology research is in the area of skin cancer, whereas skin rashes and inflammatory conditions are ideal candidates for future research.
While acknowledging the benefits of developing AI, there are several key issues with the way data is typically collected that limit its usefulness in clinical practice.
For example, write the authors of the paper, “AI models are highly context-specific to the data on which they were trained” and may not be generalisable for people with gender, ethnicity or race that differs from those in the training set.
The field also suffers from a lack of prospective clinical trials to externally validate AI models. Meanwhile, clinicians may be influenced in their decision making by false positives that may tie up health system resources, and false negatives could have serious health consequences.
At present, say the authors, AI technology is not ready for clinical use or direct-to-consumer use. While there are clinical and consumer-level AI models currently approved and in use in various countries, including Australia, their functionality is limited and they’re used in conjunction with medical professional input.
To this end, the ACD recommends AI is used to augment rather than replace clinical judgement. Indeed, the American Academy of Dermatology in its position statement prefers to use the term “augmented intelligence”, rather than artificial intelligence, to reinforce the notion that AI detection, diagnostics and other outcomes assist rather than replace human expertise.
AI software is regulated by the TGA as Software as Medical Device (SaMD) and must have TGA pre-market approval before it’s supplied in Australia.
Regulation in Australia has lagged somewhat behind in the development and uptake of AI, although it is currently undergoing reform to be consistent with international regulations.
One key change relates to the interpretation of risk. AI models were originally considered “low risk” on the basis they don’t cause physical harm. However, new rules consider potential harm from providing users with incorrect information. They also take into account the user’s expertise, with a higher risk classification for direct-to-consumer devices compared with those designed for use by health professionals.
The authors also addressed ethical issues, noting that a recent review of guidelines on the ethical use of AI found 84 published guidelines and there was no single ethical principle common to all.
Forming the basis of the ACD ethical guidelines are principles of transparency (informing patients of AI use), justice and fairness (including images of skin of colour for training, validating and testing AI models); non-maleficence (using AI that is equal to or improves on clinician performance); responsibility (the weight of responsibility lying with AI vendors vs clinicians is currently unclear); and ensuring patient privacy.
The position statement provides advice on what to consider when choosing an AI model and the ACD recommends adopters develop knowledge and skills in the principles of AI model development and AI architecture, with a grounding in the different AI methods provided in the explanatory information paper.
