“We should not be confident in our ability to keep a super-intelligent genie locked up in its bottle forever.”
Nick Bostrom, Swedish philosopher
Earlier this week at its I/O developer conference, Google announced its AI-enabled Dermatology Assist tool that can help detect skin (as well as hair and nail) conditions with three photographs uploaded from a user’s phone from different angles. These pictures are accompanied by a series of questions such as duration of the condition and other symptoms.
This Google AI app received its CE (conformité européenne) mark as a class I medical device in Europe and is scheduled to be launched later in 2021. This is one of the first instances of an AI tool in healthcare that will be available to lay people rather than only clinicians.
This tool, developed with a de-identified dataset of about 65,000 images with diagnosed conditions and a deep learning system, has a portfolio of close to 300 conditions along with frequent questions and answers relevant to the condition that have been reviewed by dermatologists. This tool is not meant to be used as a diagnostic device that would potentially delay appropriate diagnosis (including biopsy) and treatment by a dermatologist.
There are several important issues that are very relevant to this development of AI tools in healthcare:
Public education. There are close to 10 billion searches for skin, hair, and nail issues on Google’s search engine on an annual basis with about 2 billion people worldwide with skin conditions. One potential benefit of a tool such as this is expediting the diagnosis of a lethal skin condition that would have had a delayed or even missed diagnosis.
In addition, it is exceedingly difficult, if not impossible, for a lay person to be able to accurately describe a skin condition using the Google search bar so this tool can facilitate the query. Public education using this AI tool for skin lesions can lower the barriers due to access, and this increased access is particularly useful in regions in the world with little or no access for even primary care.
Diagnosis. One potential disadvantage is an incorrect labeling of the skin lesion that would lead to an early false reassurance and therefore ultimately a delayed or missed diagnosis of a lethal condition. In addition, there is also a good chance that while few patients will have appropriate workup for a suspicious lesion, many will be undergoing unnecessary referrals to dermatologists. It behooves the health system to have an ongoing assessment and follow-up of these patients as well as providing feedback to this AI tool.
Medical legal framework. The medical legal framework will need to be carefully delineated for both the clinician and the AI technology partner for this and other upcoming AI-enabled medical tools in the future. The learned intermediary rule asserts that if the clinician has been given adequate information about a product’s risks, it is the clinicians duty to convey that information including its warning to the patient. If there is direct consumer marketing or if the clinicians’ role is limited, however, this rule cannot be invoked.
In this new paradigm of delivering healthcare in the form of AI-enabled tools, the traditional clinician-patient dyad is possibly changed and there will not only be legal, but ethical, regulatory, and financial consequences.
In addition, there needs to be an effective coupling between AI and clinical outcome to achieve true synergy. This AI tool, for it to be impactful, will need to result ultimately in not only more expedient diagnoses, but improved patient outcomes and increased healthcare value.
