The design shown in the demo required one person to take three photos of their flaws from different angles and distances. Users can choose to add information, such as the affected body parts and when they have problems. Click “Submit” to compress the image to Google. Then, the application will display “recommended conditions” to illustrate possible conditions with an image. Clicking on one will display a series of key information, such as symptoms, infectivity, and treatment options. Google said that the app has received training on “hundreds of thousands of skin images”, can identify 288 diseases, including skin cancer, and covers approximately 90% of common dermatological web searches.
The FDA exempts some health software that it considers “lower risk”-such as “health” recommendations such as diabetes management or information about health symptoms-from obtaining medical device approval. It requires the approval of others, such as applications that provide specific diagnostics or applications that are used as medical devices such as stethoscopes. The boundary between applications that require a license and those that do not require a license is difficult to determine because medical software and its management rules are relatively new.
Bradley Thompson, Epstein Becker Green’s regulatory attorney, will ask clients some key questions when trying to determine whether they need FDA approval. They include how the output of the software is presented to a person and whether the company makes a specific medical claim.
Google’s app does not highlight a possible skin condition based on a person’s photo. It displays a warning “The suggested condition listed here is not a medical diagnosis.” A company spokesperson likened the application to a search engine, displaying the results for people to read carefully and draw their own conclusions.
However, Google also emphasized the medical characteristics of skin applications. Health director DeSalvo said that Google developed the application because there are not enough skin experts to help everyone with skin diseases.Google’s Blog post Associated the application with a peer-reviewed study in which the company’s technology was compared with doctors and stated that “our artificial intelligence system can achieve the same accuracy as a dermatologist certified by the US board of directors.
This kind of bragging caught the attention of Thompson’s lawyer. “This does show that this is at least comparable to the capabilities of human doctors,” he said-the FDA may be interested in such claims.
Daneshjou, a dermatologist and researcher at Stanford University, also believes that Google’s app may provide consumers and regulators with medical expertise, not just search results. She said the app may be considered a “high-risk” device and requires FDA approval because some skin diseases such as melanoma can be dangerous.
Daneshjou contributed a A recent study This has raised concerns about the thoroughness of the FDA’s review of artificial intelligence health software. She said that it may be too early to open artificial intelligence dermatology tools to consumers. “If patients think the algorithm is as effective as a board-certified dermatologist, they may be more confident in it,” she said. This may cause people to seek unnecessary biopsies or treatment from doctors, or not make important visits.
Daneshjou said that Google should also disclose more information about how it tests its technology on different skin tones. So far, the company’s artificial intelligence dermatology research involves relatively few people with darker skin.
Google said that these publications do not represent its latest data or image recognition models, which have been improved. The spokesperson said that the design and disclaimer of the dermatology application are based on user experience research; other research is ongoing, and the company will also study how people use the service after the service is launched in Europe.
When deploying other promising artificial intelligence health software outside the laboratory, Google faces practical challenges. In 2018, the company began testing a system that can detect eye diseases in diabetic patients in clinics in Thailand. In 2020, the company released a study on the launch, claiming that the system has Rejected more than 20% Patient images due to issues such as variable lighting and nurses’ practical limitations.
Update, June 23, 11:30 AM EST: This article has been updated to include comments from the FDA.
More exciting connected stories