The design shown in the demo requires a person to snap three photos of their blemish from different angles and distances. The user can optionally add information such as the body part affected and how long they’ve had the problem. Tapping “Submit” zips the images off to Google. The app then displays “Suggested conditions,” showing possible conditions illustrated by images. Tapping on one brings up a list of key information such as symptoms, contagiousness, and treatment options. Google says the app was trained on “hundreds of thousands of skin images” and can identify 288 conditions, including skin cancers, covering roughly 90 percent of common dermatology web searches.
The FDA exempts some health software it deems “lower risk”—such as “wellness” advice like diabetes management or information about health symptoms—from medical device approvals. It requires approval for others, such as those offering specific diagnoses, or apps that function as medical devices such as a stethoscope. The line between apps that need clearance and those that do not is hard to pinpoint because medical software and the rules governing it are relatively new.
Bradley Thompson, a regulatory lawyer with Epstein Becker Green, asks clients a handful of key questions when trying to determine whether they’ll need FDA sign-off. They include how the software’s output is presented to a person and whether a company makes specific medical claims.
Google’s app does not highlight a single possible skin condition in response to a person’s photos, and it displays a warning that “suggested conditions listed here aren’t a medical diagnosis.” A company spokesperson likened the app to a search engine displaying results for a person to peruse and draw their own conclusions about.
Yet Google has also emphasized the skin app’s medical chops. DeSalvo, the health chief, said Google developed the app because there aren’t enough skin specialists to help every person with skin conditions. Google’s blog post links the app to peer reviewed studies in which the company’s technology was compared to doctors, saying “our AI system can achieve accuracy that is on par with US board-certified dermatologists.”
That boast caught Thompson’s lawyerly eye. “That really is suggesting this is at least comparable to what a human physician can do,” he says—the type of claim that might interest the FDA.
Daneshjou, the Stanford dermatologist and researcher, also thinks Google’s app could appear to consumers and regulators to be offering medical expertise, not just search results. She says that the app might be considered a “high-risk” device, requiring FDA approval, since some skin conditions such as melanoma, can be dangerous.
Daneshjou contributed to a recent study raising concerns about how thoroughly the FDA vets AI health software, and she says it may be too early to throw open AI dermatology tools to consumers. “If a patient believes this algorithm is working as well as a board-certified dermatologist, they may have more confidence in it,” she says. That could cause people to seek unnecessary biopsies or treatment from a doctor, or not to make a crucial visit.
Google should also disclose more about how it has tested its technology on different skin tones, Daneshjou says. The company’s AI dermatology studies so far have involved relatively few people with darker skin.
Google says those publications did not represent its latest data or image recognition models, which have been improved. The spokesperson said the dermatology app’s design and disclaimers were informed by user experience studies; additional studies are underway, and the company will also research how people use the service after it is made available in Europe.
Google has faced practical challenges when deploying other promising AI health software outside the lab. In 2018 the company began testing a system capable of detecting eye disease in people with diabetes in clinics in Thailand. In 2020 the company published a study on the rollout that said the system had rejected more than 20 percent of patient images due to problems like variable lighting and practical constraints on nurses.
Updated, 6-23-21, 11:30am ET: This article has been updated with a comment from the FDA.
More Great WIRED Stories