They also don’t report the diversity of their dataset
Most algorithms designed to help people identify skin problems do not let experts view the datasets they were developed with and share information about the skin tone or ethnicity of patients in those datasets. Didn’t. a new review. The authors argue that this can make it difficult for people to evaluate programs before using them and to understand whether they may not work as well for certain groups of people.
These types of devices use pictures of skin conditions to teach the system to recognize the same conditions in new images. One can upload a picture of a rash or mole and the tool can tell what type of rash or mole it was.
Paper, . published in JAMA Dermatology, analyzed 70 studies that developed either a new Deep learning models or tested existing algorithms on a new set of data. Taken together, the models were developed or tested using more than 1 million images of skin problems. Only a quarter of those images were available for review by experts or the public, the analysis found. Fourteen studies included information about the ethnicity or race of patients in their data, and only seven described their skin types.
The rest did not share a demographic breakdown of their patients. “I highly suspect that these datasets are not diverse, but there is no way of knowing,” said Roxana Daneshjou, a clinical scholar in dermatology at Stanford University. Twitter.
The analysis also examined whether models aimed at identifying skin cancer were trained on images where cancer was confirmed with a skin sample sent to a lab – the “gold standard” for making sure the diagnosis was correct. Of the included studies, 56 claimed to have identified those conditions, but only Of them, 36 met the gold standard. Which could not have been less accurate, the authors say.
Included in the review an algorithm From Google, which has developed a tool designed to help people identify skin conditions. The company plans to build a pilot version of its web tool, which lets people upload pictures of a skin problem and receive a list of possible conditions, later this year. According to the analysis, the Google paper included a breakdown of skin type and an ethnicity, but did not provide publicly used data or models. It also did not use gold standard methods for assessing certain types of skin cancer, including melanoma and basal cell carcinoma.
Medical algorithms are only as good as the data they were developed with, and they may not be as effective when used in situations different from the ones they were trained on. Experts therefore argue that the data, or the description of that data, should be freely available: “The data that is used to train and test a model can determine its applicability and generalizability. Therefore, A clear understanding of the data set characteristics … is important,” the authors wrote.
The lack of transparency is a consistent problem with medical algorithms. As of February 2021, most AI products approved by the Food and Drug Administration (FDA) do not report critical information about the data they were developed with. state news Investigation. The FDA told state news Its new “action plan” for AI emphasizes greater transparency.
The limitations do not mean that most dermatology algorithms are useless, wrote Philipp Tschandl, a researcher at the Medical University of Vienna, in an accompanying editorial. Therapists are also not perfect and have their own biases or knowledge gaps that may undermine their interpretation of a skin problem. “We know this and still manage to practice medicine well,” he wrote. “We need to find interpretability, smart screening and risk mitigation methods to allow algorithms to work safely and equitably in medicine.”