HealthDay News — A direct-to-consumer machine learning model for detecting skin cancers may not adequately recognize rare, but aggressive, skin cancers, according to a study presented at the annual meeting of the European Academy of Dermatology and Venereology, held virtually from Sept. 29 to Oct 2.
Lloyd Steele, M.B.Ch.B., from the University of London, and colleagues assessed the performance of a machine learning model for Merkel cell carcinoma (MCC) and amelanotic melanoma. The performance of a direct-to-consumer model, which is available in Europe as a certified medical device, was assessed using a set of images that included 28 MCCs, 35 amelanotic melanomas, 28 seborrheic keratoses, and 25 hemangiomas.
The researchers found that the direct-to-consumer app incorrectly classified five of 28 MCCs (17.9 percent) and seven of 35 amelanotic melanomas (22.9 percent) as low risk. Nearly two-thirds of benign lesions (62.2 percent) were classified as high risk. The model’s sensitivity for detecting malignancy was 79.4 percent, with a specificity of 37.7 percent.
“In order to improve, machine learning model evaluations should consider the spectrum of diseases that will be seen in practice,” Steele said in a statement. “At the moment, most of the performance of those models is driven by the imaging data available, which is particularly scarce when it comes to rare skin cancers.”