by Paul Govern
Chronic graft-versus-host disease (cGVHD) is a leading cause of morbidity and mortality in allogeneic stem cell and bone marrow transplant recipients. When studying therapies, the proportion of skin affected by cGVHD is a key measure of patient response. Estimates from these visual evaluations vary considerably from one researcher to the next. Artificial intelligence could streamline and standardize these evaluations.
Using annotated 3D photographs of 36 research participants, a team at Vanderbilt University Medical Center trained a machine learning algorithm to identify areas of skin affected by cGVHD. As reported in the British Journal of Haematology, Andrew McNeil, PhD, Eric Tkaczyk, MD, PhD, and colleagues found that their algorithm performed on a par with clinicians.
In a blinded test, a dermatologist scored the algorithm’s performance as comparable in most cases with the project’s human annotator, a cGVHD trainee. The algorithm’s accuracy was also found comparable to that of in-person evaluations by clinicians as measured in a previous study.
Also on the study were Kelsey Parks, Xiaoqi Liu, PhD, Inga Saknite, PhD, Fuyao Chen, Tahsin Reasat, MS, Austin Cronin, Lee Wheless, MD, PhD, and Benoit Dawant, PhD. The study was supported in part by the Department of Veterans Affairs (IK2 CX001785) and the National Institutes of Health (CA090625, AR074589).
