Artificial intelligence to estimate extent of cGVHD from patient photos - PROJECT SUMMARY / ABSTRACT Lack of reliable assessment tools is a major obstacle to validating novel treatments for chronic graft-versus- host disease (cGVHD), the leading cause of long-term morbidity and mortality after allogeneic hematopoietic stem cell transplantation (HCT) to cure life-threatening blood diseases. Skin erythema is a key cGVHD biomarker, but standard of care in-person evaluations are subjective, prone to error, and costly. We will validate an artificial intelligence (AI) technology for accurate, efficient, and easily deployable measurement of cGVHD erythema from photographs in diverse patient populations. In our preparatory study, AI technology achieved human-level performance under controlled photography conditions. We propose to refine and validate this AI technology in a multicenter cGVHD patient cohort of unprecedented size, leveraging experts in dermatology, transplant medicine, medical imaging, artificial intelligence, biomedical informatics, and data science. A unique database of over 11,000 photographs and clinical information from ethnically and phenotypically diverse patients will be assembled from five major cancer centers: Fred Hutchinson Cancer Center, Mayo Clinic, the National Institutes of Health, University of Pennsylvania, and Vanderbilt University Medical Center. Accuracy to measure cGVHD erythema will be determined relative to expert dermatologist-level assessments. We will quantify and overcome potential biases for the AI including skin tone, gender, photography conditions, and disease severity. We will compare the prognostic value of AI and human assessments as biomarkers of mortality. Finally, we will prospectively benchmark the accuracy of AI measurements against standard in-person clinical trial assessments (NIH Skin Scoring) for patients recruited at all five cancer centers. The proposal could improve patient care and telemedicine through: consistent cGVHD scoring equivalent to specialist examination; visualizing cutaneous changes for quality assurance in observational and therapeutic studies; enabling frequent longitudinal monitoring at home or in clinic; and relieving the burden of time- consuming skin area assessment on patient care providers. 1
