PracticeUpdate Conference Series_WORLDSymposium 2019

Algorithm That Incorporates Facial Features Helps Identify Mucopolysaccharidoses The approach could be valuable in countries where access to diagnosis is limited. A n algorithm incorporating frontal facial images has suc- ceeded in suggesting mucopolysaccharidosis (MPS) as a possible diagnosis. This outcome of an evaluation of the algorithm was reported at the WORLDSymposium 2019.

Roberto Giugliani, MD, of the Hospital de Clínicas de Porto Alegre and the Universidade Federal do Rio Grande do Sul in Porto Alegre, Brazil, and col- leagues set out to determine whether a tool for facial recognition could identify patients with MPS III and IV. Dr. Giugliani told Elsevier’s PracticeUp- date , “New technologies like this one are needed to identify patients earlier,

Dr. Roberto Giugliani

especially in locations where medical resources are scarce. Such tools, easy to apply and requiring only a smartphone, are helpful diagnostic aids in the feld of rare diseases." The investigators took the wide spectrum of clinical manifesta- tions of mucopolysaccharidoses into consideration and explored the benefits of noninvasive phenotypic characterization using frontal pictures. The team evaluated version 18.1.14 of the algorithm with pictures from 25 known patients with MPS IIIB, n=16, or IVA, n=9. A deidentified case was created with frontal pictures of each patient, including date of birth, height, weight and head circumference. The phenotype was refined with the following descriptors: coarse facial features, hepatomegaly, short stature, intellectual disability, psychomotor deterioration, abnormal heart valve morphology, cardiomyopathy, hernia, hearing impairment, kyphosis, sple- nomegaly, genu valgum, corneal opacity, dysostosis multiplex, respiratory distress, and pectus carinatum, according to the clinical examination. Face2Gene scores were classified as exhibiting low, medium, and high similarity. Overall, 1 patient (4%) was scorable for MPS, 9 patients (36%) scored lowest, 14 (56%) medium, and one (4%) highest. Among patients with MPS IIIB, 15 (94%) were suggested to suffer from MPS within the top fve syndromes. Ten (62%) were sug- gested to suffer from MPS IIIB as one of the top fve suggested diagnoses. Among patients with MPS IVA, all were suggested as harboring a MPS. None, however, was suggested. Other than MPS, the most common suggested diagnoses were: Angelman, DiGeorge, and Williams-Beuren syndrome. Dr. Giugliani explained that the Face2Gene algorithm extracts mathematical facial descriptors from patients’ frontal pictures and compares them with syndrome gestalts with the goal of suggesting potential diagnoses.

Deep learning algorithms build syndrome-specifc, computational- based classifers (syndrome gestalts). Proprietary technology converts a patient photo into de-identifed mathematical facial descriptors (facial descriptors). The patient’s facial descriptor is compared with syndrome gestalts to quantify similarity (gestalt scores), resulting in a prioritized list of syndromes with similar morphology. Artifcial intelligence suggests likely phenotypic traits and genes to assist in feature annotation and syndrome prioritization. Dr. Giugliani concluded that the algorithm incorporating frontal facial images succeeded in suggesting MPS as a possible diag- nosis in most patients based on frontal images only with minimal clinical data and without testing. Nonetheless, improvements are still necessary. The approach could be valuable in low-income countries where access to diagnosis is limited.

www.practiceupdate.com/c/80420

WORLDSymposium 2019 • PRACTICEUPDATE CONFERENCE SERIES 5