Computer face-matching technology using two-dimensional photographs accurately matches the facial gestalt of unrelated individuals with the same syndromic form of intellectual disability

Abstract

BackgroundMassively parallel genetic sequencing allows rapid testing of known intellectual disability (ID) genes. However, the discovery of novel syndromic ID genes requires molecular confirmation in at least a second or a cluster of individuals with an overlapping phenotype or similar facial gestalt. Using computer face-matching technology we report an automated approach to matching the faces of non-identical individuals with the same genetic syndrome within a database of 3681 images [1600 images of one of 10 genetic syndrome subgroups together with 2081 control images]. Using the leave-one-out method, two research questions were specified:Using two-dimensional (2D) photographs of individuals with one of 10 genetic syndromes within a database of images, did the technology correctly identify more than expected by chance: i) a top match? ii) at least one match within the top five matches? or iii) at least one in the top 10 with an individual from the same syndrome subgroup?Was there concordance between correct technology-based matches and whether two out of three clinical geneticists would have considered the diagnosis based on the image alone? ResultsThe computer face-matching technology correctly identifies a top match, at least one correct match in the top five and at least one in the top 10 more than expected by chance (P < 0.00001). There was low agreement between the technology and clinicians, with higher accuracy of the technology when results were discordant (P < 0.01) for all syndromes except Kabuki syndrome.ConclusionsAlthough the accuracy of the computer face-matching technology was tested on images of individuals with known syndromic forms of intellectual disability, the results of this pilot study illustrate the potential utility of face-matching technology within deep phenotyping platforms to facilitate the interpretation of DNA sequencing data for individuals who remain undiagnosed despite testing the known developmental disorder genes.Electronic supplementary materialThe online version of this article (10.1186/s12896-017-0410-1) contains supplementary material, which is available to authorized users.

Severe neurocognitive and growth disorders due to variation in THOC2 , an essential component of nuclear mRNA export machinery

“…We used computerized face-matching technology to specifically evaluate the cohort to assess if a characteristic facial gestalt was evident across individuals with pathogenic or likely pathogenic variants across our original and this expanded clinical cohort (Supp. Figure S4) (Dudding-Byth, et al, 2017). Although a clearly recognizable facial gestalt was not obvious, there are some similarities.…”

The Discovery of a LEMD2-Associated Nuclear Envelopathy with Early Progeroid Appearance Suggests Advanced Applications for AI-Driven Facial Phenotyping

“…For diagnoses with sufficiently distinct facial phenotypes, algorithms employing machine learning generally fare well in prioritizing potential diagnoses, sometimes with even better accuracy than that of seasoned geneticists. 19,25,26 Another not-yet-exploited utility of these tech-nologies is their potential to detect similarities of novel phenotypes to known disorders and to thereby identify a potential molecular etiology, such as the disruption of a signaling pathway or a specific cellular compartment. In our case, DeepGestalt recognized the facial similarity of individuals with the LEMD2 mutation c.1436C>T to individuals with another nuclear envelopathy, HGPS.…”

De Novo and Inherited Pathogenic Variants in KDM3B Cause Intellectual Disability, Short Stature, and Facial Dysmorphism

“…Facial recognition software is able to objectively detect subtle features which might be difficult to identify by eye. [16][17][18] In the past, facial recognition software has been used to objectively measure dysmorphic features in individuals with various syndromes. [17][18][19] We collected and analyzed photographs of 13 out of 14 children.…”

“…[16][17][18] In the past, facial recognition software has been used to objectively measure dysmorphic features in individuals with various syndromes. [17][18][19] We collected and analyzed photographs of 13 out of 14 children. The three affected parents were deliberately excluded to avoid a bias caused by facial features that were shared with their children and unrelated to the variant identified.…”