Knowledge gap: Using computer vision applied to 3D medical imaging for clinical diagnostics hae already been proposed for schyzophrenia or autism. However, no approach has been proposed to take advantage of the numerous regular photos available online.

Strategy: * In this paper, they propose a novel approach, called 'Clinical Face Phenotype Space', able to analyze ordinary photographs and predict corresponding genetic disorders. * This model has been train on about 3,000 images, representing 8 major disorders, described using 36 feature points and their pairwise distance. * They address the problem of variability in everyday pictures, such as lighting, pose and facial expression, by using Metric Learning, which is related to Principal Component Analysis and a good way to change data representation by compressing uninformative dimensions, reducing their impact.

Main findings: * In addition to classify very well the 8 major disorders, they test it with 75 disorders not used to train the model and shows that it helps to narrow diagnostic search space. * They also derive probabilistic ranking, returning confidence for several similar syndromes, for a given picture. * Their approach computes a score that estimates the relative closeness of pictures and possibly detect new clusters and ultra-rare disorders.

Strengths: * All the code seems to be available online (Python and Matlab mixture). The 3,000 pictures are also available. * The use of ordinary pictures can be challenging and I think they address data variability in a nice way.

Weaknesses: * I found one of their analysis not very convincing: they also test whether disorders caused by mutations in similar genes result in clusters in their phenotype space, which could make sense. For only monogenic syndromes, they derive protein-protein interaction network and measure similarity as the path distance between genes. They compared 4 groups: distance =1 (30), 2 (153), 3 (152) or (4 + unknown, 3,482), where unknown regroups: no gene associated with syndrome, multigenic syndrome, no known protein-protein interaction.