JMIR Publications

ABSTRACT

Background:

Collectively 5% of the population have a genetic disease. Many of them feature characteristics that can be detected by facial phenotyping. Face2Gene CLINIC is an online app for facial phenotyping of patients with genetic syndromes. DeepGestalt, the neural network driving Face2Gene, automatically prioritizes syndrome suggestions based on ordinary patient photographs potentially improving the diagnostic process. Hitherto, studies on DeepGestalt’s quality highlighted its sensitivity in syndromic patients. However, determining a diagnostic methodology’s accuracy also requires testing of negative controls.

Objective:

The goal of this study was to evaluate DeepGestalt's accuracy on photos of individuals both with and without a genetic syndrome. Moreover, we sought to propose a machine-learning based framework for the automated differentiation of DeepGestalt’s output on such images.

Methods:

Frontal facial images of a convenience sample of individuals with a diagnosis of a genetic syndrome established clinically or molecularly were reanalyzed. Each photo was matched by age, sex and ethnicity to a picture featuring an individual without a genetic syndrome. Absence of a facial gestalt suggestive of a genetic syndrome was determined by physicians working in medical genetics. Photos were taken from online reports or made by us for the purpose of this study. Facial phenotype was analyzed by DeepGestalt v. 19.1.7 accessed via Face2Gene CLINIC. Furthermore, we designed linear support vector machines (SVMs) using Python 3.7 to automatically differentiate the two classes of photographs based on DeepGestalt's results lists.

Results:

We included data of 323 syndromic patients covering 17 different syndromes, and matched them to an equal number of facial images without a genetic syndrome analyzing a total of 646 pictures. We confirm DeepGestalt’s high sensitivity (top-10-sensitivity 295/323, 91%). DeepGestalt’s syndrome suggestions in individuals without a craniofacially dysmorphic syndrome followed a non-random distribution. 17 syndromes appeared in the top-30-suggestions of more than 50% of non-dysmorphic images. DeepGestalt’s top-1-scores differed between syndromic and control images (AUROC 0.72, 95% CI 0.68-0.76; P<.001). A linear SVM running on DeepGestalt’s result vectors showed stronger differences (AUROC 0.89, 95% CI 0.87-0.92; P<.001).

Conclusions:

DeepGestalt fairly separates images of individuals with and without a genetic syndrome. This separation can be significantly improved by SVMs running on top of DeepGestalt thus supporting the diagnostic process of patients with a genetic syndrome. Our findings facilitate the critical interpretation of DeepGestalt’s results and may help enhance it and similar computer-aided facial phenotyping tools.