Facial analysis for GPIBDs
Computer assisted facial recognition
Patients with mutations in genes of the GPI anchor biosynthesis pathway often share a characteristic facial gestalt. Yet, only very experienced clinicians were able to identify and describe the typical characteristics of a patient's face. In this research project, we are exploring ways to improve the diagnostic process by means of artificial intelligence (AI), neural networks, and algorithms.
PEDIA research project where we are working on prioritization of exome data by image analysis of facial photographs, we have collected over a hundred photos of GPIBD cases. Using a neural network (DeepGestalt) developed by FDNA (Face2Gene) we were able to achieve an accuracy of over 50% in correct gene prediction for the five most prevalent GPIBDs: PIGA, PIGN, PIGT, PIGV, and PGAP3. Moreover, we identified gene-specific substructures in the GPI pathway. The remarkable information content of human faces advocates for use of computer-assisted syndromology in the definition of disease entities.
Challenge
Large data sets are optimal for training and validating the applied AI. However, GPIBDs are a very rare type of monogenic disorder, with a prevalence ranging between 1/50.000 – 1/150.000. Due to the rarity of the syndrome we can not rely on large data sets of images. In addition, analysis performed by AI may be biased by ethnic background, age, sex, and uneven sample size. Therefore, we are exploring alternative methods to identify the
Novel approach
We want to increase the prediction accuracy of the mutated gene in affected individuals by training a neural network with frontal face photos of healthy family members. Thereby, we want to reduce familial similarity and increase syndromic features in the face.
We are investigating whether analysis of facial images form families, including affected patients, their parents, and - if available - healthy siblings, can improve the accuracy of predicting the correct syndrome / mutated gene.
Our final goal is the improvement of the diagnostic yield for individuals suspected with GPIBD.
Please find more information about the study here: proband information.
Participants of the study are asked to fill out a consent form
and may send images of their family members to pedia@uni-bonn.de