AI models outperform pediatricians in diagnosing genetic syndromes using facial recognition

Facial recognition models, particularly those incorporating pretrained foundation models (PFMs) and CosFace loss functions, significantly enhance the accuracy of diagnosing Williams-Beuren syndrome, Noonan syndrome, and Alagille syndrome, according to a study. These models outperform traditional diagnostic methods, including those used by experienced pediatricians, and show promise as effective tools for the clinical screening of these genetic disorders.

In a recent study, 3297 facial photos from children with Williams-Beuren syndrome (n = 174), Noonan syndrome (n = 235), Alagille syndrome (n = 51), and those without genetic syndromes (n = 1206) were analyzed. The photos were divided into 5 subsets for training and testing, with a 4:1 ratio. Researchers utilized the ResNet-100 architecture to develop 4 facial recognition models: 2 with pretrained PFMs and 2 without, each employing either ArcFace or CosFace loss functions.

Using ResNet-100 with PFM and CosFace loss function achieved the highest accuracy (84.8%). Pretraining with a PFM notably improved performance, increasing accuracy from 78.5% to 84.5% for ArcFace and from 79.8% to 84.8% for CosFace. Both loss functions showed similar performance with PFM, outperforming the accuracy of 5 pediatricians, among whom the most experienced one achieved an accuracy of 70%.

Reference
Shen JJ, Chen QC, Huang YL, et al. Facial recognition models for identifying genetic syndromes associated with pulmonary stenosis in children. Postgrad Med J. 2024;qgae095. doi: 10.1093/postmj/qgae095. Epub ahead of print. PMID: 39075977.

Rare Disease 360® is the Official Media Partner and Official Publication of The Alagille Syndrome Alliance (Alagille.org).