Identifying facial phenotypes of genetic disorders using deep learning

Area of Science:

• Medical Genetics
• Computer Vision
• Artificial Intelligence

Background:

• Syndromic genetic conditions affect 8% of the population, often presenting with recognizable facial features.
• Facial analysis technologies show promise but are limited to identifying only a few phenotypes.
• Clinical genetics requires consideration of hundreds of possible diagnoses.

Purpose of the Study:

• To develop and evaluate DeepGestalt, a facial image analysis framework for identifying a wide range of syndromic genetic conditions.
• To assess DeepGestalt's performance against expert clinicians in syndrome identification tasks.

Main Methods:

• Utilized computer vision and deep-learning algorithms to create the DeepGestalt framework.
• Trained the model on a dataset of over 17,000 images across more than 200 syndromes.
• Conducted experiments comparing DeepGestalt's accuracy to that of clinicians.

Main Results:

• DeepGestalt outperformed clinicians in distinguishing between syndromes and identifying subtypes.
• Achieved 91% top-10 accuracy in identifying the correct syndrome from 502 clinical images.
• Demonstrated the framework's ability to quantify similarities to hundreds of syndromes.

Conclusions:

• DeepGestalt significantly advances the potential of facial analysis in syndromic genetic condition diagnosis.
• The framework offers substantial value for phenotypic evaluations, genetic testing, and precision medicine.
• Community-driven data curation enhances the robustness and applicability of AI in clinical genetics.