2020 Update on prenatal phenotyping

Designing dysmorphology machine learning

The cohort included 679 individuals with 105 different genetic disorders. All individuals had a previously confirmed molecular diagnosis that would be detected by ES. Each individual had a frontal facial photograph analyzed and detailed clinical features documented in HPO terms extracted by 2 clinicians.

A facial analysis software called DeepGestalt, trained on 17,000 patient images, was used to create a Gestalt score. Each individual had 4 different predicted gene scoring approaches:

• a molecular deleteriousness score
• facial analysis with the Gestalt score
• a combination of molecular deleteriousness score and HPO-based gene-prioritization tool (termed semantic similarity score)
• the PEDIA score, which included all 3 prior approaches.

A type of machine learning algorithm (support vector machine, or SVM) was applied, validated, and used to prioritize genes based on the combined scores.

AI seemed to improve diagnostic accuracy

Utilizing the combination of machine learning, HPO terms, and facial analysis software greatly improved the accuracy of variant classification predictions over any approach alone.

Using only the sequence variant and molecular deleteriousness score, the causative variant was ranked in the top 10 of all identified variants in less than 45% of cases. Adding the HPO-based gene prioritization tools increased the accuracy to 63% to 94%. Use of the PEDIA score, which incorporated all 3, increased the accuracy to 99% for the top 10 ranking.

Even more impressive improvements were made in the top 1 ranking accuracy rate, which went from 36% to 74% without facial image information to 86% to 89% with inclusion of DeepGestalt scores.

Study strengths and limitations

This study’s innovative application of facial analysis and machine learning, combined with HPO-driven variant classification, showed added benefit. To achieve this with available patient photographs and thorough phenotyping, previously diagnosed patients were used. Because complete ES information was not available for those patients, their known pathogenic variant was inserted into randomly selected exomes from the 1000 Genomes Project (healthy individuals). The authors additionally noted that the PEDIA score performance was diminished for rare disorders in which limited data were available.