BARCELONA – Exome sequencing has a diagnostic rate on the order of 25 to 40 percent, depending on the patient cohort and whether or not a patient had prior testing. Automated facial recognition software may help to increase this yield for disorders that come with facial dysmorphology, according to research presented at the European Society of Human Genetics annual meeting here this week.
In a talk on Sunday, Martin Mensah, a researcher at the institute for medical genetics and human genetics at Charité in Berlin, presented results from a study to test the performance of automated facial recognition software called Facial Dysmorphology Novel Analysis, available from FDNA, a company based in Boston, for predicting a patient's genetic disorder. Along with other deep phenotyping tools, such software might be able to help prioritize what genes to analyze for pathogenic variants in exome sequencing data.
The algorithm analyzes a patient's facial features from a photograph and compares it against a database of so-called "syndrome masks" – typical facial characteristics of a syndrome that were generated by training the software with photos from patients with confirmed diagnoses. It then creates a heat map of similarity and ranks the syndromes in order of greatest facial likeness.
The researchers tested the software on 130 patients with confirmed molecular diagnoses that covered 72 different monogenic disorders. In 11 percent of cases, the first hit delivered by the tool was the correct diagnosis, and in another 21 percent, the correct diagnosis was among the top 10 hits.
Some cases would have been easy to recognize by a clinician, and the software had no problems with those, either. Other cases were trickier, but the tool still delivered the correct diagnosis – for example, a patient with Rothmund-Thomson syndrome wore makeup in the photograph to disguise the absence of eye brows but the software was still able to pick up the disorder. The correct diagnosis did not always make it near the top, though: for a patient with CHARGE syndrome, for example, it was only ranked ninth.
In addition, the software can get thrown off – when Mensah uploaded a photo of himself, for example, it suggested two intellectual disability syndromes. However, the tool might have been stumbled by facial features related to his ethnic background, which is Central European and African, because the algorithm was trained on patients with European ethnicity, he said.
In 41 percent of cases, the software did not work – possible reasons, Mensah suggested, are that the database did not include a syndrome mask or there were not enough reference images for the disorder in question — photos from about 30 patients are needed to create a mask. Other possible reasons could be that the photograph of the patient was unsuitable, or that the syndrome does not have typical facial features.
The researchers also compared the facial recognition software with a feature match tool that predicts disorders from expert-annotated features and found that there was good correlation between the two, with 67 cases being solved by both tools. Overall, the feature match tool worked better, but the main reason was the absence of syndrome masks for some disorders, so the facial recognition software could not match them. Also, for nine cases, only the facial analysis suggested the correct diagnosis, which would have been missed by the feature match tool.
The researchers also tested how the FDNA technology could help with the analysis of exome sequencing data. For those patients where they did not have exome data available, they spiked the patient's pathogenic variant into exome data derived from the 1,000 Genomes Project. They then filtered the exomes for rare variants with effects on proteins and found that in combination, exome data and facial recognition-based predictions often delivered the correct diagnosis. For example, in the case of a patient with Smith-Lemli-Opitz syndrome, facial dysmorphology software initially ranked the diagnosis fifth, but with exome data, it became the first hit.
The Berlin team is currently planning a larger study, Mensah said, and is looking for collaborators able to contribute data from patients with confirmed molecular diagnoses and, if possible, exome data.