NEW YORK (GenomeWeb) – A number of research groups have explored how automated facial image analysis technology can help with the differential diagnosis of inherited diseases, including prioritizing exome sequencing variants for analysis and grouping patients with mutations in different genes.
At the European Society of Human Genetics annual meeting in Copenhagen this week, several groups presented projects that involved automated facial image analysis. Most teams have been collaborating with FDNA, a Boston-based company that has commercialized facial recognition software for clinicians and researchers in genetics. In addition, a team from the University of Oxford in the UK has launched a project, called Minerva & Me, that uses a different facial analysis software.
Researchers at the Institute of Medical Genetics and Human Genetics at Charité University Medicine Berlin have been testing FDNA's technology in a study called Prioritization of Exome Data by Image Analysis (PEDIA). Last year at ESHG, the group presented early results from that project.
FDNA's software, called Face2Gene, analyzes a frontal facial photograph of a patient for characteristic patterns, extracting a set of de-identified markers. It then compares those data to a database of patterns that were generated from photos of patients with known syndromes, a process the company refers to as "gestalt match". Based on similarity, the algorithm then comes up with a list of suggested diagnoses. For use by clinicians, the Face2Gene software also takes clinical features into account.
At the ESHG meeting, Peter Krawitz, a group leader at the Institute of Medical Genetics and Human Genetics, explained that rare variants from exome sequencing data are prioritized for analysis using clinical features of the patient that are encoded in human phenotype oncology (HPO) terms. The idea is that disease genes are associated with characteristic phenotypic abnormalities. His group wanted to find out how much FDNA's facial matching software could add to the performance of this approach.
Initial tests with patients with confirmed diagnoses, he said, showed varying results. For a patient with cleidocranial dysplasia, for example, FDNA ranked the correct diagnosis first, but for a patient with Smith-Lemli-Opitz syndrome, it only ranked it ninth.
Surprisingly, the software was able to detect Crouzon syndrome in the mother of a patient with that syndrome, even though she is only a carrier of the underlying gene defect.
For the first phase of the PEDIA study, which will close next month, the researchers analyzed photos from more than 450 syndromic patients who had confirmed molecular diagnoses, submitted by more than 20 institutions. According to Krawitz, using facial pattern matching alone, the correct diagnosis was ranked number one 18 percent of the time, and using both genetic variant data — spiked into exome data — and clinical features, it ranked first 65 percent of the time. Combining both approaches, the correct disease became the first pick 81 percent of the time.
For the second phase of the PEDIA study, which will run until the end of this year, the researchers are looking for unsolved cases — patients with facial phenotypes who do not have a molecular diagnosis yet — and are offering researchers submitting such patients exome sequencing and a PEDIA analysis. The plan is to group patients with similar facial phenotypes in order to help find a molecular diagnosis for them.
Krawitz cautioned, though, that the technology does not work well in ethnic backgrounds other than European at the moment because fewer photos from patients with other ethnicities have been published. "Ethnic background, sex, and age are all potential confounders that need to be analyzed," he said.
In addition to the PEDIA study, Krawitz and his colleagues have used the Face2Gene software to sub-categorize patients with the same disorder who have mutations in different genes. For this project, they analyzed facial images of patients with a pathway disorder called glycosylphosphatidylinositol biosynthesis defect (GPI-BD).
According to Alexej Knaus, a researcher at the Institute for Medical Genetics and Human Genetics who presented the work during an FDNA-sponsored conference workshop, GPI-BD has been associated with recessive mutations in 14 genes that are part of the same pathway. Patients share a number of clinical phenotypes, among them intellectual disability and seizures, but they also have non-overlapping clinical features.
For their study, the researchers uploaded photos of patients with a mutation in one of five different genes to Face2Gene. For each patient group with the same affected gene, they created a profile, which Knaus said requires about 10 images per group. They found that the software was able to discriminate between patients with different gene defects with a mean accuracy of 51 percent, whereas trained physicians were able to distinguish them with a mean accuracy of 40 percent. This is better than getting the right answer by chance, which would happen 20 percent of the time. However, there was a high error rate for telling some groups apart, Knaus said, because their facial features are too similar.
In an unrelated study, presented on a poster, researchers at Magdeburg University in Germany looked at how well the FDNA software could distinguish between patients with RASopathies who had germline mutations in one of five different genes in the RAS-MAPK pathway. Here, the program could pick out the correct gene with a mean accuracy of 61 percent, compared with 20 percent by chance.
Other research groups presenting posters at the meeting tested the ability of the Face2Gene tool to aid with diagnosis, receiving different results. A team from the University of Siena in Italy, for example, had the software analyze photos from more than 440 patient cases, among them 60 with confirmed diagnoses. For a third of the confirmed cases, it ranked the correct diagnosis first, and for almost half the cases, it was among the first 10 suggestions. Part of the problem, the researchers wrote, is that the FDNA database does not contain profiles for several syndromes yet.
Meanwhile, a group from the Argon Institute for Health Research in Zaragoza, Spain, uploaded photos from 91 patients with developmental delay or intellectual disability to Face2Gene, 21 of them with molecular diagnoses, of which the software recognized seven. The group is waiting for molecular diagnoses for the remaining 70 patients and wrote that it "will consider Face2Gene a useful tool if in at least 10 percent of patients, one of the suggested syndrome matches does coincide with the molecular diagnosis of the patient."
Another group, led by researchers at Jena University Hospital in Germany, assessed how well the FDNA software could discriminate between patients with Emanuel syndrome; patients with Pallister-Killian syndrome, who have similar cytogenetic features; patients with other syndromes; and unaffected controls. They found that the tool was able to pick out the correct syndrome with a mean accuracy of almost 90 percent, more than three times better than by chance.
A team from the Technion-Israel Institute of Technology compared the performance of FDNA's algorithm against four geneticists for finding matches in pairs of photos from diagnosed, undiagnosed, and control patients and found that the software performed slightly better. "Future applications of this technology could complement next-generation sequencing in undiagnosed patients and lead to the discovery of rare novel syndromes," they wrote.
Finally, researchers from the University of Rennes in France and their colleagues applied the Face2Gene software to patients with autism spectrum disorder (ASD). For their study, they had clinical geneticists and the software analyze photos of 79 patients with ASD and intellectual disability. In total, the study identified 10 genetic syndromes in these patients that were confirmed molecularly. Of these, five were initially suspected by the geneticists and five by the Face2Gene software, of which just three overlapped. Interestingly, the software was unable to generate a specific facial profile for ASD, which the researchers said highlights the clinical heterogeneity of the disease.
In the meantime, researchers at the University of Oxford in the UK have launched a research project, called Minerva and Me, that aims to train software on photos of patients with confirmed diagnoses, an approach that appears to be similar to FDNA's. The project is led by Oxford researcher Christoffer Nellåker. Three years ago, his group published a paper in eLife in which the researchers described an algorithm that extracts phenotypic information from facial photographs and uses machine learning to generate models of facial characteristics for 91 disorders. Like FDNA's software, it provides a list of possible diagnoses for an individual patient.