An artificial intelligence algorithm called DeepGestalt has been developed that makes more accurate predictions than experts in detecting genetic diseases.
Many genetic diseases bring with them a distinctive facial phenotype. (One of the best-known examples of this condition is Down Syndrome.) Many genetic diseases are not easily diagnosed by clinicians because they are extremely rare. That’s why people with a genetic condition (and their families, of course) have to go through quite a long and traumatic diagnostic process before they realize what the problem is.
Wanting to find a solution to this problem, the genomics and artificial intelligence company FDNA began to identify, organize and analyze complex physiological data in humans. The company has created a facial analysis system they call DeepGestalt. By looking at facial images, DeepGestalt can detect genetic conditions much more accurately than doctors. The results of the FDNA study have been published in Nature Medicine.
More successful than experts
The company used a dataset of 500,000 face photos of 10,000 people selected from the internet to train the algorithm. Next, they mixed photos of people with a single genetic disease and people with multiple genetic diseases, and tested how well the system could diagnose. In this way, two tests were performed. Cornelia de Lange Syndrome was present in the first test and Angelman Syndrome was present in the other test. Both of these are developmental disorders that bring with them cognitive and motor impairment. As a result, DeepGestalt predicted more than 90 percent of the time in both tests. For experts, this rate was around 70-75%.
Next, DeepGestalt was shown photos of people with Noonan Syndrome (a syndrome with varying effects depending on which of five different genes are mutated) to test whether it could tell the difference between people with the same disease but different genotypes. This time, DeepGestalt was only able to achieve 65 percent accuracy. The final test was to diagnose hundreds of photos of people with 216 different diseases. In this test, the system achieved 90 percent accuracy.
Artificial intelligence cannot teach humans
This algorithm works by dividing the face into several regions and evaluating how well each region matches each syndrome. The system then aggregates and evaluates all parts of the face to determine which syndrome is appropriate for the person in the photograph. According to the authors of the study, like many artificial intelligence systems, DeepGestalt cannot provide accurate information and predictions about facial features that affect its classification. So it’s like a black box. He can pass the experts in genetic diagnosis based on the phenotype, but he cannot teach the experts how to do it.
Source: Ars Technica
Genome science and artificial intelligence company FDNA has developed an algorithm called DeepGestalt that diagnoses genetic diseases by looking at photos. A dataset of 500,000 photos taken from the internet was used to train the algorithm. In one of the various tests applied, it was seen that artificial intelligence achieved 90 percent accuracy when diagnosing a genetic disease, while this rate was around 70-75% for experts. DeepGestalt works by dividing the face into different regions and evaluating how much each region fits with which syndrome. Then the face is evaluated as a whole and diagnosed.