Artificial intelligence can become a powerful ally of science and health with the technology called DeepGestalt, which can accurately identify some rare genetic diseases by just analyzing a patient’s face. The study on the software was published last week in the journal Nature Medicine.
According to research, 8% of the population has significant genetic diseases and many of them can be identified because of their recognizable facial features. An example of this is Angelman syndrome, which affects the nervous system causing physical aspects such as wide mouth and spaced teeth, strabismus or a protruding tongue.
In his three essays, DeepGestalt was able to identify a series of syndromes just like that, looking at patients’ faces. The basis for successful testing in such a challenging field and with little data available were “cutting-edge algorithms such as deep learning,” said Yaron Gurovich, who, in addition to leading the research, is also FDNA’s chief technology officer , an artificial intelligence and precision medicine company.
Gurovich also argued that the success of the trials also opens space for future research and applications, as well as the identification of new yet unidentified syndromes. In the study, the authors also demonstrated concern with face analyzes that could discriminate against people who have preexisting conditions or who develop medical complications.
The team trained DeepGestalt using 17,000 facial images taken from a database of patients diagnosed with more than 200 different genetic disorders. With this, it was discovered that AI technology has outperformed clinicians in two distinct sets of tests. In one, it was necessary to identify a specific syndrome among 502 selected images.
At each test, the AI listed the possible syndromes and identified the main and correct in their top 10 suggestions 91% of the time. In the other type of test, it was necessary to identify different genetic subtypes of Noonan syndrome, which presents a series of distinct characteristics and health problems, such as heart defects. At this stage, the algorithm had a success rate of 64%.
For comparison, clinicians who examined the images of patients with Noonan syndrome were able to identify the disorder only 20% of the time. “We showed that this system can be used in medical settings,” Gurovich said of the results, even without explaining exactly what facial features DeepGestalt took into account in its predictions.
However, to help researchers better understand, the technology recreates a heat-based image map, analyzing which regions of the face contributed to the classification of diseases, according to Gurovich. It is worth remembering that all images used in the trials were from patients who had already been diagnosed with one condition. In addition, the AI did not analyze whether each had a genetic disorder, but rather whether any of the possible ones had already been identified.