A new artificial intelligence (AI) technology that can identify rare genetic diseases through analyzing an image of a patient's face could help cut diagnosis times for rare diseases and provide more personalized care, according to a new study published online Jan. 7 in Nature Medicine.
The deep learning algorithms, developed by Boston-based startup FDNA Inc. and collectively called DeepGestalt, pinpoint recognizable characteristics and patterns in facial photos of patients to determine a possible set of genetic mutations.
Researchers led by Yaron Gurovich, chief technology officer of FDNA, trained DeepGestalt on a dataset of images curated through Face2Gene, FDNA’s facial phenotyping smartphone app. The dataset compiled over 17,000 facial images of patients diagnosed with 216 different genetic syndromes.
To help researchers better understand which facial features led to its prediction, the technology pinpoints what regions of the face determined its classification of diseases on a heat map.
In three clinical trials, DeeGestalt outperformed clinicians in accurately identifying genetic diseases from hundreds of facial images.
“The increased ability to describe phenotype in a standardized way opens the door to future research and applications, and the identification of new genetic syndromes,” Gurovich said in a prepared statement. “It demonstrates how one can successfully apply state of the art algorithms, such as deep learning, to a challenging field where the available data is small, unbalanced in terms of available patients per condition, and where the need to support a large amount of conditions is great.”
In the first two trials, researchers had the algorithms identify a target syndrome from 502 images and found that it identified the correct syndrome in its top 10 suggestions with 91 percent accuracy, according to study findings.
For a third trial, the algorithms identified genetic subtypes of Noonan Syndrome, which causes unseal facial features, bleeding problems and developmental defects. The algorithms achieved a 64 percent accuracy, which was substantially higher than the 20 percent of clinicians in previous studies who correctly identified the syndrome from facial images of patients, according to the researchers.
Bias and the future
Face2Gene, a facial phenotyping smartphone app released to health care providers by FDNA in 2014, is powered by DeepGestalt and currently has 150,000 images in its database. The startup has noted, however, that the tool is intended to be used as an aid and not for definitive diagnostic purposes.
Recently the app has been criticized for its ethnic bias in training data sets that contain mostly Caucasian faces, which FDNA plans to address.
“We know this problem needs to be addressed and as we move forward, we’re able to have less and less bias,” Gurovich told Nature.