“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,” said Yaron Gurovich, chief technology officer at FDNA, an artificial intelligence and precision medicine company, who led the research.
This opens the door for future research and applications, and the identification of new genetic syndromes, he added.
But with facial images being easily accessible, this could lead to payers and employers potentially analyzing facial images and discriminating against individuals who have pre-existing conditions or developing medical complications, the authors warned.
Gurovich and his team trained DeepGestalt, a deep learning algorithm, by using 17,000 facial images of patients from a database of patients diagnosed with over 200 distinct genetic syndromes.
The team found that the AI technology outperformed clinicians in two separate sets of tests to identify a target syndrome among 502 chosen images. In each test, the AI proposed a list of potential syndromes and identified the correct syndrome in its top 10 suggestions 91% of the time.
“We showed that this system can be used in clinical settings,” Gurovich said of the results.
The technology works by applying the deep learning algorithm to the facial characteristics of the image provided, then producing a list of possible syndromes.
It does not explain which facial features led to its prediction, the study said. To help the researchers better understand, the technology produces a heat map visualization looking at what regions of the face contributed to the classification of diseases, explained Gurovich.
All the images used in the trials were from patients already diagnosed with a condition; the technology didn’t identify whether each patient had a genetic disorder, but identified possible disorders that had already been diagnosed.
One difficulty, mentioned Gurovich, is that the performance of AI system is hard to measure. “The reason it is hard is because there are not enough publicly available benchmarks,” he said.
Jorge Cardoso, senior lecturer in artificial medical intelligence at the school of biomedical engineering and imaging sciences at King’s College London, described the technology as “very interesting.”
He added, in an email to CNN, that “the collection of increasingly large and well curated medical datasets has enabled AI tools to predict genetic mutations from imaging phenotypes reducing the burden of healthcare systems and improving the way we care for patients.” Phenotypes are observable characteristics.
“While several limitations still need to be addressed to ensure the proposed algorithms are robust in the hospital environment, clinically accurate, and applicable to different age groups and ethnic populations, the potential of AI in healthcare is immense,” said Cardoso, who was not involved in the research.
Peter McOwan, professor of computer science at Queen Mary University of London, said in an email to CNN: “This is yet another fantastic potentially life changing application of AI tech. When we see so many negative stories round AI technology it’s good to be reminded of the real benefits it can provide to humanity.”