Getting tested for Alzheimer’s disease may one day be as simple as having your vision checked.
Getting tested for Alzheimer’s disease may one day be as simple as having your vision checked.
Retispec has developed an artificial-intelligence algorithm that it says can analyze results from an eye scanner and detect signs of Alzheimer’s up to 20 years before symptoms develop. The tool is part of broader work by startups and researchers to use AI to uncover the mysteries of the disease that afflicts more than seven million Americans.
hello! You are reading a premium article
Retispec has developed an artificial-intelligence algorithm that it says can analyze results from an eye scanner and detect signs of Alzheimer’s up to 20 years before symptoms develop. The tool is part of broader work by startups and researchers to use AI to uncover the mysteries of the disease that afflicts more than seven million Americans.
For years, people have studied the individual symptoms of Alzheimer’s, including brain inflammation and neurodegeneration, but the exact causes of the disease remain unknown. Researchers say AI could usher in a new era in diagnosing neurological diseases that are difficult to identify, let alone treat.
“There’s still a lot we don’t fundamentally understand about the brain and how it works,” said Eliav Shaked, co-founder of Toronto-based RetiSpec. “The power in AI is that it’s able to connect the dots. “Can help.”
Another company, Sacramento, Calif.-based NeuroVision, aims to use machine learning to develop retinal scans and blood tests to identify people at risk of developing Alzheimer’s and other forms of dementia. The company’s AI model analyzes eye scans for anomalies, such as buildup of certain proteins or bent-shaped blood vessels, which are associated with Alzheimer’s, said Steven Verduner, co-founder of NeuroVision.
It may be difficult for people to recognize such signals in the scan. Many scans contain dark areas, and plaque deposits may be very small. The human eye can’t distinguish them very well, Verduner said.
“The algorithm works better,” he said.
At the University of Arizona College of Medicine in Tucson, Rui Chang, associate professor of neurology, created an AI model that aims to identify genetic triggers associated with Alzheimer’s. Chang said the traditional approach researchers follow is too slow.
“It’s like looking at the forest one tree at a time,” he said.
AI can absorb entire forests of information at once and find patterns that people cannot. The model took two months to identify 6,000 gene targets that, if destroyed or suppressed, could change the way Alzheimer’s develops. Chang said the device cut a decade off his research.
Chang founded a company called Path-Biotech, which will begin clinical trials next year based on his AI research.
Alzheimer’s was the sixth leading cause of death in the US in 2021, not including COVID-19.
The Food and Drug Administration in July approved a drug, Lecambi, that removes amyloid, the sticky plaque that collects in the brains of Alzheimer’s patients. But existing techniques to identify the disease are expensive and difficult.
People with symptoms can have a spinal tap or PET scan to see if they have high levels of amyloid and tangled threads of the protein tau, which is also commonly found in Alzheimer’s patients. The scans are very accurate, making them the gold standard of diagnosis, said Katherine Bornbaum, chief business officer of RetiSpec. Compared to autopsy results – still the only way to tell definitively whether a patient had Alzheimer’s at the time of his or her death – PET scans have a diagnostic success rate closer to 90%.
But the machines are not widely available, the scans are expensive, and it can take weeks to get a diagnosis. Insurers do not routinely cover scans and they can cost around $6,000.
Artificial intelligence technologies can speed up the process of getting a diagnosis and make it cheaper. For example, Retispec’s AI reads scans from a camera that can be connected to machines already available in most optometrists’ offices. The camera measures a wider range of the spectrum than the human eye, allowing the AI to detect unique optical signatures that match the presence of amyloid in the brain. The model, which returns results instantly, was 80% accurate in detecting such signatures in a recent study of 271 patients.
Matt Laming, a research fellow at Massachusetts General Hospital, said AI tools in medical research can perform well in clinical trials, but fare poorly in difficult real-life situations.
“Biotech AI models are finicky,” he said.
AI learns better from massive amounts of data, Laming said. For example, AI models like ChatGPT are good at analyzing and copying writing, because they learn from text collected on the Internet.
Medical data is comparatively scarce and proprietary. This means AI in biotech has a more limited sample to learn from and its results may be more easily thrown off by the wider variation in cases encountered in the clinic compared to more controlled laboratory settings.
“When it comes to AI fundamentally changing the way we do medicine, I don’t think that’s going to happen,” Laming said.
Chang, at the University of Arizona, said he has attempted to overcome this problem by using mathematical models that reduce errors and improve prediction accuracy. RetiSpec said the company has obtained samples from 14 research partners, from whom it collects samples from racially and socio-economically diverse communities. NeuroVision said it took samples from different data sets and tested them against others to minimize errors.
“The most important thing we’ve done is to make sure the AI isn’t harmed by going in and out of the garbage,” said RetiSpec’s Shaked.
Write to Vipal Monga at [email protected]