In an initial study, Canary analyzed speech recordings with AI and identified early stage Alzheimer’s with 96 percent accuracy.
Dave Arnold retired in his 60s and began spending time volunteering in local schools. But then he started misplacing items, forgetting appointments and losing his sense of direction. Eventually he was diagnosed with early stage Alzheimer’s.
“Hearing the diagnosis made me very emotional and tearful,” he said. “I immediately thought of all my mom had experienced.” His mother suffered with the condition for years before passing away. Over the last year, Arnold has worked for the Alzheimer’s Association as one of its early stage advisors, sharing his insights to help others in the initial stages of the disease.
Arnold was diagnosed sooner than many others. It's important to find out early, when interventions can make the most difference. One promising avenue is looking at how people talk. Research has shown that Alzheimer’s affects a part of the brain that controls speech, resulting in small changes before people show other signs of the disease.
Now, Canary Speech, a company based in Utah, is using AI to examine elements like the pitch of a person’s voice and their pauses. In an initial study, Canary analyzed speech recordings with AI and identified early stage Alzheimer’s with 96 percent accuracy.
Developing the AI model
Canary Speech’s CEO, Henry O’Connell, met cofounder Jeff Adams about 40 years before they started the company. Back when they first crossed paths, they were both living in Bethesda, Maryland; O’Connell was a research fellow at the National Institutes of Health studying rare neurological diseases, while Adams was working to decode spy messages. Later on, Adams would specialize in building mathematical models to analyze speech and sound as a team leader in developing Amazon's Alexa.
It wasn't until 2015 that they decided to make use of the fit between their backgrounds. ““We established Canary Speech in 2017 to build a product that could be used in multiple languages in clinical environments,” O'Connell says.
The need is growing. About 55 million people worldwide currently live with Alzheimer’s, a number that is expected to double by 2050. Some scientists think the disease results from a buildup of plaque in the brain. It causes mild memory loss at first and, over time, this issue get worse while other symptoms, such as disorientation and hallucinations, can develop. Treatment to manage the disease is more effective in the earlier stages, but detection is difficult since mild symptoms are often attributed to the normal aging process.
O’Connell and Adams specialize in the complex ways that Alzheimer’s effects how people speak. Using AI, their mathematical model analyzes 15 million data points every minute, focusing on certain features of speech such as pitch, pauses and elongation of words. It also pays attention to how the vibrations of vocal cords change in different stages of the disease.
To create their model, the team used a type of machine learning called deep neural nets, which looks at multiple layers of data - in this case, the multiple features of a person’s speech patterns.
“Deep neural nets allow us to look at much, much larger data sets built out of millions of elements,” O’Connell explained. “Through machine learning and AI, we’ve identified features that are very sensitive to an Alzheimer’s patient versus [people without the disease] and also very sensitive to mild cognitive impairment, early stage and moderate Alzheimer's.” Based on their learnings, Canary is able to classify the disease stage very quickly, O’Connell said.
“When we’re listening to sublanguage elements, we’re really analyzing the direct result of changes in the brain in the physical body,” O’Connell said. “The brain controls your vocal cords: how fast they vibrate, the expansion of them, the contraction.” These factors, along with where people put their tongues when talking, function subconsciously and result in subtle changes in the sounds of speech.
Further testing is needed
In an initial trial, Canary analyzed speech recordings from phone calls to a large U.S. health insurer. They looked at the audio recordings of 651 policyholders who had early stage Alzheimer’s and 1018 who did not have the condition, aiming for a representative sample of age, gender and race. They used this data to create their first diagnostic model and found that it was 96 percent accurate in identifying Alzheimer’s.
Christian Herff, an assistant professor of neuroscience at Maastricht University in the Netherlands, praised this approach while adding that further testing is needed to assess its effectiveness.
“I think the general idea of identifying increased risk for cognitive impairment based on speech characteristics is very feasible, particularly when change in a user’s voice is monitored, for example, by recording speech every year,” Herff said. He noted that this can only be a first indication, not a full diagnosis. The accuracy still needs to be validated in studies that follows individuals over a period of time, he said.
Toby Walsh, a professor of artificial intelligence at the University of New South Wales, also thinks Canary’s tool has potential but highlights that Canary could diagnose some people who don’t really have the disease. “This is an interesting and promising application of AI,” he said, “but these tools need to be used carefully. Imagine the anxiety of being misdiagnosed with Alzheimer’s.”
As with many other AI tools, privacy and bias are additional issues to monitor closely, Walsh said.
Other languages
A related issue is that not everyone is fluent in English. Mahnaz Arvaneh, a senior lecturer in automatic control and systems engineering at the University of Sheffield, said this could be a blind spot.
“The system may not be very accurate for those who have English as their second language as their speaking patterns would be different, and any issue might be because of language deficiency rather than cognitive issues,” Arvaneh said.
The team is expanding to multiple languages starting with Japanese and Spanish. The elements of the model that make up the algorithm are very similar, but they need to be validated and retrained in a different language, which will require access to more data.
Recently, Canary analyzed the phone calls of 233 Japanese patients who had mild cognitive impairment and 704 healthy people. Using an English model they were able to identify the Japanese patients who had mild cognitive impairment with 78 percent accuracy. They also developed a model in Japanese that was 45 percent accurate, and they’re continuing to train it with more data.
The future
Canary is using their model to look at other diseases like Huntington’s and Parkinson’s. They’re also collaborating with pharmaceuticals to validate potential therapies for Alzheimer’s. By looking at speech patterns over time, Canary can get an indication of how well these drugs are working.
Dave Arnold and his wife dance at his nephew’s wedding in Rochester, New York, ten years ago, before his Alzheimer's diagnosis.
Dave Arnold
Ultimately, they want to integrate their tool into everyday life. “We want it to be used in a smartphone, or a teleconference call so that individuals could be examined in their home,” O’Connell said. “We could follow them over time and work with clinical teams and hospitals to improve the evaluation of patients and contribute towards an accurate diagnosis.”
Arnold, the patient with early stage Alzheimer’s, sees great promise. “The process of getting a diagnosis is already filled with so much anxiety,” he said. “Anything that can be done to make it easier and less stressful would be a good thing, as long as it’s proven accurate.”
Avalanche rescue dogs train to find and dig out people buried in snow slides
By the time he is three, Lume, like Huckleberry, will be a fully trained avy pup and will join seven other avy dogs on Breckenridge ski patrol team. Some of the team members, both human and canine, are also certified to work with Colorado Rapid Avalanche Deployment, a coordinated response team that works with the Summit County Sheriff’s office for avalanche emergencies outside of the ski slopes’ boundaries.
There have been 19 avalanche deaths in the U.S. this season, according to avalanche.org, which tracks slides; eight in Colorado. During the entirety of last season there were 17. Avalanche season runs from November through June, but avalanches can occur year-round.
High tech and high stakes
Complementing avy dogs’ ability to smell people buried in a slide, avalanche detection, rescue and recovery is becoming increasingly high tech. There are transceivers, signal locators, ground scanners and drones, which are considered “games changers” by many in avalanche rescue and recovery
For a person buried in an avalanche, the chance of survival plummets after 20 minutes, so every moment counts.
A drone can provide thermal imaging of objects caught in a slide; what looks like a rock from far away might be a human with a heat signature. Transceivers, also known as beacons, send a signal from an avalanche victim to a companion. Signal locators, like RECCO reflectors which are often sewn directly into gear, can echo back a radar signal sent by a detector; most ski resorts have RECCO detector units.
Research suggests that Ground Penetrating Radar (GPR), an electromagnetic tool used by geophysicists to pull images from inside the ground, could be used to locate an avalanche victim. A new study from the Department of Energy’s Sandia National Laboratories suggests that a computer program developed to pinpoint the source of a chemical or biological terrorist attack could also be used to find someone submerged in an avalanche. The search algorithm allows for small robots (described as cockroach-sized) to “swarm” a search area. Researchers say that this distributed optimization algorithm can help find avalanche victims four times faster than current search mechanisms. For a person buried in an avalanche, the chance of survival plummets after 20 minutes, so every moment counts.
An avy dog in training is picking up scent
Sarah McLear
While rescue gear has been evolving, predicting when a slab will fall remains an emerging science — kind of where weather forecasting science was in the 1980s. Avalanche forecasting still relies on documenting avalanches by going out and looking,” says Ethan Greene, director of the Colorado Avalanche Information Center (CAIC). “So if there's a big snowstorm, and as you might remember, most avalanches happened during snowstorms, we could have 10,000 avalanches that release and we document 50,” says Greene. “Avalanche forecasting is essentially pattern recognition,” he adds--and understanding the layering structure of snow.
However, determining where the hazards lie can be tricky. While a dense layer of snow over a softer, weaker layer may be a recipe for an avalanche, there’s so much variability in snowpack that no one formula can predict the trigger. Further, observing and measuring snow at a single point may not be representative of all nearby slopes. Finally, there’s not enough historical data to help avalanche scientists create better prediction models.
That, however, may be changing.
Last year, an international group of researchers created computer simulations of snow cover using 16 years of meteorological data to forecast avalanche hazards, publishing their research in Cold Regions Science and Technology. They believe their models, which categorize different kinds of avalanches, can support forecasting and determine whether the avalanche is natural (caused by temperature changes, wind, additional snowfall) or artificial (triggered by a human or animal).
With smell receptors ranging from 800 million for an average dog, to 4 billion for scent hounds, canines remain key to finding people caught in slides.
With data from two sites in British Columbia and one in Switzerland, researchers built computer simulations of five different avalanche types. “In terms of real time avalanche forecasting, this has potential to fill in a lot of data gaps, where we don't have field observations of what the snow looks like,” says Simon Horton, a postdoctoral fellow with the Simon Fraser University Centre for Natural Hazards Research and a forecaster with Avalanche Canada, who participated in the study. While complex models that simulate snowpack layers have been around for a few decades, they weren’t easy to apply until recently. “It's been difficult to find out how to apply that to actual decision-making and improving safety,” says Horton. If you can derive avalanche problem types from simulated snowpack properties, he says, you’ll learn “a lot about how you want to manage that risk.”
The five categories include “new snow,” which is unstable and slides down the slope, “wet snow,” when rain or heat makes it liquidly, as well as “wind-drifted snow,” “persistent weak layers” and “old snow.” “That's when there's some type of deeply buried weak layer in the snow that releases without any real change in the weather,” Horton explains. “These ones tend to cause the most accidents.” One step by a person on that structurally weak layer of snow will cause a slide. Horton is hopeful that computer simulations of avalanche types can be used by scientists in different snow climates to help predict hazard levels.
Greene is doubtful. “If you have six slopes that are lined up next to each other, and you're going to try to predict which one avalanches and the exact dimensions and what time, that's going to be really hard to do. And I think it's going to be a long time before we're able to do that,” says Greene.
What both researchers do agree on, though, is that what avalanche prediction really needs is better imagery through satellite detection. “Just being able to count the number of avalanches that are out there will have a huge impact on what we do,” Greene says. “[Satellites] will change what we do, dramatically.” In a 2022 paper, scientists at the University of Aberdeen in England used satellites to study two deadly Himalayan avalanches. The imaging helped them determine that sediment from a 2016 ice avalanche plus subsequent snow avalanches contributed to the 2021 avalanche that caused a flash flood, killing over 200 people. The researchers say that understanding the avalanches characteristics through satellite imagery can inform them how one such event increases the magnitude of another in the same area.
Avy dogs trainers hide in dug-out holes in the snow, teaching the dogs to find buried victims
Sarah McLear
Lifesaving combo: human tech and Mother Nature’s gear
Even as avalanche forecasting evolves, dogs with their built-in rescue mechanisms will remain invaluable. With smell receptors ranging from 800 million for an average dog, to 4 billion for scent hounds, canines remain key to finding people caught in slides. (Humans in comparison, have a meager 12 million.) A new study published in the Journal of Neuroscience revealed that in dogs smell and vision are connected in the brain, which has not been found in other animals. “They can detect the smell of their owner's fingerprints on a glass slide six weeks after they touched it,” says Nicholas Dodman, professor emeritus at Cummings School of Veterinary Medicine at Tufts University. “And they can track from a boat where a box filled with meat was buried in the water, 100 feet below,” says Dodman, who is also co-founder and president of the Center for Canine Behavior Studies.
Another recent study from Queens College in Belfast, United Kingdom, further confirms that dogs can smell when humans are stressed. They can also detect the smell of a person’s breath and the smell of the skin cells of a deceased person.
The emerging avalanche-predicting human-made tech and the incredible nature-made tech of dogs’ olfactory talents is the lifesaving “equipment” that Leffler believes in. Even when human-made technology develops further, it will be most efficient when used together with the millions of dogs’ smell receptors, Leffler believes. “It is a combination of technology and the avalanche dog that will always be effective in finding an avalanche victim.”