Canary Speech builds AI technology to diagnose Alzheimer's disease
It is known that Alzheimer's disease is difficult to diagnose. As a rule, doctors use a combination of cognitive tests, brain imaging and behavioral observation, which can be expensive and time-consuming. But what if a quick voice sample, which is easy to take from a person at home, could help identify a patient with Alzheimer's disease?
Canary Speech is building exactly this technology. Using deep learning, its algorithms analyze short voice samples for signs of Alzheimer's disease and other conditions. Deep learning provider Syntiant recently announced a partnership with Canary Speech, which will allow Canary to move technology that is mainly used in doctors' offices and hospitals into a person's home via a medical device.
Although some studies have shown that deep learning methods using voice and other types of data are very accurate at classifying people with Alzheimer's disease and other conditions in laboratory settings, results in the real world may be different. Nevertheless, artificial intelligence and deep learning methods could become useful tools for making difficult diagnoses.
Most people think that Alzheimer's disease, the most common form of dementia, affects memory. But research shows that Alzheimer's can affect speech and language even in the earliest stages, before most symptoms become noticeable. While people usually cannot detect these subtle effects, a deep learning model trained on the voices of tens of thousands of people with and without these conditions can tell these factors apart.
“What interests you is what the central nervous system tells you, which is conveyed through speech production. This is what Canary Speech does — we analyze this dataset”
Until now, the algorithm has been cloud-based, but Canary's partnership with Syntiant enables a chip-based application that runs faster and has more memory and storage capacity.
The new technology is designed to be built into a wearable device and analyzes a 20- or 30-second speech sample in less than a second for conditions such as Alzheimer's disease, as well as anxiety, depression and even general energy levels. O'Connell said Canary's system is about 92.5% accurate when it comes to correctly distinguishing the voices of people with and without Alzheimer's disease.
Some research suggests that conditions such as depression and anxiety affect speech, and O'Connell said Canary is working on testing and improving the accuracy of its algorithms for detecting these conditions.
According to Frank Rudzicz, an associate professor of computer science at the University of Toronto and co-founder of Winterlight Labs, which makes a product similar to Canary Speech, other voice technologies have had similar success. In a 2016 study, Rudzicz and other researchers used simple machine learning methods to analyze the speech of people with and without Alzheimer's disease with an accuracy of about 81%.
“With deep learning, you simply feed raw data into these deep neural networks, and the deep neural networks automatically create their own internal representations”
Like all deep learning algorithms, this creates a "black box", which means it is impossible to know exactly which aspects of speech the algorithm focuses on. According to him, thanks to deep learning, the accuracy of these algorithms has exceeded 90%.
Previously, programmers used deep learning alongside medical brain imaging such as MRI. In studies, many of these methods are equally accurate, typically with accuracy above 90%.
In a study conducted in December 2021, programmers successfully trained the algorithm to distinguish not only the brains of cognitively normal people from those with Alzheimer's disease, but also people with mild cognitive impairment, which in many cases are early precursors to Alzheimer's, whose brains were either the same as healthy people's or more similar to those with Alzheimer's. Distinguishing these subtypes is especially important because not everyone with mild cognitive impairment develops Alzheimer's disease.
“We want to have methods for stratifying people along the Alzheimer's disease continuum. These are subjects who may develop Alzheimer's disease”
According to Dayan, identifying these patients as early as possible will likely be crucial for effectively treating their conditions. As a rule, scan-based deep learning has an equally high level of effectiveness, at least in classification studies conducted in the laboratory.
It is not yet known whether these technologies will be as effective in the real world, but they are likely to work well. Scientists said additional research is needed to know for sure.
Another cause for concern, according to Dayan, is the potential biases that recent research has shown AI may hide if the data on which the algorithm is trained lacks diversity. For example, Rudzicz said an algorithm trained on speech samples from people in Toronto may not work as well in rural areas.
O'Connell said Canary Speech's algorithm analyzes non-linguistic elements of speech, and that they have versions of the technology used in other countries, such as Japan and China, which are trained using data from native speakers: "We test our model and train it in this system, in this environment, to improve performance".
Although Canary's partnership with Syntiant could make real-time remote monitoring possible, O'Connell believes that a formal diagnosis should come from a doctor, and this technology serves as another tool for making a diagnosis. Dayan agrees: "I hope that in the coming years AI will help doctors, but will not completely replace them".
