An early diagnosis can make all the difference for patients with Alzheimer’s disease. Can the rapidly advancing technology of artificial intelligence (AI) catch what doctors might miss? Dr. Mallar Chakravarty, a computational neuroscientist at McGill University’s Douglas Mental Health University Institute, and his team have created an algorithm that hopes to go even further: predicting whether a patient will develop Alzheimer’s five years before any major symptoms appear.
- By plugging simple data such as a patient’s genetic makeup, blood test results and MRI scans into a machine, the team assesses whether the patient could develop Alzheimer’s within five years
- Their goal is to identify high-risk patients and encourage early prevention of Alzheimer’s through lifestyle changes before major symptoms appear
Being Patient spoke to Chakravarty about how machine learning is transforming the way conditions like Alzheimer’s could be diagnosed, how doctors analyze MRI data to assess Alzheimer’s risk and what AI may be capable of in 10 years.
Being Patient: Traditionally, when people want to determine whether they have Alzheimer’s, they can get a PET scan or a spinal tap, but most people don’t do that because those methods are not covered by insurance. How is machine learning changing the ways people could determine whether they have Alzheimer’s?
Dr. Chakravarty: Even before that happens, many people who believe they are at risk will go to their family doctor or a memory specialist to let them know that they’ve noticed changes in their cognition. For example, they can’t remember appointments or how to do simple tasks. Oftentimes, in elderly populations, this is diagnosed first and foremost as geriatric depression because these early signs share a lot of features with this disorder. Certainly what we see in the studies that we’ve done and when we do patient and subject recruitment is a lot of their general practitioners think they have geriatric depression. Slowly, after years have gone by, they realize that they may have memory impairment. Traditionally, most patients will undergo a battery of neuropsychological evaluations once they realize that there is cognitive decline. That doesn’t guarantee they’ll get an Alzheimer’s diagnosis. That just tells you if you’re below average on cognition, or if taken over time, it can reveal your cognition is in decline. It’s only after that, where there’s some suspected dementia, that a doctor will typically prescribe a PET scan or in other cases, a spinal tap, to look for amyloid or tau in the blood.
What we’re trying to do is to say OK, what’s really important is catching these people at the early stages of cognitive decline. We’ve developed an algorithm that takes some simple metrics, like results from tests that are taken in the clinic routinely, looking at whether people have certain copies of a genetic variant called ApoE and some MRI data. We take that data and feed it into our algorithm to predict whether you’re at a high risk for decline or not.
Being Patient: Are you primarily targeting the ApoE4 group, who are genetically predisposed to Alzheimer’s?
Dr. Chakravarty: No. The nice thing about these kinds of algorithms is that they’re not limited to just one specific bit of information. The algorithms take a composite picture of what the short-term change in brain anatomy looks like based on an MRI, or knowing if someone has certain copies of ApoE, and assess whether that makes their risk worse. Also, knowing that this is their brain signature and this is what their cognitive data looked like, does that make a difference at all? I think all those things together is what we’re trying to look at. We’re targeting anyone who is elderly and thinks they may have a risk.
Being Patient: We know that with a PET scan, you can use a dye to see where plaques and tangles are in the brain. What can you see with an MRI?
Dr. Chakravarty: With an MRI, you can get a really detailed analysis of brain anatomy. We know that neurodegeneration or brain atrophy plays a big role in the path towards Alzheimer’s disease. We look for genetic signatures [groups of genes with unique characteristics] on the individual subject level and say, OK, does this anatomical profile fit into a set of other profiles that we’ve already learned about and does it indicate that this individual is at risk for Alzheimer’s disease?
Being Patient: How far away are we from predicting whether or not healthy individuals’ cognition is declining as they age and what does that mean for the future?
Dr. Chakravarty: We’re seeing this incredible growth in technology that supports AI, so that’s very important. We’ve also adapted a lot of the algorithms we’re using from algorithms used by companies like Google or Facebook, where they were initially implemented for facial recognition software, for example. And in a way, that’s easier to do because Google or Facebook have the whole internet to draw from to do this testing. These algorithms will only learn what we tell them to learn, so they’re only as good as the problem that we teach them to solve. It won’t learn anything more or anything less. In our case, in terms of health care data, you’re somewhat limited by where the data comes from and how much data you have access to. One major limitation is that different clinicians or sites have various names for the same thing, so how do you standardize that across sites to train your algorithm? I think that is going to be a really important thing to overcome in the future. These types of methods hold a lot of promise though.
Being Patient: You mentioned a combination of MRI and looking at the shape of the hippocampus. What do we know about the shape of the hippocampus, the part of the brain that’s responsible for memory, and predicting cognitive decline? Could machine learning apply this information as well?
Dr. Chakravarty: This relates to two different studies that we did a few years apart. In this current study, where we’re using machine learning, we didn’t look at the shape of the hippocampus. But in previous studies that we’ve done, we’ve indicated that it seems like the shape of the hippocampus in individuals who otherwise aren’t showing much cognitive decline, is a really interesting biomarker that we can use to help predict decline. It might be more robust than more traditional markers, like volume. The reason for that is that the shape — at least as it appears on an MRI — may be more indicative of earlier changes that don’t show up as a volume metric change, indicating possible rewiring with the rest of the brain and so on. That is one factor that we could consider in the future, in combination with the artificial intelligence work that we do.
Being Patient: Are you testing people for various forms of dementia, like frontotemporal dementia?
Dr. Chakravarty: Not necessarily in my lab, but I think people are. More and more, I think the dementia world has moved away — and rightfully so — from trying to cure or treat the disease after we know it’s there. People have started to get into primary prevention, so prevention for people in their 50s or 60s, when we know that risk is starting to go up. The ApoE4 crowd, if we know that they’re at higher risk, we can target them, or anyone who appears to be at a high risk, and try to reverse any disease that could appear down the road.
Being Patient: Using AI for this purpose is relatively new. How do you know whether AI will help identify the people who are at a greater risk for Alzheimer’s?
Dr. Chakravarty: AI has been around for a very long time. The granddaddies of this field proposed these algorithms that we’re using now almost 40–45 years ago and at the time, they were kind of laughed out of computer science because people thought it would never work on a larger scale. The theory and the backbone has been there for a long time. The proof is in the pudding in a lot of ways, like with self-driving cars, Netflix, Facebook — they’ve all shown that AI can do a good job of predicting behaviors and preferences for individuals. There is proof and it’s very strong from different digital domains. In cancer research, AI is starting to make significant inroads because it’s one of the biomedical fields where we’ve done a much better job in terms of precision medicine. In metabolic disorders, it’s starting to happen too. I think one of the main differences between those diseases and the brain is that we know a lot more about those disorders than we do about normative brain functioning and brain disorders. As those two things come together, it will help the algorithms make better predictions.
Being Patient: These algorithms get better with more data, so what do you think we’ll know in 10 years that we don’t know today?
Dr. Chakravarty: What we know already is remarkable, compared to 10 years before, in neuroscience, clinical neuroscience and Alzheimer’s disease. There have been a lot of changes in the way we have approached Alzheimer’s disease in the last 10 years. I think with AI, modern medicine and hopefully, the confluence of the two, we have a long way to go. Technologies like brain imaging or genotyping with regular genetics and understanding lifestyle risk factors like depression, diet or obesity and how those things come together is where AI will play a significant role in the future and it’s a very exciting thing to think about.