There’s no clinical test for Alzheimer’s disease, so physicians diagnose it by conducting assessments of patients’ cognitive decline. But it’s particularly difficult for them to identify mild cognitive impairment (MCI), an early stage of dementia when symptoms are less obvious. And it’s even harder to predict which MCI patients will develop Alzheimer’s disease (not all of them do).
So it makes sense that researchers might try to apply deep learning to this challenge. “There's a lot of interest in some type of test that would say, ‘This person will go on to develop Alzheimer’s and this person will not,’” says Pamela Greenwood, a psychology professor at George Mason University, in Fairfax, Va.
Recently, collaborators from Harvard University, Massachusetts General Hospital, and China’s Huazhong University of Science and Technology designed a program that combines fMRI brain scans with clinical data to make this prediction. They presented the work, which has not yet been published, in May at the IEEE International Conference on Communications in Kuala Lumpur, Malaysia.
“We try to find the disease at its very early stage,” says Quanzheng Li, principle investigator from Massachusetts General Hospital’s Center for Clinical Data Science. “A lot of people try to use traditional machine learning to do this, but the result is not that great because it's a very difficult problem to solve.”
After initial tests, they say their deep learning program, when paired with a special fMRI dataset, is about 20 percent more accurate than other classification methods using a more basic dataset. However, when those traditional classifiers also used the special dataset, they showed similar gains in accuracy.
Javier Escudero, a biomedical engineer at the University of Edinburgh, says that means the new program may not be all that much better than the old ones; it may have just been working with better data.
If that’s true, then other experts who want to apply deep learning to diagnosing Alzheimer’s disease may want to take a close look at the data they incorporate into their analyses. According to this latest work, fMRI scans that are processed to show relationships between areas of the brain provide a more nuanced view of the condition than those which merely record measurements over time.
For now, the Harvard-led team is among the first to try to combine fMRI scans and deep learning into a program that could predict an MCI patient’s chance of developing Alzheimer’s disease. The fMRI scans used in their analysis were taken when patients were at rest. As with any fMRI scans, they reveal where electrical signals are flashing in the brain and how these areas relate to one another.
The term for this relationship is functional connectivity, and it changes as patients develop MCI. This is because signals rely on the flow of oxygen to neurons, but the accumulation of tau proteins in Alzheimer’s disease patients strangles these neurons, causing regions of the brain to atrophy.
The group wanted to see if they could use these changes in functional connectivity to predict Alzheimer’s disease. They began with data from 93 MCI patients and 101 normal patients provided by the Alzheimer’s Disease Neuroimaging Initiative. Based on a time series of 130 fMRI measurements taken from each of 90 regions within participants’ brains, the researchers could tell where signals flashed over a period of time.
Next, in a crucial step, the group processed this dataset to create a secondary measurement of the strength of these signals in brain regions relative to each other. In other words, they constructed a map of functional connectivity that showed which areas and signals were most closely related to each other.
Lastly, the team built a deep learning program that could interpret the strength of these patterns, and, when combined with clinical data on age, gender, and genetic risk factors, predict whether a person would develop Alzheimer’s disease.
In the end, the team says, its program—using the specially processed dataset of functional connectivity—could predict whether the patients in their cohort would progress to Alzheimer’s disease with nearly 90 percent accuracy.
Li says the program is almost accurate enough to be helpful in a clinical setting. “When the method reaches about 90 percent, it’s very useful,” he says. “We are not there yet but we are very close.”
Even with laboratory techniques such as testing for too many proteins in cerebrospinal fluid, experts’ predictions of which MCI patients will develop Alzheimer’s disease are only about 65 percent accurate. That means some people go undiagnosed, while others worry needlessly about developing the disease.
But after reviewing a draft version of the Harvard team’s research, Dinggang Shen, who has done similar work on cognitive computing at the University of North Carolina at Chapel Hill, is skeptical.
“Nobody in the field can get to 80 to 90 percent,” he says. “That's impossible, just based on this method.” (The authors admit that there were several typos in an early draft shared with Shen, but insist that the accuracy is correct).
The results represent a roughly 20 percent improvement over other classifiers they tested, which used only time series fMRI data, and not the functional connectivity map. However, the accuracy of those other classifiers improved by about 16 percent when they also used the functional connectivity map.
The takeaway, Escudero says, should be that programs which assess functional connectivity, or the strengths of signal associations within the brain, are much better at predicting patients’ likelihood of developing Alzheimer’s disease than those which only measure brain signals over time. “It seems clear that the biggest gain comes from the consideration of the connectivity data,” he says.
This newest experiment is part of a wide range of efforts to apply deep learning or artificial intelligence to assist clinicians with complex decisions. Perhaps most famously, IBM’s Watson aims to guide doctors faced with mountains of medical records and many volumes of literature.
George Mason’s Greenwood provides important perspective, pointing out that because there’s still no cure for Alzheimer’s disease, the usefulness of any such prediction tool is limited for this particular disease. The new program would also need to undergo thorough peer review and far more testing before it could ever be used in a clinical diagnosis.