Researchers from the University of Massachusetts Amherst have received a two-year NIH grant for $278,118 to develop new deep-learning models for the early prediction of Alzheimer’s using clinical data, including brain MRIs taken in real-world settings.
The ultimate goal of this research is to enable earlier detection of Alzheimer’s—ideally two years or more before the onset of symptoms—and identify patient populations at risk for developing the condition using MRI data, allowing researchers to test interventions and medications that interrupt the course of the disease. To create these predictive models, the researchers will use multimodal clinical data, including brain MRIs.
“This research brings us closer to putting people in clinical trials at a point where the brain biology is still intact and something can be done,” says Madalina (Ina) Fiterau, PhD, assistant professor in the Manning College of Information and Computer Sciences at UMass Amherst and principal investigator and project leader of the study. “Sixty percent of a patient’s brain matter disappears by the time of diagnosis, and at that stage it’s irretrievable. What we would like to do is identify those changes early, at least two years before onset, and then, based on that, figure out which treatments work.”
According to the study’s other principal investigator, Joyita Dutta, PhD, investigator, associate professor of biomedical engineering, “We would not have been able to say this three years back, but now that many new drug candidates are emerging, we are at the point where forecasting techniques can actually be deployed to identify potential subjects for a disease-modifying therapy.”
Previous research has aimed to create deep-learning models for predicting degenerative disease using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Unfortunately, this poses a challenge for making the findings generalizable since these ADNI-based models use engineered data, notes Fiterau. For instance, instead of incorporating actual images of the cerebral cortex, other studies use software to extract average cortical thickness.
“ADNI contains a vast set of specialized features that are extracted from the brain images that require a lot of feature engineering and domain expertise,” explains Fiterau. “Data collected in the wild is not going to have the specialized features. You’re going to have the MRI scans, but no annotation,” she says. “That’s the purpose of the grant: figuring out how to take a model that’s been trained on this specialized, carefully curated data set and see what its performance can be on real data collected in the wild.”
This research is important because it enables predictive algorithms to use standard MRIs instead of requiring specialized data, points out Dutta. “I work extensively with PET scans for my research but not every clinic collects PET images of Alzheimer’s patients,” she says. “At the same time, MRI tends to be the go-to imaging modality for individuals with neurological complaints. However, clinically available MRI scans are often conducted using protocols that are different from ADNI. Predictive models, therefore, need to be generalizable to ‘data collected in the wild’ in order to be practically useful.”
The research team will use deep learning to extract features from standard brain MRIs that can stand in as proxies for the specialized features from the ADNI dataset. There are some key regions that researchers know to be affected by Alzheimer’s—the hippocampus, cerebral cortex, and fluid-filled ventricle cavities—so the model will be trained to place a higher weight on these regions.
The study also aims to overcome model biases due to the demographic gaps in the ADNI data, namely the underrepresentation of minorities and the overrepresentation of highly educated individuals. In this dataset, 93% of the participants are white and 61% have 16 years of education or more (the U.S. average is 14 years).