By Steve Graff
Editor’s Note: This article was originally published by Penn Medicine News.
When the biotech company Biogen stopped a clinical trial investigating a drug known as aducanumab to slow Alzheimer’s, disappointment swept through the neurology community. “This one hurts,” one neurologist told CNN in March. “Alzheimer’s patients just can’t get a break.” The drug, which reduces amyloid, a nefarious protein believed to drive the disease, had slowed cognitive decline in animals. But studies in humans showed the intervention was too late. Damage—likely from the early buildup of the protein—had already been done, and patients’ outcomes weren’t improving as the trial progressed. Other promising drugs remain in the pipeline, but that trial—which included several Penn patients—was yet another reminder of the staggering complexity of Alzheimer’s.
“This a disease that can start up to 20 years before any symptoms become obvious,” said Yong Fan, PhD, an assistant professor of Radiology in the Center for Biomedical Imaging Computing and Analytics at Penn. “So if you could predict the disease earlier, you would have a better chance to cure it with drugs.”
Fortunately, the field of detection has made significant strides over the years to help get there. Researchers are consistently finding better ways to track rogue proteins and predict the progression of cognitive decline with biomarkers, neuroimaging, and more recently and arguably most impressively, the rapidly advancing technology touching nearly every part of medicine today: artificial intelligence.
Sophisticated, computational algorithms may soon be helping physicians determine when and which patients will advance to Alzheimer’s with greater accuracy—and in a much faster timeframe than before.
“Two or three seconds,” Fan said. “That’s a how long it would take our deep learning algorithms to spit out a prediction score for Alzheimer’s using just an MRI [magnetic resonance imaging] scan.”
Deep learning is a type of AI where a computer is “taught” and learns based on information it’s given, so it can make an informed decision, or, in this case, an informed prediction in patients with mild cognitive impairment, or MCI, a small but noticeable change in memory and thinking, like forgetting conversations or misplacing car keys. The condition affects nearly 20 percent of adults over the age of 65.
Up to 20 percent of them will eventually develop Alzheimer’s, but the problem is, physicians have a difficult time telling who those patients are and when it will happen to them. That limits their ability to treat the disease, as well as researchers investigating new ways to slow or stop it.
In a recent study in Alzheimer’s & Dementia, Fan, an expert in AI and deep learning in several disease types, Hongming Li, PhD, a research associate of Radiology, and their co-author David A. Wolk, MD, co-director of the Penn Memory Center, put their deep learning methods to the test using the MRI scans of over 2,100 patients who were cognitively “normal,” had Alzheimer’s, or MCI from the Alzheimer’s Disease Neuroimaging Initiative and the Australian Imaging Biomarkers Lifestyle Study of Aging.
The researchers programmed the AI to “see subtle changes and patterns that we would miss,” Fan said, in the brain’s hippocampus. It wasn’t looking for proteins like amyloid or tau or volume of the hippocampus, which are frequently used as markers of cognitive decline. Rather, it was looking for changes in brain tissue texture over time, from diagnosis of MCI to dementia.
The AI essentially studied the MRI images to become an expert predictor of who will develop Alzheimer’s disease – and when – by learning the difference between the progression of a “normal” brain versus a diseased one. They used 800 patients to build the deep learning model, and the rest to validate it.
In the end, it performed better than prediction models based on anatomical features, such as hippocampal volume and shape, with a concordance index of 0.762 in one cohort and 0.781 in another. That score shows how accurate an algorithm or a biomarker, for example, is at predicting an event over a period of time—in this case, dementia. “Anything over 0.7 is great,” Fan said.
The researchers later factored in clinical measures, such as age, education, mutation status (APOE4 is associated with an increased risk of Alzheimer’s), and cognition scores, into the prediction model, and produced an even better index: 0.864.
“Even with just a scan of a patient diagnosed with MCI at baseline, the computer would be able to predict if and when they progress to Alzheimer’ disease,” Fan said. “And it can do it really fast.”
Developing and implementing software that can predict a risk for Alzheimer’s in less than a few seconds would be practice changing. Today, in the clinic, real-time prediction using MRI scans hasn’t become a reality. It’s too time consuming: Analyzing a raw image using conventional, computational methods can take up to 10 hours or more.
The deep learning approach is also cost effective, an important benefit of a technology that’s still being investigated and to help identify patients for future drug trials.
“We focused on MRI scans, because MRI data is readily available and much cheaper than PET scans,” Fan said. “In the clinical research setting, if you want to enroll a subset of patients who will convert to Alzheimer’s to test a drug’s effects, at this moment, you would have to use an expensive PET scan.”
The next step is to go back even further in time, before a patient develops MCI, since Alzheimer’s can start long before it affects a person’s memory or thinking.
“We are working to develop a new prediction model to predict any cognitive decline, not just Alzheimer’s dementia,” Fan said. “Because that might be too late.”