By Deborah Borfitz
May 28, 2020 | A multidisciplinary team of researchers has developed a machine learning (ML) algorithm shown to be “slightly better” than 11 neurologists at diagnosing Alzheimer’s disease in a head-to-head comparison. The neurologists were all co-authors on a paper describing the feat in Brain.
Their interest in collaborating on the project has everything to do with the dearth of neurologists around the world, including places like Nebraska where only a handful of board-certified neurologists serve the entire state, says corresponding author Vijaya B. Kolachalama, Ph.D., assistant professor of medicine at Boston University School of Medicine (BUSM). Two of them are part of the study.
The predictive algorithm was given one specific task—to decide, based on an MRI scan of the brain if a person did or did not have Alzheimer’s disease, he says. It was trained on 417 clinically diagnosed Alzheimer’s disease and cognitively normal subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, including MRI scans, information on their age and gender as well as their cognitive function test score.
The same parameters were used by the neurologists offering their diagnostic impression on 80 randomly selected cases from the ADNI dataset that were not used for model training, Kolachalama says.
The 21-member team also includes Rhoda Au, professor of anatomy and neurobiology, neurology, and epidemiology at BUSM and current director of neuropsychology at the Framingham Heart Study (FHS). The population-based study is a joint project of the National Heart, Lung and Blood Institute and BU, and has been ongoing for several decades.
The unique FHS dataset (102 patients) was one of three independent cohorts used to validate the ML model, says Kolachalama. The others were the Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (382 patients) and the National Alzheimer’s Coordinating Center (582 patients).
Performance of the ML model was consistent across datasets, he adds. Impressively, it also predicted high-risk cerebral regions that closely tracked with post-mortem histopathological findings in a handful of cases where that data was available.
Translational Impact
Diagnosing Alzheimer’s disease is of course infinitely more complex than a ML model could ever face on its own, despite “hype” to the contrary in the mainstream media, says Kolachalama. The journey typically begins with an individual walking into a memory clinic complaining of forgetfulness and difficulty in doing routine activities, recounted with the help of a family member. In addition to taking a detailed history, a neurologist might conduct tests of memory, problem solving, attention, counting and language, plus a battery of medical tests and imaging studies to rule out other possible causes of the problem.
Although Alzheimer’s disease is by far the most common cause of dementia worldwide, dozens of neurological diseases can mimic its symptoms, he notes. And it can only be confirmed after death through examination of brain tissue in an autopsy.
The only reason the ML model could achieve “neurologist-level” performance is because it was designed to apply computational speed to the know-how of the human experts, says Kolachalama. It would not have been possible to build a model of this caliber without a broad multidisciplinary team that comprises expertise in artificial intelligence, neurology, image processing, study and cohort design and predictive modeling.
Other researchers have trained Al on images to predict disease, including Alzheimer’s, he continues, but the breadth of this team and the multiple sets of validation truly set this effort apart. Collecting data is incredibly complex because the imaging, clinical, and demographic information doesn’t come pre-packaged, deidentified and ready for ingestion by a computer program.
It’s also why peer reviewers for Brain seemed to appreciate the translational impact of the team’s deep learning framework. “This is only a first step,” Kolachalama says. “We’re now working toward making the algorithm more usable at the point of care,” which could involve even more rounds of validation.
“If computers can accurately detect debilitating conditions such as Alzheimer's disease using readily available data such as a brain MRI scan, then such technologies have a wide-reaching potential, especially in resource-limited settings,” says Kolachalama. The algorithm, in addition to accurately predicting the risk of the mind-robbing condition, can also “generate interpretable and intuitive visualizations of individual Alzheimer's disease risk en route to accurate diagnosis.”
Ultimately, researchers hope to extend their efforts to making predictions about who will develop Alzheimer’s in the future and creating models for other types of neurodegenerative diseases. More collaborators—both neurologists and computer scientists—are always welcome, he adds.