IBM uses machine learning to detect early Alzheimer’s in blood samples

Long before memory loss occurs in patients with Alzheimer’s, the protein amyloid-beta—which has been implicated in the formation of brain tangles that characterize the disease—starts building up in the spinal fluid. Problem is, detecting the protein there requires an invasive procedure that’s done under anesthesia, making it impossible to deploy on a widespread basis for the early detection of Alzheimer’s.

Researchers at IBM in Australia have developed an alternative to spinal fluid testing: a blood test that they say could predict the buildup of amyloid-beta in spinal fluid with up to 77% accuracy.

The team members used machine learning to pinpoint four proteins that can be measured in the blood of people who have a high genetic risk of Alzheimer’s. They believe that testing for those proteins could be used to identify patients who are not yet symptomatic but may benefit from early treatment, and to improve the selection of patients for clinical trials of drug candidates.

They published their findings in the journal Scientific Reports.

The IBM researchers started by downloading data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), a long-term imaging study of older adults that’s examining whether MRI and PET images can be combined with other measures to predict early-stage Alzheimer’s. They examined 566 people in the ADNI study, 182 of whom were carriers of APOEε4, a gene variant that raises the risk of late-onset Alzheimer’s.

RELATED: Fighting ALS, dementia and more by trapping toxic brain proteins

The IBM researchers were able to obtain information on nearly 400 proteins that were measured in blood samples from the ADNI cohort, but they wanted to develop a simple test that could use as few proteins as possible to reflect the buildup of amyloid-beta. Using machine learning, they determined that the proteins Chromogranin-A (CGA), Aβ1–42 (AB42), Eotaxin 3, and Apolipoprotein E (APOE) were the best predictors. They validated the model by comparing it to blood tests using different variables, as well as a predictive model that just used age and APOEε4 status.

“While all models making use of blood analytes outperformed the base model of age and APOEε4, models that made use of the protein level measurements consistently achieved the strongest predictive performance,” the IBM researchers wrote in the study.

One potential use of the four-protein blood test may be to improve the selection of patients for clinical trials of Alzheimer’s drugs, the IBM researchers proposed. That’s because studies have shown that changes in amyloid levels in cerebral spinal fluid can occur as long as 10 years before PET imaging turns up signs of Alzheimer’s. So pharma companies developing drugs to slow or halt the progression of the disease could use a low-cost, minimally invasive way to find clinical trial participants who face a high risk of Alzheimer’s but are many years removed from developing symptoms of the disease, they said.

Clinical trial failures have sidelined several experimental Alzheimer’s treatments. Most recently, Roche put the brakes on two trials of crenezumab, its anti-Abeta antibody, because an interim data analysis concluded they were unlikely to succeed. Then Biogen spooked its investors by disclosing in February that it’s adding 510 patients to a phase 3 trial of a similar antibody, aducanumab, due to “variability in the primary endpoint” seen in an analysis.

The high failure rate of these and other Alzheimer’s trials “may be because the people enrolled are in the latest stages of the disease, and have likely already suffered a level of brain tissue loss that cannot easily be repaired,” suggested Ben Gaudey, a staff researcher on the genomics research team at IBM, in a blog post.

The IBM team envisions a day when clinicians and clinical trial investigators will be able to use simple blood tests to identify the patients facing the highest risk of developing the disease. Toward that end, the researchers are working on a similar test to detect another major Alzheimer’s biomarker, the protein tau. And they believe their machine-learning model will be able to be applied to other diseases that can be detected in spinal fluid before symptoms emerge.

“As our population lives longer, neurodegenerative diseases such as Parkinson’s, Alzheimer’s and Huntington’s are affecting millions of people around the world,” Gaudy said. “While these mysterious and crippling diseases do not yet have a cure, the answer to slowing their growth may lie in prevention.”