Machine-learning program shows promise for Alzheimer’s diagnosis
Researchers have developed a new machine-learning program that appears to outperform other methods for diagnosing Alzheimer’s disease…
Researchers have developed a new machine-learning program that appears to outperform other methods for diagnosing Alzheimer’s disease before symptoms begin to interfere with everyday living, initial testing shows.
The computer program integrates a range of Alzheimer’s disease indicators, including mild cognitive impairment. In two successive stages, the algorithm selects the most pertinent to predict who has Alzheimer’s.
The team developed what it calls Cascaded Multi-view Canonical Correlation (CaMCCo) algorithm, which integrates measurements from magnetic resonance imaging (MRI) scans, features of the hippocampus, glucose metabolism rates in the brain, proteomics, genomics, mild cognitive impairment and other parameters.
The program then assesses the variables in a two-stage cascade. First, the algorithm selects the parameters that best distinguish between someone who’s healthy and someone who’s not. Second, the algorithm selects from the unhealthy variables those that best distinguish who has mild cognitive impairment and who has Alzheimer’s disease.
“Many papers compare the healthy to those with the disease, but there’s a continuum,” said Anant Madabhushi, the F. Alex Nason Professor II of biomedical engineering at Case Western Reserve University. “We deliberately included mild cognitive impairment, which can be a precursor to Alzheimer’s, but not always.”
Dr Madabhushi’s lab has repeatedly found that integrating dissimilar information is valuable for identifying cancers. This is the first time he and his team have done so for diagnosis and characterisation of Alzheimer’s disease.
“The algorithm assumes each parameter provides a different view of the disease, as if each were a different set of coloured spectacles,” said Dr Madabhushi.
In predicting which patients in the study had Alzheimer’s disease, CaMCCo outperformed individual indicators as well as methods that combine them all without selective assessment. It also was better at predicting who had mild cognitive impairment than other methods that combine multiple indicators.
The researchers continue to validate and fine-tune the approach with data from multiple sites. They also plan to use the software in an observational mode: As a collaborating neurologist compiles tests on patients, the computer would run the data.
