How AI can spot Alzheimer’s disease five years before diagnosis

Despite all the hype, artificial intelligence (AI) is not ready to replace doctors or automate brain surgery. But it can analyse immense amounts of data to help us better study, diagnose, treat and even prevent disease. Our hope is that it could reverse two decades of failed experimental therapies for Alzheimer’s disease.

Some of the most promising AI applications in medicine are in machine learning, a type of intelligence that uses algorithms to learn from patterns in data. Similar to how Siri learns your voice and Facebook can predict what videos you’ll like, learning algorithms can scan the retinas of diabetics and predict sight loss or sequence the genomes of entire towns to flag disease-causing genetic mutations.

We wanted to know what a new unbiased type of AI called “multilayer clustering” could tell us about Alzheimer’s. The push for Alzheimer’s treatments has been exciting and disappointing. In 20 years, we have yielded a greater understanding of the disease, yet failed to produce a successful preventive therapy.

The many undefined subtypes of mild cognitive impairment, a precursor to Alzheimer’s, are partly to blame. When trials are testing one drug on numerous types of cognitive impairment, should we be surprised when more than 99% of those trials fail?

Scientists could much more precisely test the impact of investigational drugs if we could better identify and group those with similar brain changes and cognitive impairment. To do this, we applied a multilayer clustering algorithm to analyse dozens of data points from a US study called the Alzheimer’s Disease Neuroimaging Initiative, including cognition tests, brain scans and spinal fluid biomarkers. The samples were taken from 562 people with mild cognitive impairment followed for up to five years.

AI identified two groups of interest: those whose cogn­­­ition declined very rapidly and were at highest risk for Alzheimer’s; and those who saw little or no decline in their symptoms. A third group remained unclustered.

Brain scans of ‘rapid decliners’ showed twice the rate of atrophy as those for whom the disease was slow moving. Rapid decliners also progressed from mild cognitive impairment to dementia at five times the rate of those with a slow-moving disease. The full findings were published in July in Scientific Reports.

Looking at common traits of rapid decliners, it’s possible that in the future such algorithms could help doctors identify those most at risk from Alzheimer’s years before diagnosis.

These findings have direct implications for the design of future trials. We have known bits and pieces of this information – that there are dozens of genes that put people at risk or that certain brain changes put people at risk – but an unbiased tool such as an AI algorithm could put all of these pieces of information together, painting a picture that scientists might not have otherwise seen.