Researchers at Worcester Polytechnic Institute trained a machine learning system to read anatomical changes in the brain and predict Alzheimer's disease with close to 93% accuracy. The work, published in the journal Neuroscience, also found that the pattern of change - specifically the loss of brain volume - is not uniform across patients but varies by age and by sex.
That second finding may matter as much as the accuracy figure. If the brain shrinks in different places, or at different rates, depending on who the patient is, then a single one-size-fits-all reading of a scan risks missing disease in some groups and overcalling it in others. Age- and sex-specific patterns give researchers a more precise baseline for what counts as an abnormal amount of tissue loss.
The practical problem the WPI team is aiming at is timing. Early Alzheimer's is notoriously hard to catch because the first symptoms look like ordinary aging, and by the time the picture is unmistakable, considerable neuronal damage has already occurred. The researchers argue that algorithms can sift through far more scan data than a clinician reasonably can, picking up subtle structural shifts and sorting patients into cognitive states. Alzheimer's kills neurons, and the resulting loss of brain tissue is what these models are learning to detect.
An estimated 6.9 million Americans aged 65 and older are living with Alzheimer's disease, so even modest gains in how early it is identified would affect a large population. The team frames this as research that has informed the field and may eventually lead to tools clinicians can use, rather than a diagnostic ready for the clinic today.