Machine learning spots Alzheimer's in brain scans with 93% accuracy, revealing sex-based differences

Published in Neuroscience, 2026

A 93% detection rate from anatomy alone

An estimated 6.9 million Americans age 65 and older live with Alzheimer's disease. Early diagnosis remains difficult because initial symptoms often look like normal aging. A research team at Worcester Polytechnic Institute has demonstrated that machine learning can distinguish between healthy brains, mild cognitive impairment, and Alzheimer's disease with 92.87% accuracy, using nothing but anatomical measurements from MRI scans.

The study, published in the journal Neuroscience, analyzed 815 MRI scans from the Alzheimer's Disease Neuroimaging Initiative, a multicenter project containing brain scans from people aged 69 to 84. The scans covered individuals with normal cognitive function, mild cognitive impairment, and diagnosed Alzheimer's disease.

Narrowing 95 brain regions to the ones that matter

Rather than feeding entire MRI images into a neural network, which would demand enormous computing power, the WPI team took a two-step approach. First, they used machine learning to extract volume measurements from 95 distinct brain regions across all 815 scans. Then they deployed an algorithm to identify which volume differences between healthy individuals and those with cognitive impairment or Alzheimer's were most predictive.

Three structures emerged as top predictors across all age and sex categories: the hippocampus, responsible for memory and learning; the amygdala, which controls emotions; and the entorhinal cortex, a hub for memory, navigation, and perception that is among the first brain regions affected by Alzheimer's.

Volume loss in the right hippocampus appeared in both males and females in the youngest age group studied (69 to 76), suggesting it may be particularly important for early detection.

Where men's and women's brains diverge

The sex-specific patterns were unexpected. In women, significant volume loss occurred in the left middle temporal cortex, a region involved in language, memory, and visual perception. In men, volume loss was more pronounced in the right entorhinal cortex.

Benjamin Nephew, assistant research professor in the Department of Biology and Biotechnology and lead researcher, said the degree of these differences was surprising. He suggested they may relate to interactions between Alzheimer's progression and changes in sex hormones. Some researchers have linked Alzheimer's risk to estrogen loss in women and testosterone decline in men as they age.

These findings carry practical implications. If diagnostic tools do not account for sex-based differences in brain atrophy patterns, they may miss early signs in some patients or misclassify others.

Building toward a generalizable model

A key challenge in this type of research is ensuring that the biomarkers identified are not artifacts of a specific dataset. Nephew described the goal as building a generalizable model, one where the identified patterns are universal across patient populations rather than unique to the Alzheimer's Disease Neuroimaging Initiative data.

The team is following up by evaluating deep learning models and examining other factors that may interact with brain atrophy and Alzheimer's, including diabetes. The research has attracted WPI students from biology, biotechnology, neuroscience, psychology, computer science, and bioinformatics.

Limitations worth noting

The study used MRI scans from a specific age range (69 to 84) and a specific dataset. Whether the same patterns hold in younger populations or in datasets from different demographics is an open question. The study also identifies correlations between brain volume loss and disease status; it does not establish that these volume changes cause cognitive decline.

Machine learning models are only as good as the data they train on. The Alzheimer's Disease Neuroimaging Initiative is a well-curated dataset, but it may not fully represent the diversity of the broader patient population. Validating these findings across multiple independent datasets will be essential before any clinical application.