Skip to content

Machine Learning Model Identifies Alzheimer's Disease–like Structural Brain Patterns in Younger, Asymptomatic Adults

Key findings

  • MRI patterns of brain atrophy found in individuals with Alzheimer's disease (AD) are extremely subtle at young ages, and they overlap with typical age-related structural changes
  • Researchers at Massachusetts General Hospital trained a computer classifier to factor out typical age effects, and also weight patterns of atrophy due to AD, when classifying structural brain patterns on MRIs
  • In a cohort used for training, the classifier was 94% accurate in distinguishing 136 individuals with AD from 268 healthy controls, and it detected AD-like brain changes in individuals who had mild cognitive impairment
  • The classifier then detected AD-like structural brain patterns in an independent cohort, including 8% of participants who were 40 to 59 years old, and AD-like patterns were associated with subtle reductions in performance on cognitive tests
  • While the work requires further validation, the findings suggest that this type of classifier might be used to determine which younger adults are at risk of AD and help select candidates for clinical trials of very early interventions

Recent advances in brain MRI show promise in allowing early identification of individuals at risk of future Alzheimer's disease (AD). These methods have primarily been applied to people >55 years old.

Detecting structural abnormalities in younger adults is more challenging. The patterns of brain atrophy found in individuals with AD are extremely subtle at young ages, and they overlap with typical age-related structural changes.

Binyin Li, MD, at the time of the study a research fellow in the Martinos Center for Biomedical Imaging at Massachusetts General Hospital, David H. Salat, PhD, director of the Brain Aging and Dementia Laboratory at the center, and colleagues developed a novel procedure for factoring out typical age effects when classifying MRIs. In Human Brain Mapping, they report identifying very early AD-like atrophy patterns in undiagnosed and younger, asymptomatic adults.

Creation of a Machine Classifier

To model age-related trends in brain structure, the researchers used brain MRIs from 136 participants in the Human Connectome Project Aging (HCP-A), which enrolled generally healthy adults 36 to >90 years old.

They weighted AD structural deterioration with patterns quantified from 136 individuals with AD and 268 healthy controls in the Alzheimer's Disease Neuroimaging Initiative (ADNI). The team then trained a support vector machine to classify individuals with brains that most resembled atrophy in AD.

Accuracy in the ADNI Cohort

In the ADNI cohort, the classifier was 94% accurate in distinguishing individuals with AD from controls.

The team also validated the classifier on 180 individuals from the ADNI who had early mild cognitive impairment and 96 who had late MCI. In both groups, individuals with AD-like changes according to the classifier had significantly worse scores on general cognition, memory and trail-making than those who were free of AD-like changes. They also had significantly greater amyloid-beta and tau, and significantly lower brain metabolism.

Accuracy in an Independent Cohort

A separate cohort of 136 adults from the HCP-A was used to determine whether the classifier could correctly identify undiagnosed individuals at high risk of AD:

  • 9% were classified as having AD-like changes (ADL group) and 91% were not (NADL group)
  • The ADL group included five of 65 participants between 40 and 59 years old (8%)
  • The ADL group had significantly worse scores than the NADL group on cognitive tests, and an AD risk score derived from the classifier significantly correlated with test scores

Hope for the Future

Because of certain limitations in the study procedure, the researchers describe the results as a proof-of-concept only. They have more research underway, but these results do suggest the classifier identified individuals with clinically meaningful brain changes. The classifier may prove capable of determining which younger adults are at risk of AD and help select candidates for clinical trials of very early interventions.

accuracy of a machine classifier in distinguishing individuals with Alzheimer's disease from healthy controls

of individuals 40 to 59 years old had Alzheimer's disease–like changes detected by a machine classifier

Learn more about the Martinos Center for Biomedical Imaging

Learn more about research in the Department of Radiology


Ibai Diez, PhD, and Jorge Sepulcre, PhD, DMSc, MD, of the Gordon Center for Medical Imaging, explain how neuroimaging findings are being combined with genetic data—research that might lead to disease-modifying therapies for neurodegenerative diseases.


Valentina Perosa, MD, and Susanne J. van Veluw, PhD, of the Department of Neurology, and colleagues have developed the first convolutional neural networks that can identify and quantify biologically relevant histopathological markers of both Alzheimer's disease and cerebral amyloid angiopathy.