Summary: A new AI model can quantify the pace at which a person’s brain is aging by analyzing longitudinal MRI scans. This approach offers a powerful, noninvasive way to detect accelerating brain aging, identify the regions most affected, and link those changes to declines in cognitive function — potentially enabling earlier diagnosis and personalized interventions for dementia and Alzheimer’s disease.
USC researchers report that a faster pace of brain aging strongly correlates with worsening cognitive performance. By tracking anatomical changes over time rather than relying on a single scan, the model yields more precise measures of ongoing neuroanatomic decline and can highlight individuals who may benefit from early treatment or monitoring.
Key Facts
- Longitudinal brain tracking: The model analyzes baseline and follow-up MRI scans to measure how quickly brain tissue changes over time, producing a more accurate assessment than single-scan methods.
- Predicts cognitive decline: Higher rates of brain aging are associated with measurable declines in memory, processing speed, executive function, and other cognitive domains.
- Potential for earlier diagnosis: By identifying accelerated brain aging before symptoms appear, the tool could help flag people at elevated risk for Alzheimer’s and other neurodegenerative conditions.
Source: USC
Overview
A newly developed artificial intelligence tool estimates the temporal pace of brain aging from magnetic resonance imaging (MRI) scans. Unlike prior cross-sectional approaches that estimate a single “brain age” from one scan, this longitudinal model compares scans taken at different times for the same person, allowing clinicians and researchers to measure how quickly a brain is aging right now.
Andrei Irimia, associate professor at the USC Leonard Davis School of Gerontology and coauthor of the study, explains that biological brain age differs from chronological age: two people born the same year may have brains that appear very different in terms of structural aging. Traditional blood-based measures of biological age, such as DNA methylation clocks, do not reliably reflect brain aging because the blood–brain barrier prevents many cellular changes in the brain from being detectable in peripheral blood.
Sampling brain tissue directly is invasive and impractical for routine monitoring, so MRI-based approaches offer a practical alternative. Previous AI models estimated brain age from a single scan by comparing anatomy to large reference datasets, but those cross-sectional estimates could not show whether brain aging was accelerating or slowing over time.
A longitudinal 3D-CNN for measuring brain aging
To address these limitations, researchers developed a three-dimensional convolutional neural network (3D-CNN) trained on longitudinal MRI data. Built in collaboration with Paul Bogdan at the USC Viterbi School of Engineering, the model was trained and validated on more than 3,000 scans from cognitively normal adults. By comparing baseline and follow-up images for the same individual, the model calculates the pace of brain aging and produces interpretable saliency maps that highlight the anatomical regions most influential in its estimates.
When tested on an independent cohort that included 104 cognitively normal adults and 140 patients with Alzheimer’s disease, the model’s estimated aging rates aligned closely with changes on cognitive tests administered at both time points. This alignment suggests the model can function as an early biomarker of neurocognitive decline in both healthy and impaired populations.
Irimia notes that the model not only measures anatomy but links those anatomical changes to cognitive outcomes: individuals with higher rates of brain aging tended to show larger declines in memory, processing speed, and executive function. The saliency maps also revealed regional differences in aging rates, which varied by sex and life decade and may help explain divergent disease trajectories between men and women.
Performance and implications
The longitudinal model computes the pace of brain aging with high precision in its test set: a mean absolute error (MAE) of 0.16 years (about 7% mean error), substantially outperforming the best cross-sectional model, which had an MAE of 1.85 years (around 83% error). By mapping where aging accelerates, the method can better characterize healthy aging and disease progression and may inform personalized treatment decisions in the future.
Researchers emphasize the tool’s potential to identify people whose brains are aging faster than typical before cognitive symptoms become evident. Early identification could improve the timing and effectiveness of interventions, including preventive strategies and therapies for Alzheimer’s disease, which may be more effective if started before significant pathology accumulates.
Future directions
Ongoing work aims to link regional aging patterns to genetic, environmental, and lifestyle factors to understand why certain pathologies emerge in particular brain regions. The researchers are also exploring how the model might contribute to individualized Alzheimer’s risk estimates and to selecting treatments that best match a person’s neuroanatomic trajectory.
The study’s authors include Andrei Irimia, Paul Bogdan, first author Chenzhong Yin, Heng Ping, Phoebe E. Imms, Nahian F. Chowdhury, Nikhil N. Chaudhari, and Haoqing Wang.
Funding: Support came from multiple sources including the National Institutes of Health (grants R01 NS 100973, RF1 AG 082201, R01 AG 079957), the Department of Defense (contract W81XWH-18-1-0413), the National Science Foundation (CAREER Award CPS/CNS-1453860 and grants MCB-1936775, CNS-1932620), the U.S. Army Research Office (grant W911NF-23-1-0111), DARPA awards, an Intel Faculty Award, Northtrop Grumman, the Hanson-Thorell Research Scholarship Fund, USC undergraduate research programs, and anonymous donors.
About this AI and brain aging research news
Author: Elizabeth Newcomb
Source: USC
Contact: Elizabeth Newcomb – USC
Image credit: Neuroscience News
Original Research: Closed access. Title: Deep learning to quantify the pace of brain aging in relation to neurocognitive changes — Andrei Irimia et al., Proceedings of the National Academy of Sciences. DOI: 10.1073/pnas.2413442122
Abstract (summary)
Brain age (BA) estimated from MRI differs from chronological age (CA) and has been used to evaluate cumulative neuroanatomic aging. However, BA is cross-sectional and does not capture the recent temporal trends of brain aging. The pace P of brain aging better reflects contemporaneous aging but has been difficult to measure noninvasively. The study introduces a three-dimensional convolutional neural network (3D-CNN) trained on longitudinal MRI to estimate P. The longitudinal model was trained on MRIs from 2,055 cognitively normal adults, validated on 1,304 cognitively normal adults, and applied to an independent cohort of 104 cognitively normal adults and 140 patients with Alzheimer’s disease. In testing, the model estimated P with a mean absolute error of 0.16 years (7% mean error), outperforming the most accurate cross-sectional model (MAE 1.85 years, 83% error). Combined with interpretable saliency mapping, the model identifies regional variations in aging rates that differ by sex, decade of life, and cognitive status, and its estimates of P associate significantly with changes across cognitive domains, underscoring its utility for studying neuroanatomic and neurocognitive aging trajectories.