Skip to navigation Skip to main content Skip to footer

Approved Research

Brain age versus chronological age: a large-scale, reproducible study

Principal Investigator: Professor Alexander Lundervold
Approved Research ID: 59172
Approval date: June 7th 2021

Lay summary

The human brain is constantly changing throughout the lifespan. A progression of brain substance loss is expected, as well as volumetric, morphometric and signal changes in a variety of structures. The biological aging of the brain therefore reflect the individual's chronological aging.

So-called "brain age" models attempt to use this fact to predict an individual's chronological age directly from imaging data. Earlier work has shown that machine learning methods can be used to construct accurate brain age models directly from neuroimaging data from healthy persons, e.g. MRI recordings of their brains.

In individuals with brain disorders, such as Alzheimer's disease, one expects biological and cognitive deviations from normal aging. This can potentially be detected as a gap between the predicted age and the actual chronological age, the so-called "brain age gap". This has also been shown to be relevant to other diseases and conditions, e.g. neuropsychiatric and neurodevelopmental disorders, obesity and traumatic brain injuries, and the potential clinical impact of being able to assess the risk of age-related disease has motivated a lot of research into the construction of brain age models.

In our project we will create a fast (i.e minutes) and accurate end-to-end pipeline for brain age prediction directly from MRI examinations, using state-of-the-art deep learning techniques.

Our investigations indicate that there are still significant challenges related to robustness of brain age models. In our preliminary studies we have constructed a brain age model trained on a large heterogeneous collection of data. By careful evaluation of the model performance in various experiments with different setups of training and test data, we have strong indications that enlarging the imaging data used for training the model would significantly improve its performance and robustness.

All the details necessary to reproduce and extend our findings will be made available to other researchers. The project will therefore provide a stepping-stone towards bringing brain age models closer to practical usefulness.

Our project is already well underway, and we aim to complete the main steps by Q2 2021.