Brain age explained: what scans from 48,598 people showed across nine disorders, and what the gap means

A brain scan can be boiled down to a single number: how old the brain looks. Software trained on thousands of scans from healthy people learns what a typical 30-year-old brain and a typical 70-year-old brain look like, then guesses the age of any new brain it is shown. That guess, minus the person’s real age, is the brain age gap.

The gap has become one of the most studied measures in brain imaging. A study published in July in the Public Library of Science (PLOS) journal PLOS Medicine, and publicized by the publisher at the end of September, compared it across nine conditions at once, from autism to Alzheimer’s disease, using one method throughout.

The gap was widest in Alzheimer’s disease, intermediate in addiction and psychiatric illness, and absent in autism and attention-deficit/hyperactivity disorder (ADHD). What that ranking means is less settled than the phrase “brain aging” suggests.

What brain age is and how the gap is calculated

The idea rests on a regular pattern. Across adult life the brain’s gray matter, the tissue that holds the bodies of nerve cells, shrinks in a fairly predictable way. A computer model shown enough scans with known ages can learn that pattern well enough to estimate age from a scan alone.

Researchers call the difference between the estimate and the truth the predicted age difference (PAD), or the brain age gap. A positive PAD means the brain appears older than would typically be expected. A negative one means it looks younger.

The estimates are imprecise for any one person. In the new study, the models’ guesses for healthy people missed by an average of about 1.5 to 5 years, depending on the group, before a statistical correction, described below, was applied.

What the nine-disorder brain scan study did

The research was led by Shile Qi of the Nanjing University of Aeronautics and Astronautics. Its raw material was magnetic resonance imaging (MRI), the scanning method that produces detailed pictures of soft tissue. The team assembled structural MRI data from 45,900 healthy controls and 2,698 patients, all of it collected earlier by other projects. Structural means scans of anatomy, and controls are people without a diagnosis. That makes 48,598 people in all.

The scans used to train the age-guessing models were obtained from 3 consortia, the largest of them the UK Biobank, a long-running British health study, which supplied 39,679. The rest of the healthy scans served as comparison groups.

The patients fell into four families:

• Developmental conditions: 344 people with ADHD and 484 with autism.
• Psychiatric disorders: 152 with schizophrenia, 143 with bipolar disorder and 258 with major depression.
• Addiction: 155 with alcohol use disorder and 144 with tobacco use disorder, with those who met both definitions also analyzed as a group of their own.
• Dementia: 361 with Alzheimer’s disease and 657 with mild cognitive impairment, a milder decline in memory and thinking that can come before it.

Each patient group was compared with a set of healthy people matched for age and sex. Because each group had its own model and its own comparison set, the differences were evaluated using standardized effect sizes rather than direct comparisons of PAD values. A standardized effect size expresses the distance between two groups relative to how much individuals vary. By the usual rule of thumb, 0.2 is small, 0.5 is moderate and 0.8 is large.

Which disorders showed the largest brain age gap

Every patient group but two looked older than its comparison group.

Condition	Effect size
Alzheimer’s disease	0.97
Alcohol and tobacco use disorders together	0.84
Tobacco use disorder	0.72
Alcohol use disorder	0.62
Schizophrenia	0.53
Bipolar disorder	0.46
Mild cognitive impairment	0.45
Major depression	0.28
Autism	0.06
ADHD	0.01

The authors sum up the order as different degrees of aging, with the highest effects in dementia, followed by addiction and psychiatric disorders, but not in developmental disorders. The tidy summary hides some overlap. Mild cognitive impairment, which the authors file under dementia, came in below all three addiction groups, schizophrenia and, narrowly, bipolar disorder.

Even the larger effects leave the groups heavily overlapping. As a rough guide, if the gaps are spread in a bell curve, an effect of 0.5 implies that about 69% of patients have an older-looking brain than the average healthy person, against the 50% expected if there were no difference at all.

Within a diagnosis, the gap said little about how ill someone was. In the two dementia groups, people with wider gaps scored somewhat lower on a standard test of memory and thinking, though the links were weak. In schizophrenia, bipolar disorder, depression and alcohol use disorder, the gap bore no relation to symptom scores.

The researchers also split the groups by age and by sex, and report, for instance, that the gap in dementia was most marked in the youngest patients. They add their own warning that these analyses should be interpreted as exploratory.

A second kind of scan, which tracks brain activity instead of anatomy, showed a gap in almost none of the groups.

Where in the brain the age gap appeared

The models can be asked which regions drove an estimate. One area that stood out was the prefrontal cortex, which showed higher PAD across multiple brain disorders. The prefrontal cortex is the front-most part of the brain’s outer layer, involved in planning and self-control.

Beyond that shared region, the patterns diverged. In psychiatric disorders the gap was concentrated in the frontal and temporal lobes, at the front and sides of the brain, and in dementia at the front and the back. Addiction showed a different pattern, involving deeper structures and two networks of regions that work together. The authors say these regional maps should be interpreted with caution, because neighboring regions tend to shrink together and are hard to tell apart.

They also asked which genes are unusually active in the regions involved, and report different biological processes for each family of disorders. That part of the analysis rests on post-mortem tissue from a limited number of donors, six in all, none of them among the people scanned.

How the result compares with earlier brain age studies

It matches them on the overall order, less so on the size of some gaps. In 2019 an Oslo-led consortium published a similar comparison in Nature Neuroscience, using structural MRI data from 45,615 individuals aged 3 to 96. Its effect sizes were 1.03 for dementia, 0.51 for schizophrenia, 0.41 for mild cognitive impairment and 0.29 for bipolar disorder. In children with autism or ADHD, there were no signs of a negative brain age gap, which would have pointed to delayed development, and the positive gaps were negligible.

Two international consortia have put the gaps in years. For schizophrenia, one consortium’s study included data from 26 cohorts worldwide, cohorts being study groups, with 2,803 patients and 2,598 healthy people. Patients’ brains looked 3.55 years older on average, an effect size of 0.48. Three authors of the new paper were among that study’s many contributors.

For depression, a consortium compared 2,675 patients with 2,126 controls and found a gap of 1.08 years, an effect size of 0.14. The authors stressed how modest that was, writing that substantial variation within each group and overlap between groups were observed. The new study’s figure for depression, 0.28 from 258 patients, is twice the consortium’s.

Addiction is where the new work adds most. The Oslo analysis did not include it, and alcohol and tobacco use had previously been examined in separate, smaller studies.

Does an older-looking brain mean a faster-aging brain?

Not necessarily, and this is the main caution in reading any brain age result. One scan cannot distinguish a brain that has been shrinking quickly from one that was always a little different.

A team at the University of Oslo tested the assumption directly. Using the UK Biobank and a second set of European studies, they compared people’s brain age gap at one scan with how much their brains went on to change, which they could see because the same people were scanned again later. They found no link between the gap and the rate of brain change measured longitudinally, meaning over time in the same individuals.

What the gap did track were things in place long before the scan. Brain age in adulthood was associated with the congenital factors of birth weight and polygenic scores of brain age. Congenital means present from birth, and a polygenic score adds up many small inherited influences. The authors concluded that results like these cast doubt on single-scan brain age figures. In their words, the results question their validity as markers of ongoing change within a person’s brain.

Part of the gap is inherited, on the 2019 study’s evidence too. It estimated that common genetic variants accounted for 24% of the variation in brain age gap between people.

The repeat-scan team allowed one exception that matters here. In disease groups such as Alzheimer’s patients, they wrote, variation in brain age might reflect to a greater extent prevailing loss of brain structure. That fits the new study’s ranking, in which the one disorder defined by progressive loss of brain tissue stands at the top.

The new study shares the one-scan design. Its authors acknowledge that using such snapshot data to analyze a dynamic process is not optimal.

There are statistical traps as well. A group led from the University of Pennsylvania showed that the gap depends on age itself, so that any group differences on the brain age gap could simply be due to group differences on age. Put plainly, two groups can differ on the gap just because they differ in age. The common fix, removing the effect of age afterward, makes a model’s accuracy look better than it is. The new study applied such a correction and reports its accuracy both ways. Its patient and comparison groups were matched on age, and age was included in the group comparisons, an approach the Pennsylvania authors say keeps group differences independent of age. Their wider conclusion was that further theoretical work is warranted to determine the best way to quantify deviation from normality. In other words, there is no agreed best way yet to measure how far a brain departs from the norm.

What a wider brain age gap has been linked to

The measure is not empty. In one of the best-known studies, a model trained on neuroimaging data from a large healthy reference sample was applied to 669 members of the Lothian Birth Cohort 1936, a group of Scottish people all born in the same year and scanned in their early 70s.

Those whose brains looked older had weaker grip strength, poorer lung function, slower walking speed, lower fluid intelligence and a higher risk of dying. Fluid intelligence means on-the-spot reasoning. Each extra year of brain age came with a 6.1% relative increase in the risk of death between age 72 and 80.

That is an association in older adults from the general population, and it does not show that the gap itself shortens life. It does suggest the number captures something about general health.

Limits of the brain age study

Cause. The release from the journal’s publisher is direct about this. The results do not show that these conditions directly cause accelerated brain aging. The study is correlational, and some conditions, particularly psychiatric disorders and addiction, frequently occur together.

Which condition is responsible. The authors list this as a limitation of their own: psychiatric disorders and addiction have high comorbidity, and these potential confounders were not considered. Comorbidity means having more than one condition at once, and confounder means an outside factor that can produce a misleading link. A person counted under schizophrenia may also smoke heavily, and smoking had one of the larger effects in the study.

Everything else that shapes a brain. Ethnicity, other chronic diseases, lifestyle and environmental exposures were not available across all the data sets. As a result, the authors write, residual confounding effects cannot be fully excluded.

Small patient groups. The headline total is large because of the healthy scans. Each condition was represented by between 143 and 657 people.

Anything about one person. These are group averages with wide overlap. The authors say only that the patterns could be tested in future studies for their usefulness as biomarkers, meaning measurable signs that might one day help with diagnosis.

The authors’ own summary is modest. “Different neurological disorders appear to leave different signatures on the brain aging clock, which may help researchers better understand the neural and biological pathways involved in these conditions,” they said in the publisher’s release.

Across nine conditions, brains looked oldest for their age in Alzheimer’s disease and no different in autism or ADHD, in a comparison of single scans that cannot say whether any of those brains were aging faster.