An old road atlas lies on the table. On its pages, lines connect cities, and the thickness of each line tells you how busy the traffic once was. A neural network looks at such a map and names the age of the car that drove along it. That is roughly how a model trained to predict a person's age from the map of their brain connections works.
Instead of cities, this map has 84 brain regions, and instead of roads, bundles of nerve fibers connecting them. The thickness of each connection depends on how many fibers pass through it. Such a map is called a connectome. It is built from diffusion MRI scans, which show how water moves along neural pathways.
Scientists trained such models on data from four different groups of participants: healthy older adults, patients with Alzheimer's disease, and people with mild cognitive impairment. In total, 1093 MRI sessions. The model was trained separately for each group, because the scanners and imaging protocols differed.
Age prediction accuracy varied: in one group the model was off by an average of 4.7 years, in another by 9.75 years. But the key measure was not age itself, but the difference between predicted and actual age. It was called the brain age gap. If the model says the brain is 70 years old while the person is 65, then the brain looks older than its years.
The problem is that the brain age gap carries a systematic bias: for younger people the model tends to err in one direction, for older people in the other. So the scientists subtracted the expected trend from the gap and obtained the corrected gap, or cBAG. It was precisely this measure that turned out to be linked to white matter integrity. In three of the four groups, people with a higher cBAG had lower fractional anisotropy in the hippocampus, a region important for memory.
In a longitudinal analysis, where the same people were scanned several times, a higher baseline cBAG predicted a subsequent decline in fractional anisotropy and in local network efficiency. In other words, a brain that looks older really does lose the quality of its wiring faster.
The hardest part turned out to be transferring the model between groups. When a model trained on one group was applied to another, the age error reached 80 years. The model could make a person almost a century older. After linear recalibration the error dropped to 8 years, but the ranking of people by age still was not restored. This means the model picks up on quirks of a particular scanner, not just biology.
The old atlas is back on the table. Now you can see that its lines are not just roads but a record of how they wore down. The neural network reads that record, but it needs to learn from maps drawn by a single cartographer so as not to confuse the style of the drawing with the state of the roads.
When a model trained on the AD-DECODE cohort was transferred to other cohorts, the age prediction error reached 80 years, meaning the model could make a person almost a century older. After linear recalibration, the error dropped to 8 years, but the ordering of people by age still was not recovered.
The idea of a biological age distinct from chronological age echoes H. G. Wells's The Time Machine, where time flows unevenly for different beings. Here, each person's brain ages at its own pace, and the model tries to capture that individual trajectory from the structure of connections.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.