What Brain-PAD means, and what it does not

What brain age is

A brain age model learns the relationship between brain images and age in a large group of healthy adults. Once trained, it can be applied to a new scan and return an estimated age. The idea is simple: brains change with age in predictable ways (grey matter thins, ventricles widen, white matter changes), and a model can summarize thousands of those changes in one number.

The difference between the estimate and the person's actual age is the Brain-PAD (brain predicted age difference). Across many studies, groups with neurodegenerative disease, and some groups with other health conditions, show higher average Brain-PAD than comparison groups. That is why the measure has become popular as a research outcome and as a candidate biomarker (Cole and Franke, 2017).

How it is estimated from MRI

Most brain-age models start from a T1-weighted MRI scan, the standard structural image collected in almost every imaging study. From that scan, a model may use every voxel, regional measures such as cortical thickness and volume from morphometry, or small patches of tissue.

Our own work compared these choices. A patch-based procedure estimated age accurately across adulthood (Beheshti et al., 2019), and a direct comparison found that the choice between voxel-wise and region-wise inputs changes accuracy (Beheshti et al., 2022b). A separate evaluation of many machine-learning algorithms showed that the algorithm matters less than careful evaluation on data the model has not seen (Beheshti et al., 2022).

The practical lesson for a study is that the model, its training data and its evaluation should be reported as carefully as any other measurement instrument. A brain-age value without that information is hard to interpret and hard to compare across studies.

Why age-bias adjustment matters

Every brain-age model shares one statistical habit: it pulls estimates toward the average age of its training data. Young people tend to be estimated as older than they are and older people as younger. This is regression toward the mean, and it means raw Brain-PAD is correlated with age itself.

If that bias is ignored, a group that is simply older or younger than another can appear to have a different Brain-PAD for no biological reason. Several adjustment schemes exist. We proposed and tested a robust one that estimates the bias in a separate set of healthy people and removes it before analysis (bias adjustment; Beheshti et al., 2019b).

Adjustment is necessary but not sufficient. In the main analysis, actual age should normally still be included as a covariate, and the adjustment method should be stated in the paper so a reviewer can follow it.

How reliable it is over time

A measure that changes from one scan to the next for no reason cannot track real change. When the same people were scanned repeatedly, patch-wise brain age was highly consistent over time (Beheshti et al., 2021). That makes brain age a reasonable candidate for longitudinal studies, where the question is whether change differs between groups.

Reliability is a property of a model in a population and a scanner setting, not a fixed fact. A study that uses different scanners, protocols or age ranges should check reliability in its own data or justify why published estimates apply. A test–retest subsample is often the simplest way to do that (test–retest reliability).

Reliability also shapes sample size. When a measure is noisy, real differences between groups are diluted, and a study needs more participants to detect them. Reporting the reliability you assumed, and where the estimate came from, lets a reviewer check that the study was planned for the measure it actually used. In a mixed-effects model of repeated scans, reliability and true change can be separated explicitly, which is one reason longitudinal designs are often more informative than a single scan per person.

What a positive Brain-PAD can and cannot say

At the group level, a higher average Brain-PAD in one group than another says that, on average, the brains in that group look older than expected for their age, according to this model. That is a useful, testable research finding.

For one person, a positive value is much harder to interpret. Individual estimates carry error of several years, the model may not suit that person's age range or scanner, and the measure is not tied to any specific disease. Brain-PAD is a research measure, not a diagnosis, and Cedar does not use it to assess individuals.

A checklist for using brain age in a study

1. State the research question that brain age answers better than regional measures.
2. Describe the model: input features, training data, age range and software version.
3. Report accuracy on data the model was not trained on (mean absolute error and correlation).
4. Adjust for age bias, name the method, and keep age as a covariate in the main analysis.
5. Check reliability in your setting or justify the published estimate.
6. Account for scanner and site differences in multi-site data.
7. Report group-level results with confidence intervals, and avoid individual interpretation.
8. Confirm that the dataset's data-use agreement permits the analysis.

References

1. Cole JH, Franke K. Predicting age using neuroimaging: innovative brain ageing biomarkers. Trends in Neurosciences. 2017. doi:10.1016/j.tins.2017.10.001 (external site)
2. Beheshti I, Gravel P, Potvin O, Dieumegarde L, Duchesne S. A novel patch-based procedure for estimating brain age across adulthood. NeuroImage. 2019. doi:10.1016/j.neuroimage.2019.05.025 (external site)
3. Beheshti I, Nugent S, Potvin O, Duchesne S. Bias-adjustment in neuroimaging-based brain age frameworks: A robust scheme. NeuroImage: Clinical. 2019. doi:10.1016/j.nicl.2019.102063 (external site)
4. Beheshti I, Potvin O, Duchesne S. Patch-wise brain age longitudinal reliability. Human Brain Mapping. 2021. doi:10.1002/hbm.25253 (external site)
5. Beheshti I, Ganaie MA, Paliwal V, Rastogi A, Razzak I, Tanveer M. Predicting Brain Age Using Machine Learning Algorithms: A Comprehensive Evaluation. IEEE Journal of Biomedical and Health Informatics. 2022. doi:10.1109/JBHI.2021.3083187 (external site)
6. Beheshti I, Maikusa N, Matsuda H. The accuracy of T1-weighted voxel-wise and region-wise metrics for brain age estimation. Computer Methods and Programs in Biomedicine. 2022. doi:10.1016/j.cmpb.2021.106585 (external site)