Brain age models are usually evaluated by prediction accuracy, but that is not enough for a clinical biomarker: you need correction for age bias, biological validity, interpretability, and robustness to transfer across samples. Nearby in the archive are papers on age-related brain maps, remodeling in Alzheimer's disease, and neurointerfaces with Riemannian geometry, but none combines all these requirements in a single pipeline on a graph representation of the connectome. This work places itself in the broader search for reliable aging biomarkers, but stands out for its systematic validation: it separates prediction accuracy from biological interpretation and explicitly tests transportability across cohorts.
Diffusion MRI from four cohorts (ADNI, Duke/UNC ADRC, HABS-HD, AD-DECODE) was processed into connectomes: 84 brain regions became graph nodes, and connections between them became edges weighted by the number of tracts. For each cohort, a graph neural network with three message-passing layers and edge weights was trained separately. Five feature sets were evaluated: imaging only, imaging plus demographics, imaging plus biomarkers, the full set, and the full set without cardiovascular variables. For fair evaluation, cross-validation with grouping by participant was used so that sessions from the same person did not fall into both training and validation. Raw age predictions were used for accuracy metrics, while the gap corrected for the age trend (cBAG) was used for biological and longitudinal analyses.
Prediction accuracy depended on the cohort more than on the feature set: mean absolute error ranged from 4.70 years in HABS-HD to 9.75 years in AD-DECODE. The graph neural network did not outperform classical models (Ridge, ElasticNet, XGBoost) on the same vector features, but it preserved the network structure for interpretation. The corrected gap cBAG was consistently correlated with reduced fractional anisotropy in white matter: in three cohorts, the association with hippocampal FA was significant (r from −0.15 to −0.44). In the longitudinal ADNI analysis, a higher baseline cBAG predicted subsequent decline in FA and local efficiency (β up to −0.285, q = 0.0016). Transferring models between cohorts produced huge errors (MAE up to 80.57 years), but linear recalibration reduced the median external MAE from 17.39 to 8.22 years, without improving rank correlation (it remained 0.171).
The work shows that connectome-based brain age can be a meaningful biomarker if prediction accuracy is separated from biological interpretation. Unlike neighboring studies, which often stop at accuracy or a single cohort, this one systematically tests bias correction, longitudinal stability, and transportability. The main conclusion: cBAG primarily reflects white matter integrity, not cognitive status or genetic risk, and its value lies not in classifying disease but in tracking an individual's aging trajectory.
The field could move toward harmonizing connectomes across scanners and protocols, as well as domain adaptation methods for model transfer. Combining data from multiple centers in federated learning is promising for increasing sample diversity without compromising privacy. A separate path is testing the stability of SHAP attributions under retraining and on independent data to understand which connections truly matter for aging.
The results will influence the development of aging biomarkers for clinical trials and population neuroscience, where measures robust to changes in equipment and participant composition are needed. The methodology for assessing transportability and calibration will be useful in any machine learning project on medical images.
The next step is to train a model on pooled harmonized data from several cohorts and test whether external transportability improves. It is also worth evaluating the stability of SHAP attributions under bootstrap resampling and on independent datasets.
This work directly addresses an unresolved problem in neuroimaging: how to distinguish a true biological aging signal from data collection artifacts and systematic model errors. It shows that even with good internal accuracy, transfer to another cohort can completely destroy predictive power, and recalibration alone is not enough to recover individual differences. This raises the question of which connectome properties are truly transportable across populations.
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.