Retrained on Nigerian scans, a brain-age model reaches 9.82 years of error, then 8.78 on a second cohort

Brain-age prediction is being retrained for the population it serves instead of applying Global North models directly to Nigerian adults. Existing brain-age models were trained and validated almost exclusively on Global North data from North America, Europe and Australia, which differ from Nigerian clinical imaging in population, scanner and acquisition protocol. Among 309 neurologically normal Nigerian adults, the population-specific model reached an uncorrected mean absolute error of 9.82 years on the 47-participant held-out test set; on the 44-participant head-to-head set it reached 10.24 years, against 10.52 years for PyBrainAge, 12.33 for SynthBA, 17.51 for ENIGMA-Brain Age and 7.87 for SPARROW.

The route is to turn routine clinical scans into usable three-dimensional volumetric features, then predict age from them with a lightweight regression model. Gaussian process regression is a kernel-based Bayesian regression that uses a radial basis function kernel to describe a smooth relationship between brain-structure features and age, with small training and inference cost. The data are 343 clinical T1-weighted images from five 1.5T scanners in AfNiA; SynthSR synthesized 1 mm isotropic volumes from 2D acquisitions, FreeSurfer segmentation yielded 187 cortical and subcortical volume features via the Destrieux atlas, and 309 participants entered training, validation and testing.

Age-bias correction sharply lowers internal test error, but the advantage narrows on an independent cohort. Uncorrected, the brain age gap correlated significantly with chronological age with a slope of -0.56, overestimating age in younger participants and underestimating it in older ones. After correction with slope and intercept estimated on the validation set, internal test error fell from 9.82 to 6.39 years; on a separate 25-person Nigerian cohort it was 8.78 years corrected and 13.29 years uncorrected, and every model including this one had a negative R2 there.

The findings rest on clinical T1-weighted MRI from 309 neurologically normal Nigerian adults, with external validation limited to one separate 25-person Nigerian cohort; the authors propose expanding imaging data through the Africa Neuroimaging Archive to further develop the population-specific model for identifying accelerated brain ageing.

The authors note that sample size limited model complexity and may raise prediction error across all models, that imbalanced sex distribution prevented a robust sex-stratified analysis, that dependence on FreeSurfer preprocessing remains a major constraint, and that the external cohort may differ from the training distribution in unmeasured ways, with applicability beyond Nigeria still untested.