Tracing Sources of Brain Age Prediction Degradation in Low-Resource Clinical Settings

Deep brain-age models are trained predominantly on high quality structural MRI from high-resource cohorts, yet remain understudied in low-resource clinical settings. Here, we report the first zero-shot evaluation of brain-age imaging models on sub-Saharan African clinical MRI, testing three frozen backbones on Thick slice acquisitions T1 scans from four Nigerian scanners. In our evaluation, MAE increased from 4.1-4.6 years on held-out ADNI data to 12.1-13.0 years on the Nigerian cohort, with systematic overestimation of $+8.2$ to $+10.8$ years. Our analyses suggest that this generalization gap reflects two distinct imaging-related effects rather than population differences alone as no single acquisition perturbation reproduced the offset in degraded ADNI scans, meanwhile, attention-map similarity are different between ADNI test dataset and Nigerian dataset. Finaly, we fine-tuned the model with data removing registration on native and degraded ADNI scans and sucessfully reduced the Nigerian offset to $+1.7$ years (CI $[-2.6,+5.5]$) and MAE from $+13.0$ to $+9.3$ years, on Nigerian age labels.