Deep learning-derived age of hippocampus-centred regions is influenced by APOE genotype and modifiable risk factors

Brain age has been widely investigated by using the whole brain image. However, the age of some specific brain regions, such as those related to the hippocampus, remains underexplored. This study developed age prediction models for left and right hippocampus-centred regions of interest (hippocampus ROI) using three-dimensional convolutional neural networks (3D-CNN) based on MRI scans from 31,370 healthy participants in the UK Biobank. The hippocampus ROI age (HA) gap was calculated by subtracting chronological age from predicted HA. Additionally, the longitudinal change rate of the HA gap was estimated in 3,893 participants with imaging data at two time points over an average follow-up of 2.63 years. The models achieved state-of-the-art performance (mean absolute error (MAE): 2.47 to 2.84 years). Cross-sectional analysis revealed that APOE ε4 homozygotes had a greater HA gap compared to APOE ε4 non-carriers. Participants with hypertension, diabetes, heavy alcohol consumption, or smoking also exhibited larger HA gap. Transfer learning applied to an independent dataset confirmed similar trends in some variables, though findings were not statistically significant. Interestingly, longitudinal analysis showed that APOE ε4 homozygotes had a higher annual change rate in the left HA gap compared to APOE ε2 homozygotes. Occlusion analysis saliency maps indicated that regions around the hippocampus, including the thalamus, pallidum, nearby cerebral cortex, and white matter, significantly contributed to the age prediction. The left HA gap emerges as a potential biomarker linked to the APOE genotype and an indicator of health.