Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session C

Poster A8 - Interpretable deep learning reveals differences in brain aging according to genetic risk of Alzheimer's Disease

Friday, October 25, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Interpretable deep neural networks (DNNs) estimate brain age from magnetic resonance images (MRI) to produce ‘saliency maps’. When a DNN is trained to predict brain age (BA), saliency maps represent neuroanatomical features according to their contributions to BA estimation. Polygenic risk scores (PRSs) calculated from common genetic risk variants reveal differences in the risk of mild cognitive impairment (MCI) and Alzheimer’s disease (AD) [2]. We grouped patients by high- vs. low-PRS (PRS > 1.25 and PRS < 0.75, respectively) and compared their average saliencies to reveal differences in brain aging according to genetic risk of AD. This revealed differences in neuroanatomic brain aging according to PRS, which can complement preclinical screenings for risk of cognitive decline.

Materials and

Methods:  Data Acquisition
Genetic data and MRI scans are obtained from the UK Biobank database for low-PRS (N=6835, 3562 females, mean age = 55.3 years, SD = 7.6 years) and high-PRS (N=2311, 1208 females, mean age = 55.4 years, SD = 7.5 years) participants.
Saliency Maps
Saliency is an estimate of each brain voxel's contribution to the DNN's BA estimate. Saliency is computed as the partial derivative of the CNN output (BA) with respect to the input brain MRI. We use TensorFlow's gradientTape context manager to measure the gradient of prediction (BA) with respect to input MRI intensity at each voxel.
We project saliency volumes onto the brain's surface to summarize 3D information on a 2D surface. Our in-house projection tool maps saliency volumes onto FreeSurfer ico7 mesh-based surfaces by averaging the voxels that connect a white matter vertex to its corresponding pial vertex [4]. Volume-to-surface mapping averages and normalizes a cylinder of voxels inside the gray matter ribbon to create a saliency probability map across the cortex.
Analysis
Saliency probability maps are averaged across low- and high-PRS groups. The difference in saliency between groups is divided by the saliency of the high-PRS group to yield the fractional difference between groups. The resulting saliency map is smoothed using a FreeSurfer software package and plotted [3].

Results, Conclusions, and Discussions: Results and Discussion
Figure 1 displays average saliency differences between high-PRS and low-PRS groups on the cortex. High-PRS have, on average, higher saliency in the bilateral frontal lobes and the right parietal lobe. The medial temporal lobe, a structure associated with the onset of Alzheimer’s, presents higher saliency in the high-PRS group. The asymmetry between the left- and right-hemispheres could be a result of handedness, with the left-hemisphere of both groups presenting higher saliency values. Nonetheless, since PRS is a metric based on genetic information, its accuracy is limited by the genetic loci it samples.
Conclusion
We compare saliency across subjects with high and low PRS for AD. We find higher saliency in the frontal lobes, medial temporal lobes, and in the right parietal lobe in subjects at greater risk of AD. Future studies will examine saliency differences across subjects with modifiable AD risk factors such as BMI and alcohol consumption.

Acknowledgements (Optional):  [1] Beekly, D. L., et al. (2004). "The national Alzheimer's coordinating center (NACC) database: an Alzheimer's disease database." Alzheimer Disease & Associated Disorders 18(4): 270-277.
[2] de Rojas, I., Moreno-Grau, S., Tesi, N. et al. (2021). “Common variants in Alzheimer’s disease and risk stratification by polygenic risk scores.” Nat Commun 12, 3417.
[3] Hagler, D. J., Jr, Saygin, A. P., & Sereno, M. I. (2006). “Smoothing and cluster thresholding for cortical surface-based group analysis of fMRI data.” NeuroImage, 33(4), 1093–1103.
[4] Yin, C., Imms, P., Cheng, M., Amgalan, A., Chowdhury, N. F., Massett, R. J., et al. (2023). Anatomically interpretable deep learning of brain age captures domain-specific cognitive impairment. Proceedings of the National Academy of Sciences, 120(2), e2214634120.