Deep Learning Based Simulation of Amyloid-β accumulation
Poster No:
212
Submission Type:
Abstract Submission
Authors:
ByeongChang Jeong1, Daegyeom Kim1, Hyun-Ghang Jeong2, Cheol Han1
Institutions:
1Korea University, Sejong, Republic of Korea, 2Korea University College of Medicine, Seoul, Republic of Korea
First Author:
Co-Author(s):
Introduction:
Alzheimer's disease (AD) is a serious neurodegenerative condition marked by progressive brain tissue decline due to the accumulation of the toxic protein such as amyloid-β [1]. Recent studies [2, 3] have employed mathematical models to simulate amyloid-β accumulation in the brain, aiming to enhance understanding of AD development and progression. These models involve parameter estimation through observed data and an iterative process for parameter updates, but increased complexity can hinder this optimization. To overcome this, we converted the mathematical model into a deep learning model, combining multi-layer perceptron (MLP) and graph convolutional neural network (GCN) [4].
Methods:
We first modeled a mathematical model of amyloid-β accumulation based on the logistic growth equation and translated it into a deep learning model with MLP and GCN. The model accounts for amyloid-β generation, clearance, and spreading dynamics. The logistic growth model for each region is defined in Equation 1. In this equation, δ and γ is the generation and clearance coefficient, respectively. Akij defines the connection weight, capturing both local and transneuronal [5] spreading, where N(i) is the neighboring regions.
We translated it into a deep learning model, and the model is defined in Equation 2. Where X̂t+1 is the predicted accumulation level of all regions, Zt, and Gt correspond to clearance rate, and generation, respectively. Ãk represents the Laplacian normalized adjacency matrix of local and transneuronal A, respectively.
We used Alzheimer's Disease Neuroimaging Initiative (ADNI) data with 436 subjects to construct a graph capturing amyloid-β accumulation. In this graph, nodes represented region-of-interests (ROIs) from T1 structural magnetic resonance (MR) images, with node features represented accumulation levels from 18F-Florbetapir positron emission tomography (PET) scans. Edges indicated local and transneuronal connectivity based on structural characteristics and neuronal fibers from diffusion weighted MR images. Longitudinal data with more than two datapoints per subject were used, with 354 subjects for training and 82 subjects for testing.
We translated it into a deep learning model, and the model is defined in Equation 2. Where X̂t+1 is the predicted accumulation level of all regions, Zt, and Gt correspond to clearance rate, and generation, respectively. Ãk represents the Laplacian normalized adjacency matrix of local and transneuronal A, respectively.
We used Alzheimer's Disease Neuroimaging Initiative (ADNI) data with 436 subjects to construct a graph capturing amyloid-β accumulation. In this graph, nodes represented region-of-interests (ROIs) from T1 structural magnetic resonance (MR) images, with node features represented accumulation levels from 18F-Florbetapir positron emission tomography (PET) scans. Edges indicated local and transneuronal connectivity based on structural characteristics and neuronal fibers from diffusion weighted MR images. Longitudinal data with more than two datapoints per subject were used, with 354 subjects for training and 82 subjects for testing.
Results:
The proposed model predicted accumulation level after 2 years (Figure 1). It demonstrated a strong correlation with the real data (in test dataset, median = 0.8273, IQR = [0.7708, 0.8692]), outperforming the previous model(average 0.58) [3].
For interpretability, we examined the clearance (Zt) and generation (Gt) terms. We averaged each term over subjects and mapped the top 30% ROIs onto the brain (Figure 2). The brain regions with high clearance term were bilateral cuneus, paracentral, postcentral, precentral, supramarginal, frontal pole, insula, right superior frontal, and superior temporal. The regions with high generation term were bilateral entorhinal, fusiform, inferior temporal, lateral orbitofrontal, lingual, medial orbitofrontal, middle temporal, left posterior cingulate, and right parahippocampal.
Brain activity may play a crucial role in efficiency of amyloid-β clearance [6, 7]. Previous studies reported reduced metabolism in specific regions in early AD [8, 9]. The regions with high clearance in our model (Figure 2a) were matched with regions of normal metabolism, which exhibit relatively high brain activity. The regions with early AD amyloid-β accumulation were suspected to be connected with the default mode network and prefrontal network [10]. The identified regions (Figure 2b) also supported this.
For interpretability, we examined the clearance (Zt) and generation (Gt) terms. We averaged each term over subjects and mapped the top 30% ROIs onto the brain (Figure 2). The brain regions with high clearance term were bilateral cuneus, paracentral, postcentral, precentral, supramarginal, frontal pole, insula, right superior frontal, and superior temporal. The regions with high generation term were bilateral entorhinal, fusiform, inferior temporal, lateral orbitofrontal, lingual, medial orbitofrontal, middle temporal, left posterior cingulate, and right parahippocampal.
Brain activity may play a crucial role in efficiency of amyloid-β clearance [6, 7]. Previous studies reported reduced metabolism in specific regions in early AD [8, 9]. The regions with high clearance in our model (Figure 2a) were matched with regions of normal metabolism, which exhibit relatively high brain activity. The regions with early AD amyloid-β accumulation were suspected to be connected with the default mode network and prefrontal network [10]. The identified regions (Figure 2b) also supported this.
Conclusions:
We introduced a deep learning model to simulate amyloid-β accumulation achieving high predictive performance and interpretability. Although further investigation of each term's interpretability is needed, our model may help to understand the role of amyloid-β accumulation in AD progression.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Keywords:
Degenerative Disease
Machine Learning
Modeling
Other - simulation, amyloid-beta
1|2Indicates the priority used for review
Provide references using author date format
[2] Peraza, L. R., A. Díaz-Parra, O. Kennion, D. Moratal, J.-P. Taylor, M. Kaiser, R. Bauer and A. s. D. N. Initiative (2019). "Structural connectivity centrality changes mark the path toward Alzheimer's disease." Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring 11: 98-107.
[3] Kim, H.-R., P. Lee, S. W. Seo, J. H. Roh, M. Oh, J. S. Oh, S. J. Oh, J. S. Kim and Y. Jeong (2019). "Comparison of Amyloid β and tau spread models in Alzheimer’s disease." Cerebral Cortex 29(10): 4291-4302.
[4] Kipf, T. N. and M. Welling (2016). "Semi-supervised classification with graph convolutional networks." arXiv preprint arXiv:1609.02907.
[5] Nath, S., L. Agholme, F. R. Kurudenkandy, B. Granseth, J. Marcusson and M. Hallbeck (2012). "Spreading of neurodegenerative pathology via neuron-to-neuron transmission of β-amyloid." Journal of Neuroscience 32(26): 8767-8777.
[6] Mergenthaler, P., U. Lindauer, G. A. Dienel and A. Meisel (2013). "Sugar for the brain: the role of glucose in physiological and pathological brain function." Trends in neurosciences 36(10): 587-597.
[7] Ullah, R. and E. J. Lee (2023). "Advances in Amyloid-β Clearance in the Brain and Periphery: Implications for Neurodegenerative Diseases." Experimental Neurobiology 32(4): 216.
[8] Chételat, G., J. Arbizu, H. Barthel, V. Garibotto, I. Law, S. Morbelli, E. van de Giessen, F. Agosta, F. Barkhof and D. J. Brooks (2020). "Amyloid-PET and 18F-FDG-PET in the diagnostic investigation of Alzheimer's disease and other dementias." The Lancet Neurology 19(11): 951-962.
[9] Kantarci, K., B. F. Boeve, S. A. Przybelski, T. G. Lesnick, Q. Chen, J. Fields, C. G. Schwarz, M. L. Senjem, J. L. Gunte and C. R. Jack (2021). "FDG PET metabolic signatures distinguishing prodromal DLB and prodromal AD." NeuroImage: Clinical 31: 102754.
[10] Palmqvist, S., M. Schöll, O. Strandberg, N. Mattsson, E. Stomrud, H. Zetterberg, K. Blennow, S. Landau, W. Jagust and O. Hansson (2017). "Earliest accumulation of β-amyloid occurs within the default-mode network and concurrently affects brain connectivity." Nature communications 8(1): 1214.