Machine learning-based fMRI feature extraction of major depressive disorders
Poster No:
1441
Submission Type:
Abstract Submission
Authors:
Ye-Eun Kim1, Jiwon Lee2, Mikhail Votinov3, Lisa Wagels4, Ute Habel5, Sun-Young Kim6, Munseob Lee7, Han-Gue Jo6
Institutions:
1Kunsan National University, Kunsan, Jeollabuk do, 2Kunsan National University, Gunsan-si, Jeollabuk-do, 3Institute of Neuroscience and Medicine (INM-10), Research Centre Jülich, Jülich, Germany, 4RWTH Aachen University Hospital, Aachen, Germany, 5Department of Psychiatry, Psychotherapy and Psychosomatics, RWTH Aachen University Hospital, Aachen, North Rhine-Westphalia, 6Kunsan National University, Gunsan, Jeollabuk-do, 7Electronics and Communication Research Institute (ETRI), Deajeon, Deajeon
First Author:
Co-Author(s):
Mikhail Votinov
Institute of Neuroscience and Medicine (INM-10), Research Centre Jülich
Jülich, Germany
Institute of Neuroscience and Medicine (INM-10), Research Centre Jülich
Jülich, Germany
Ute Habel
Department of Psychiatry, Psychotherapy and Psychosomatics, RWTH Aachen University Hospital
Aachen, North Rhine-Westphalia
Department of Psychiatry, Psychotherapy and Psychosomatics, RWTH Aachen University Hospital
Aachen, North Rhine-Westphalia
Introduction:
Literature has demonstrated the applicability of graph neural network (GNN) in analyzing the connectome derived from functional magnetic resonance imaging (fMRI), particularly in identifying alterations within the functional networks of major depressive disorder (MDD) [1,2]. Building upon these findings, the current study employs a data-driven machine learning approach with GNNs to uncover the fundamental features characterizing the functional brain network of MDD.
Methods:
Resting-state functional MRI scans from a dataset of 821 MDDs were used to train a GNN model[3]. For constructing the graph, a whole-brain functional connectivity matrix was utilized, representing Fisher transformed correlation coefficients among BOLD time series across 160 Dosenbachs' atlas-based cortical brain regions[4]. Since default mode network frequently reported to be different in MDD[5], thirty-four regions of them were taken to represent default mode network within the matrix. The establishment of edges was achieved through the application of a k-nearest neighbors (KNN) algorithm, with the node feature set as a vector reflecting nodal functional connectivity. To decode the fundamental features characterizing the MDD brain network, a graph autoencoder (GAE) framework [6] was employed. GAE operates through unsupervised learning on graph-structured data, utilizing a variational autoencoder to model the latent space as a lower-dimensional vector, serving as a representation of the input dataset. Once the GAE model was trained, the connections within the input matrix significantly influencing the latent vector were identified.
Results:
Figure 1 illustrates the GAE architecture with seven fully-connected (FC) layers in the encoding phase, and three FC layers and four graph convolutional network (GCN) layers in the decoding phase. The first step involved training the GAE to derive the low-dimensional representation vector Z for the provided brain networks. This training process aimed to minimize the disparity between the input X and its corresponding output X'. Upon testing the training loss across various latent dimensions, it was found that eight dimensions yielded the lowest training loss. Subsequently, the investigation focused on identifying the top salient connections contributing to the latent vector of the trained GAE. By altering the connectivity strength of each input X ij , the variations in the latent vector concerning specific connections between brain regions i and j were inferred. The assessment involved calculating the average squared variance (MSE) between the latent vectors encoded by the input X and those encoded by X ij . These MSE values were then sorted, revealing the connections with higher MSE values, indicating a more substantial impact on the latent space. Essentially, these connections signify significant features that could characterize the given input X. Figure 2 demonstrates this sorting process. It revealed a cluster of three connections above 0.175 MSE (ventromedial prefrontal cortex and left superior frontal gyrus, medial prefrontal cortex and posterior cingulate cortex, and medial prefrontal cortex and left inferior temporal gyrus), which are the top features potentially characterizing the brain networks associated with MDD.
·Figure 1 . The architecture of graph auto encoder. Abbreviations: GCN, graph convolutional network; FC, fully- connected layer; Z, latent vector; MDD, major depressive disorder.
·Figure 2 . MSE sorted functional connectivity. The higher the MSE values implies the higher impact on the latent vector Z of the trained GAE.
Conclusions:
This study employed a data-driven GNN approach to identify the fundamental features characterizing the brain networks associated with MDD. Utilizing the autoencoder method, GAE model was trained to minimize training loss, allowing the latent vector to serve as a concise representation of the MDD brain networks. Based on data-driven analysis of a large dataset related to MDD, these findings indicate potential brain biomarkers that could be pivotal in clinical diagnosis and formulating treatment plans for psychiatric disorders.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Keywords:
FUNCTIONAL MRI
Machine Learning
Psychiatric Disorders
1|2Indicates the priority used for review
Provide references using author date format
2. Venkatapathy, S., Votinov, M., Wagels, L., Kim, S., Lee, M., Habel, U., ... & Jo, H. G. (2023). Ensemble graph neural network model for classification of major depressive disorder using whole-brain functional connectivity. Frontiers in Psychiatry, 14, 1125339.
3. Yan, C. G., Chen, X., Li, L., Castellanos, F. X., Bai, T. J., Bo, Q. J., ... & Zang, Y. F. (2019). Reduced default mode network functional connectivity in patients with recurrent major depressive disorder. Proceedings of the National Academy of Sciences, 116(18), 9078-9083.
4. Dosenbach, N. U., Nardos, B., Cohen, A. L., Fair, D. A., Power, J. D., Church, J. A., ... & Schlaggar, B. L. (2010). Prediction of individual brain maturity using fMRI. Science, 329(5997), 1358-1361.
5. Kaiser, R. H., Andrews-Hanna, J. R., Wager, T. D., & Pizzagalli, D. A. (2015). Large-scale network dysfunction in major depressive disorder: a meta-analysis of resting-state functional connectivity. JAMA psychiatry, 72(6), 603-611.
6. Liu, M., Zhang, Z., & Dunson, D. B. (2021). Graph auto-encoding brain networks with applications to analyzing large-scale brain imaging datasets. Neuroimage, 245, 118750.