Graph Convolutional Network Model for Discrimination of Major Depressive Disorder.
Poster No:
578
Submission Type:
Abstract Submission
Authors:
Kei Kamiya1, Miyuki Tajima1, Yuki Kobayashi1, Yoshimi Arai1, Risa Ogata1, Mika Yamagishi1, Hana Nishida1, Shun Kudo1, Akihiro Takamiya1,2,3, Nariko Katayama1, Bun Yamagata1, Masaru Mimura1, Jinichi Hirano1
Institutions:
1Department of Neuropsychiatry, Keio University school of Medicine, Shinjuku, Tokyo, 2Neuropsychiatry, Department of Neurosciences, Leuven Brain Institute, KU Leuven, Belgium, 3Hills Joint Research Laboratory for Future Preventive Medicine and Wellness, Keio University School, Tokyo, Japan
First Author:
Co-Author(s):
Akihiro Takamiya
Department of Neuropsychiatry, Keio University school of Medicine|Neuropsychiatry, Department of Neurosciences, Leuven Brain Institute|Hills Joint Research Laboratory for Future Preventive Medicine and Wellness, Keio University School
Shinjuku, Tokyo|KU Leuven, Belgium|Tokyo, Japan
Department of Neuropsychiatry, Keio University school of Medicine|Neuropsychiatry, Department of Neurosciences, Leuven Brain Institute|Hills Joint Research Laboratory for Future Preventive Medicine and Wellness, Keio University School
Shinjuku, Tokyo|KU Leuven, Belgium|Tokyo, Japan
Introduction:
Major depressive disorder (MDD) is considered one of the most socially and economically burdensome illnesses. A recent study of the functional connectome has revealed quantitative differences in the brain networks of individuals with MDD, contributing to identifying potential biological markers (Chai et al., 2023). Machine learning has emerged as a pivotal tool in medical research, with numerous models developed for MDD (Qin et al., 2018). However, these models often face limitations, such as limited sample sizes and a scarcity of advanced models employing deep learning techniques. Addressing these gaps, our study aims to develop a deep learning-based discrimination model for MDD utilizing Functional Connectivity (FC) data derived from a dataset comprising over a thousand samples.
Methods:
Dataset：
We have used the MDD who was diagnosed by DSM-V and Healthy Control (HC) images from the following data set: Longitudinal MRI study Identifying the Neural Substrates of Remission/Recovery in Mood Disorders (L/R) and the Strategic International Brain Science Research Promotion Program (Brain/MINDS Beyond) MRI data set.
Image acquisition and preprocessing：
We used resting state functional MRI (rsfMRI) images, which mainly consist of two different scan protocols (non-multiband based SRPB protocol (TR = 2500 msec) (https://bicr.atr.jp/rs-fmri-protocol-2/) and multiband-based HARP protocol (TR = 800 msec) (https://bicr.atr.jp/rs-fmri-protocol-2/)). rsfMRI was obtained from six different scanners. SRPB protocol images were preprocessed by fMRI prep (Esteban et al., 2019), and HARP protocol images were preprocessed by the HCP pipeline. (Glasser, et al. 2013) After preprocessing, we extracted the FC matrix based on the AAL2 (Automated Anatomical Labeling) atlas using Nilearn (https://nilearn.github.io/stable/index.html). To harmonize the scanner effect, NeuroCombat was applied for FC data (Sun et al., 2022).
Model building：
Using the preprocessed FC matrix, we built a binary prediction model for discriminating the MDD or HC. We used a Graph Convolutional Networks (GCN) model, a deep learning technique on graph-structured data. In this study, we conducted a stratified 10-fold cross-validation. For a 10-fold training/test split, the model was fit to the training data, and the predictive value was assessed using the test data over all splits (10 times). Balanced accuracy (Average of True Positive Rate and True Negative Rate), accuracy, sensitivity, specificity, and Area Under Curve (AUC) value were calculated to evaluate the overall results. The Pytorch Geometric extension library was used for model building. The flow of analysis was shown in Figure 1.
We have used the MDD who was diagnosed by DSM-V and Healthy Control (HC) images from the following data set: Longitudinal MRI study Identifying the Neural Substrates of Remission/Recovery in Mood Disorders (L/R) and the Strategic International Brain Science Research Promotion Program (Brain/MINDS Beyond) MRI data set.
Image acquisition and preprocessing：
We used resting state functional MRI (rsfMRI) images, which mainly consist of two different scan protocols (non-multiband based SRPB protocol (TR = 2500 msec) (https://bicr.atr.jp/rs-fmri-protocol-2/) and multiband-based HARP protocol (TR = 800 msec) (https://bicr.atr.jp/rs-fmri-protocol-2/)). rsfMRI was obtained from six different scanners. SRPB protocol images were preprocessed by fMRI prep (Esteban et al., 2019), and HARP protocol images were preprocessed by the HCP pipeline. (Glasser, et al. 2013) After preprocessing, we extracted the FC matrix based on the AAL2 (Automated Anatomical Labeling) atlas using Nilearn (https://nilearn.github.io/stable/index.html). To harmonize the scanner effect, NeuroCombat was applied for FC data (Sun et al., 2022).
Model building：
Using the preprocessed FC matrix, we built a binary prediction model for discriminating the MDD or HC. We used a Graph Convolutional Networks (GCN) model, a deep learning technique on graph-structured data. In this study, we conducted a stratified 10-fold cross-validation. For a 10-fold training/test split, the model was fit to the training data, and the predictive value was assessed using the test data over all splits (10 times). Balanced accuracy (Average of True Positive Rate and True Negative Rate), accuracy, sensitivity, specificity, and Area Under Curve (AUC) value were calculated to evaluate the overall results. The Pytorch Geometric extension library was used for model building. The flow of analysis was shown in Figure 1.
Results:
We finally utilized 430 MDD and 586 HC images for model building. Our final models showed 63.6 ± 5.1% balanced accuracy. The accuracy was 65.44 ± 5.01 %, sensitivity was 51.2 ± 11.2 %, specificity was 75.9 ± 10.7 %, and the AUC value was 0.636, respectively (Figure 2).
Conclusions:
The model trained using GCN showed an accuracy of 65.44%, about 8% higher than the baseline (57.6%, 586/1016). To the best of our knowledge, only one prior study has created a disease discrimination model for MDD using a sample size exceeding a thousand (Qin et al., 2018). While the accuracy of our study is lower compared to this previous research (Qin et al., 2018)., this may be due to our use of two imaging protocols with different temporal resolutions.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural) 2
fMRI Connectivity and Network Modeling
Novel Imaging Acquisition Methods:
BOLD fMRI
Keywords:
FUNCTIONAL MRI
Machine Learning
MRI
1|2Indicates the priority used for review
Provide references using author date format
Kun Qin(2022), 'Using graph convolutional network to characterize individuals with major depressive disorder across multiple imaging sites.', eBioMedicine, vol. 78
Esteban, O.(2019), 'fMRIPrep: a robust preprocessing pipeline for functional MRI.', Nature Methods, vol. 16, pp. 111–116
Glasser, M. F.(2013), 'The minimal preprocessing pipelines for the Human Connectome Project.', Neuroimage, vol. 80, pp. 105–124
Sun(2022), 'A comparison of methods to harmonize cortical thickness measurements across scanners and sites.' Neuroimage, vol. 261, pp. 1-19