Graph Neural Network: Enhancing Regression-based Prediction Performance for Depression Severity
Poster No:
1090
Submission Type:
Abstract Submission
Authors:
Ting-Ruei Lai1, Jol-Hsin Shih1, Cheng-Yu Ma1, Po-Hsien Lee2, Changwei Wu2, Ai-Ling Hsu1
Institutions:
1Chang Gung University, Taoyuan, Taiwan, 2Taipei Medical University, Taipei, Taiwan
First Author:
Co-Author(s):
Introduction:
Major depressive disorder (MDD) is a severe mental illness linked to abnormalities in brain functional networks, as measured by resting-state functional MRI (rs-fMRI)1. Recent studies highlight the superiority of graph neural networks (GNNs) over classical AI models in distinguishing MDDs from controls by leveraging connectome topology2,3. However, few AI models directly capture connectome changes linked to disorder severity4. Additionally, GNNs often underperform fully connected networks in correlating predicted and actual scores5. We hypothesize the this may stem from over-smoothing of critical features by the pooling layer in GNN architecture, which potentially reduces regression performance.
Methods:
The rs-fMRI dataset acquired from 3T scanners in the Rest-meta-MDD consortium6 were used to evaluate the performance of GNN and multi-layer perceptron (MLP) in predicting depression severity, as measured by HAMD. Data of 835 patients aged between 21 and 70 were parcellated using the Dosenbach atlas7 and region-wise functional connectivities were computed via Pearson correlation. Scanner effects were corrected using Combat harmonization8 and the dataset was split into training (80%), validation (10%), and testing (10%) subsets via age-based stratified sampling. Fig1 shows the GNN architecture that comprised three graph convolutional layers, reducing nodal features from 128 to 64, and then to 32. PairNorm regularization and ReLU activation were applied to the output of each convolutional layer. The resulting features were either flattened into 5120 features or averaged pooling into 160 features before being fed into a fully connected layer for HAMD prediction. For comparison, MLP model included three hidden layers with 223, 128, and 192 neurons, also applying BatchNorm regularization and ReLU activation5. Both models were trained using the Adam optimizer with L2 regularization and a dropout rate of 0.3. Performance was evaluated by Pearson correlation (r) and root mean square error (RMSE), with significance determined through paired t-tests on 30 independent resamples (p<0.05).
Results:
Fig2 shows that the GNN model achieved a significantly lower RMSE of 7.32±0.14 (mean±std. error of the mean) compared to MLP model (7.52±0.14) across 30 resamples, t=2.68, p=0.012. Regarding the linear relationship between true and predicted HAMD scores, no significant difference was found in the mean correlation between GNN (0.17±0.02) and MLP (0.20±0.02) models, t=-1.73, p=0.095, indicating the comparable performance of both models in capturing these relationships. In comparison to flatten-feature model, the pooling-feature model showed significantly superior accuracies (RMSE=7.12±0.13) but exhibited weaker correlations (0.13±0.02). Additionally, Fig2B further shows their correlations across the three subsets in the 13th of 30 resamples. In the training and testing subsets, the flatten-feature GNN achieved stronger positive correlations (r=0.93 and r=0.29, respectively) than the pooling-feature model (r=0.17 and r=0.11). However, in the validation subset, former exhibited a positive correlation (r=0.14), while the latter showed negative correlation (r=-0.13).
Conclusions:
This study evaluated the regression performance of GNN and MLP models in predicting HAMD scores from rs-fMRI functional connectomes. Using a 30-time resampling method, the GNN model demonstrated significantly superior predictive accuracies as indicated by a lower mean RMSE, compared to MLP model, while still maintaining comparable correlations. In comparison to the pooling-feature GNN, the flatten-feature GNN achieved enhanced predictive performance and proved to be more effective as a feature extractor in preserving the positive relationships between true and predicted HAMD scores. It successfully addressing over-smoothing issue, thereby improving regression performance.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia)
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural) 2
Keywords:
Affective Disorders
Data analysis
FUNCTIONAL MRI
Machine Learning
Other - Deep Learning; Regression; Major depressive disorder; Resting-state fMRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Qin, K., Lei, D., Pinaya, W. H. L., et al. (2022). Using graph convolutional network to characterize individuals with major depressive disorder across multiple imaging sites. EBioMedicine, 78, 103977.
3. Venkatapathy, S., Votinov, M., Wagels, L., et al. (2023). Ensemble graph neural network model for classification of major depressive disorder using whole-brain functional connectivity. Frontiers in Psychiatry, 14, 1125339.
4. Woo, C. W., Chang, L. J., Lindquist, M. A., et al. (2017). Building better biomarkers: Brain models in translational neuroimaging. Nature Neuroscience, 20(3), 365–377.
5. He, T., Kong, R., Holmes, A. J., et al. (2020). Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographics. NeuroImage, 206, 116276.
6. Chen, X., Lu, B., Li, H. X., et al. (2022). The DIRECT consortium and the REST-meta-MDD project: Towards neuroimaging biomarkers of major depressive disorder. Psychoradiology, 2(1), 32–42.
7. Dosenbach, N. U., Nardos, B., Cohen, A. L., et al. (2010). Prediction of individual brain maturity using fMRI. Science, 329(5997), 1358–1361.
8. Fortin, J. P., Cullen, N., Sheline, Y. I., et al. (2018). Harmonization of cortical thickness measurements across scanners and sites. NeuroImage, 167, 104–120.