CoRE-BOLD: Cross-Domain Robust and Equitable Ensemble for BOLD Signal Analysis

Vipul Kumar Singh, Jyotismita Barman, Sandeep Kumar, Jayadeva Jayadeva
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:961-975, 2025.

Abstract

In current neuroimaging studies aimed at analysing BOLD signals, the focus has primarily been on correlation-based features derived from time series data. Considering Major Depressive Disorder (MDD), a widespread psychiatric condition, poses a complex and poorly understood pathology. Recent research has increasingly linked MDD to disruptions in brain connectivity, as observed through functional Magnetic Resonance Imaging (fMRI). Identifying the brain regions associated with neurological disorders and cognitive processes remains a central objective in neuroimaging studies. While Graph Neural Networks (GNNs) have been widely employed to extract disease-relevant information from fMRI data, existing methods face significant limitations. These limitations include neglecting the frequency-domain characteristics of neuronal interactions, inadequately incorporating non-imaging biomarkers such as sex and age, and paying insufficient attention to bias and model stability, which leaves models prone to small perturbations. We introduce CoRE-BOLD, a unified framework addressing these gaps for MDD diagnosis in this study. CoRE-BOLD employs an ensemble of stacked networks that learn complementary representations from both correlation- and coherence-based functional connectivities. To further improve the model, we enforce orthonormality constraints on the graph convolutional filters to enhance intra-network diversity and apply a diversity-maximizing regularizer for inter-network diversity. Unlike previous studies, which incorporate non-imaging sensitive attributes as biomarkers but inadvertently introduce bias, CoRE-BOLD mitigates this through a prejudice remover regularizer, promoting fairness in representation learning across both underrepresented and favored groups. Our experimental evaluation on the REST-meta-MDD dataset demonstrates the efficacy of CoRE-BOLD as a robust and fair framework for BOLD signal analysis in MDD detection, positioning it as a promising solution for real-world medical applications. Source code of CoRE-BOLD is freely available at: https://github.com/shashivipul/CoRE-BOLD.git.

Cite this Paper


BibTeX
@InProceedings{pmlr-v259-singh25a, title = {CoRE-BOLD: Cross-Domain Robust and Equitable Ensemble for BOLD Signal Analysis}, author = {Singh, Vipul Kumar and Barman, Jyotismita and Kumar, Sandeep and Jayadeva, Jayadeva}, booktitle = {Proceedings of the 4th Machine Learning for Health Symposium}, pages = {961--975}, year = {2025}, editor = {Hegselmann, Stefan and Zhou, Helen and Healey, Elizabeth and Chang, Trenton and Ellington, Caleb and Mhasawade, Vishwali and Tonekaboni, Sana and Argaw, Peniel and Zhang, Haoran}, volume = {259}, series = {Proceedings of Machine Learning Research}, month = {15--16 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v259/main/assets/singh25a/singh25a.pdf}, url = {https://proceedings.mlr.press/v259/singh25a.html}, abstract = {In current neuroimaging studies aimed at analysing BOLD signals, the focus has primarily been on correlation-based features derived from time series data. Considering Major Depressive Disorder (MDD), a widespread psychiatric condition, poses a complex and poorly understood pathology. Recent research has increasingly linked MDD to disruptions in brain connectivity, as observed through functional Magnetic Resonance Imaging (fMRI). Identifying the brain regions associated with neurological disorders and cognitive processes remains a central objective in neuroimaging studies. While Graph Neural Networks (GNNs) have been widely employed to extract disease-relevant information from fMRI data, existing methods face significant limitations. These limitations include neglecting the frequency-domain characteristics of neuronal interactions, inadequately incorporating non-imaging biomarkers such as sex and age, and paying insufficient attention to bias and model stability, which leaves models prone to small perturbations. We introduce CoRE-BOLD, a unified framework addressing these gaps for MDD diagnosis in this study. CoRE-BOLD employs an ensemble of stacked networks that learn complementary representations from both correlation- and coherence-based functional connectivities. To further improve the model, we enforce orthonormality constraints on the graph convolutional filters to enhance intra-network diversity and apply a diversity-maximizing regularizer for inter-network diversity. Unlike previous studies, which incorporate non-imaging sensitive attributes as biomarkers but inadvertently introduce bias, CoRE-BOLD mitigates this through a prejudice remover regularizer, promoting fairness in representation learning across both underrepresented and favored groups. Our experimental evaluation on the REST-meta-MDD dataset demonstrates the efficacy of CoRE-BOLD as a robust and fair framework for BOLD signal analysis in MDD detection, positioning it as a promising solution for real-world medical applications. Source code of CoRE-BOLD is freely available at: https://github.com/shashivipul/CoRE-BOLD.git.} }
Endnote
%0 Conference Paper %T CoRE-BOLD: Cross-Domain Robust and Equitable Ensemble for BOLD Signal Analysis %A Vipul Kumar Singh %A Jyotismita Barman %A Sandeep Kumar %A Jayadeva Jayadeva %B Proceedings of the 4th Machine Learning for Health Symposium %C Proceedings of Machine Learning Research %D 2025 %E Stefan Hegselmann %E Helen Zhou %E Elizabeth Healey %E Trenton Chang %E Caleb Ellington %E Vishwali Mhasawade %E Sana Tonekaboni %E Peniel Argaw %E Haoran Zhang %F pmlr-v259-singh25a %I PMLR %P 961--975 %U https://proceedings.mlr.press/v259/singh25a.html %V 259 %X In current neuroimaging studies aimed at analysing BOLD signals, the focus has primarily been on correlation-based features derived from time series data. Considering Major Depressive Disorder (MDD), a widespread psychiatric condition, poses a complex and poorly understood pathology. Recent research has increasingly linked MDD to disruptions in brain connectivity, as observed through functional Magnetic Resonance Imaging (fMRI). Identifying the brain regions associated with neurological disorders and cognitive processes remains a central objective in neuroimaging studies. While Graph Neural Networks (GNNs) have been widely employed to extract disease-relevant information from fMRI data, existing methods face significant limitations. These limitations include neglecting the frequency-domain characteristics of neuronal interactions, inadequately incorporating non-imaging biomarkers such as sex and age, and paying insufficient attention to bias and model stability, which leaves models prone to small perturbations. We introduce CoRE-BOLD, a unified framework addressing these gaps for MDD diagnosis in this study. CoRE-BOLD employs an ensemble of stacked networks that learn complementary representations from both correlation- and coherence-based functional connectivities. To further improve the model, we enforce orthonormality constraints on the graph convolutional filters to enhance intra-network diversity and apply a diversity-maximizing regularizer for inter-network diversity. Unlike previous studies, which incorporate non-imaging sensitive attributes as biomarkers but inadvertently introduce bias, CoRE-BOLD mitigates this through a prejudice remover regularizer, promoting fairness in representation learning across both underrepresented and favored groups. Our experimental evaluation on the REST-meta-MDD dataset demonstrates the efficacy of CoRE-BOLD as a robust and fair framework for BOLD signal analysis in MDD detection, positioning it as a promising solution for real-world medical applications. Source code of CoRE-BOLD is freely available at: https://github.com/shashivipul/CoRE-BOLD.git.
APA
Singh, V.K., Barman, J., Kumar, S. & Jayadeva, J.. (2025). CoRE-BOLD: Cross-Domain Robust and Equitable Ensemble for BOLD Signal Analysis. Proceedings of the 4th Machine Learning for Health Symposium, in Proceedings of Machine Learning Research 259:961-975 Available from https://proceedings.mlr.press/v259/singh25a.html.

Related Material