Foundation Model for Functional MRI:Causal Discovery & Graph Transformer for Brain Function Analysis
Poster No:
1544
Submission Type:
Abstract Submission
Authors:
Moein Khajehnejad1, Adeel Razi2
Institutions:
1Monash University, Melborune, VIC, 2Monash University, Melbourne, VIC
First Author:
Co-Author:
Introduction:
Functional MRI (fMRI) has played central role in understanding brain function. Gleaning more comprehensive understanding of brain function is challenging because it requires integrating inherently noisy datasets and complex inter-regional dependencies. Traditional methods relying on functional connectivity do not provide mechanistic understanding as they don't capture causal dynamics and generalizable brain representations. We propose a novel framework combining causal discovery and graph-based representation learning into a foundation model. It extracts transferable features using causal graph construction and transformer-based encoders, addressing key gaps in fMRI analysis for diverse downstream tasks (e.g. case-control classification and behaviour prediction).
Methods:
Our method consists of i) Causal Graph Extraction, ii) Feature Extraction using Graph Encoding, and iii) Foundation Model Training (see Figure 1). During causal graph extraction, an encoder-decoder architecture inspired by Amortized Causal Discovery (ACD) [2] processes brain regional time-series data using a graph neural network to predict a subject-specific causal graph of directed relationships (known as a Directed Acyclic Graph), with variational inference and sparsity penalties ensuring interpretability. The feature extraction step encodes these causal graphs into node embeddings using a transformer encoder [1], where masked self-attention focuses on connected nodes, and positional encodings capture higher-order graph dependencies, integrating both local and global information. In the foundation model training step, the framework is pretrained on large-scale datasets selected from the pool of HCP, ABCD, and UK Biobank repositories of resting state and task-based functional MRI to learn generalized and transferable node embeddings, which are fine-tuned for downstream tasks such as behaviour prediction and cognitive state classification.
Results:
In a simpler experiment of using functional connectivity graphs, the transformer encoder is capable of outperforming traditional methods in tasks like age or behaviour prediction and cognitive state classification and achieving an improvement of 5-10% in classification accuracy. When incorporating causal graphs constructed by ACD, we are also able to achieve enhanced model interpretability and further gains in performance, particularly in datasets with significant noise or latent confounders. Further comparative analysis is expected to demonstrate the scalability and robustness of the proposed method across datasets, with AUROC values exceeding state-of-the-art approaches in the range of 0.1-0.2 in various tasks.
Conclusions:
This work introduces a scalable and generalizable framework for fMRI analysis that integrates causal graph discovery and graph transformer-based representation learning. By constructing causal graphs and leveraging graph attention mechanisms, the foundation model extracts meaningful brain regional features while capturing global dynamics. Pretrained on large multi-site datasets, the model is adaptable to various downstream tasks, offering state-of-the-art performance. This approach bridges the gap between causal inference and graph-based machine learning, providing a robust tool for neuroscience and clinical applications.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 2
Methods Development 1
Keywords:
Computational Neuroscience
Design and Analysis
FUNCTIONAL MRI
Machine Learning
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?
Provide references using APA citation style.
Löwe, S. (2022, June). Amortized causal discovery: Learning to infer causal graphs from time-series data. In Conference on Causal Learning and Reasoning (pp. 509-525). PMLR.