MJNMF-GAT: Multi-Task Joint Non-Negative Matrix Factorization Graph Attention Network for Understanding Adolescent Neurodevelopment

Binish Patel, Tony W. Wilson, Julia M. Stephen, Vince D. Calhoun, Yu-Ping WangIEEE Access 2025 · DOI

TL;DR

This paper introduces MJNMF-GAT, a novel deep learning framework that integrates multi-task fMRI data to better understand adolescent brain development. By combining non-negative matrix factorization with graph attention networks, the model achieves superior accuracy in predicting age and identifies key brain subnetworks involved in this developmental process.

What Problem Does This Paper Address?

Adolescence is a critical period of rapid brain development, characterized by significant changes in structure and function. Understanding these changes is crucial for identifying typical developmental trajectories and potential neurodevelopmental disorders. Functional magnetic resonance imaging (fMRI) is a powerful tool for studying brain activity, but analyzing the complex, high-dimensional, and often noisy fMRI data, especially from multiple tasks, presents a significant challenge. Existing methods often struggle to effectively integrate information from different tasks or provide interpretable insights into the underlying brain connectivity patterns. This paper aims to address these challenges by developing a model that can leverage multi-task fMRI data for more accurate and interpretable analysis of adolescent neurodevelopment.

Methods

The core of this work is the MJNMF-GAT (Multi-Task Joint Non-Negative Matrix Factorization Graph Attention Network) framework. It combines two key components:

  • Joint Non-Negative Matrix Factorization (JNMF): This technique is used to decompose multiple related fMRI data matrices (from different tasks) into a common set of non-negative factors. This allows the model to extract shared latent features or brain networks that are consistent across tasks, while also capturing task-specific coefficients that represent individual variations. This helps in dimensionality reduction and identifying robust underlying brain structures.
  • Graph Attention Networks (GATs): GATs are employed to model the brain's functional connectivity as a graph, where brain regions are nodes and connections are edges. The attention mechanism in GATs allows the model to adaptively weigh the importance of different connections, focusing on the most relevant ones for the prediction task. This is particularly useful for handling the noisy and complex nature of fMRI data.

The MJNMF-GAT framework integrates these two by using the shared basis matrix from JNMF as input to the GAT. This allows the model to learn from both the shared patterns across tasks and the specific connectivity within each task, leading to improved prediction and interpretability. For interpretability, the authors also utilize GNNExplainer, a model-agnostic method to identify the most influential nodes and edges contributing to the model's predictions.

Datasets

The study utilized data from the Philadelphia Neurodevelopmental Cohort (PNC). This dataset includes fMRI data from over 800 healthy participants aged 8 to 22 years, collected during two tasks:

  • nback fMRI: A working memory paradigm.
  • emoid fMRI: An emotion recognition paradigm.

A subset of 622 participants with data from both paradigms was selected for this study. The participants were further divided into five age groups to represent different stages of adolescent development.

Key Insights and Ideas

  • Multi-task learning is crucial for adolescent neurodevelopment: Integrating data from different fMRI tasks (working memory and emotion recognition) significantly improves the model's ability to predict age and understand brain development compared to using a single task. This suggests that different cognitive processes capture complementary aspects of brain maturation.
  • The brain's functional connectivity is a powerful "fingerprint" for age: The model successfully uses functional connectivity patterns as a basis for predicting age, highlighting how brain networks evolve predictably during adolescence.
  • GNNExplainer reveals critical subnetworks: The interpretability component of the model successfully identified specific subnetworks and connections that are most important for predicting age, offering insights into which brain regions are undergoing significant developmental changes.
  • Adolescence is not a monolithic stage: The study shows that the importance of specific functional connections for age prediction varies across different stages of adolescence, indicating a dynamic and nuanced developmental process.

Key Results

The MJNMF-GAT model demonstrated superior performance in age prediction on the PNC dataset, achieving:

  • Root Mean Square Error (RMSE): 1.9212 ± 0.1742
  • Mean Absolute Error (MAE): 1.5368 ± 0.1469
  • Correlation: 0.8032 ± 0.0469

These results were significantly better than those of several other state-of-the-art models, including LR, MLP, GCN, GAT, and others (as detailed in Table 5). Ablation studies confirmed the importance of both JNMF and the multi-task approach for optimal performance. The GNNExplainer analysis identified specific subnetworks (e.g., Default Mode Network, Visual System, Fronto-parietal Task Control) and their connections that are critical for age prediction across different adolescent stages.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Static connectivity measures: The current graph attention network uses static measures like Pearson correlation for features and cosine similarity for edge weights. This may not fully capture the dynamic relationships within brain networks.
  • Focus on task-based fMRI: The study primarily focuses on task-based fMRI data.

For future work, the authors plan to:

  • Explore alternative feature learning techniques: Investigate methods that can capture more dynamic relationships and reduce redundancy in high-order connections.
  • Extend to diverse data and populations: Analyze differences in brain connectivity across various cognitive tasks and clinical populations to assess the model's generalizability.
  • Analyze broader context of individual variability: Move beyond just task-based fMRI to understand individual differences in brain function more comprehensively.

Summary updated

BibTeX
@article{patel2025mjnmf,
  author = {Binish Patel and Tony W. Wilson and Julia M. Stephen and Vince D. Calhoun and Yu-Ping Wang},
  title = {{MJNMF-GAT: Multi-Task Joint Non-Negative Matrix Factorization Graph Attention Network for Understanding Adolescent Neurodevelopment}},
  journal = {IEEE Access},
  year = {2025},
  doi = {10.1109/access.2025.3601649},
}
Figure