An objective quantitative diagnosis of depression using a local-to-global multimodal fusion graph neural network

Shuyu Liu, Jingjing Zhou, Xuequan Zhu, Ya Zhang, Xinzhu Zhou, Shaoting Zhang, Zhi Yang, Ziji Wang, Ruoxi Wang, Yizhe Yuan, Xin Fang, Xiongying Chen, Yanfeng Wang, Ling Zhang, Gang Wang, Cheng JinPatterns 2024 · DOI

TL;DR

Researchers have developed a novel AI system, LGMF-GNN, that integrates functional MRI, structural MRI, and electronic health records to diagnose Major Depressive Disorder (MDD) with unprecedented accuracy. This system not only achieves state-of-the-art performance but also offers interpretable insights into potential biomarkers for MDD, paving the way for more objective and personalized clinical diagnoses.

What Problem Does This Paper Address?

Major Depressive Disorder (MDD) is a complex and pervasive mental health challenge, yet its diagnosis heavily relies on subjective clinical interviews and self-reported scales. This often leads to low detection rates and misdiagnosis. While neuroimaging has revealed some brain alterations associated with MDD, these findings often have small effect sizes and limited individual predictive value. Furthermore, current AI approaches often fail to integrate the multifaceted nature of MDD, neglecting the interplay of functional, structural, and demographic factors, and lack interpretability to identify diagnostic markers. There's a critical need for objective, quantitative diagnostic tools that can leverage multimodal data and provide insights into the underlying mechanisms of MDD.

Methods

The core innovation of this paper is the Local-Global Multimodal Fusion Graph Neural Network (LGMF-GNN). This system employs a hierarchical approach:

  • Local ROI GNN: This component focuses on individual brain regions. It uses a bidirectional GRU (bi-GRU) encoder to process BOLD time series from Regions of Interest (ROIs) and then a graph generator creates a learnable adjacency matrix for each subject. A GNN predictor then aggregates this information to generate local graph embeddings and classification results. This allows for a fine-grained analysis of local brain circuitry.
  • Global Subject GNN: This component takes a population-level view. It constructs three types of subject graphs: functional (based on local GNN embeddings), anatomical (based on T1 MRI features), and demographic (based on non-imaging data like age and sex). These graphs are then fused using a multimodal attention mechanism within a Snowball GCN architecture. This allows for the integration of information across individuals and modalities.

Key aspects of the LGMF-GNN include:

  • Multimodal Fusion: Seamless integration of fMRI, sMRI, and demographic data.
  • Hierarchical Learning: Combining local (ROI) and global (subject) perspectives.
  • Learnable Graph Structure: The adjacency matrix is learned, allowing for dynamic adaptation.
  • Interpretability Module: Designed to identify potential digital biomarkers.
  • Site Effect Mitigation: Techniques like ComBat harmonization and domain adversarial training are employed to address variations across different data collection sites.

Datasets

The study utilized a combination of large-scale, multi-center datasets:

  • Training and Internal Validation:
    • SRPBS dataset: 229 MDD patients vs. 228 healthy controls (HCs) from 6 sites.
  • REST-meta-MDD dataset: 814 MDD patients vs. 756 HCs from 16 sites.
  • External Independent Testing:
    • Anding dataset: 196 MDD patients vs. 177 HCs from 1 site.
  • OpenNeuro dataset: 21 MDD patients vs. 21 HCs from 1 site in Russia (representing a different population).

Collectively, these datasets comprise 2,442 participants from 24 institutions.

Key Insights and Ideas

  • The Power of Integration: Combining functional, structural, and demographic data significantly boosts diagnostic accuracy compared to using any single modality. This highlights the complex, multi-faceted nature of MDD.
  • Local-Global Synergy: A hierarchical approach that analyzes both individual brain regions (local) and population-level patterns (global) is crucial for robust MDD diagnosis. This mirrors how clinicians often integrate detailed patient information with broader medical knowledge.
  • Learnable Graphs are Key: Moving beyond static graph structures, learning the graph adjacency matrix allows the model to adapt and capture more nuanced relationships, leading to improved performance.
  • Cerebellum's Underappreciated Role: The study's interpretation revealed the cerebellum as a significant region involved in MDD, both functionally and structurally. This challenges the traditional focus on prefrontal and limbic regions and opens new avenues for research.
  • AI as a Diagnostic Assistant: The LGMF-GNN not only diagnoses MDD but can also estimate symptom severity and identify similar past cases, acting as a powerful assistant for clinicians.

Key Results

  • State-of-the-Art Performance: The LGMF-GNN achieved a 10-fold cross-validation (CV) AUROC of 80.64% ± 5.74% on the SRPBS dataset and 73.67% ± 2.67% on the REST-meta-MDD dataset.
  • Superiority over SOTA: The model showed improvements of at least 5.46% in accuracy and 7.60% in AUROC compared to state-of-the-art methods like BrainGNN on identical datasets.
  • MDD Subtype Classification: The system accurately differentiated between healthy controls and subtypes like first-episode drug-naive (FEDN) and recurrent (RECU) MDD patients.
  • Biomarker Identification: The interpretable module identified specific functional connectivity patterns (e.g., between left gyrus rectus and right cerebellar lobule VIIB) and structural alterations (e.g., gray and white matter interface changes) associated with MDD.
  • Generalizability: The model demonstrated good generalization performance on external datasets (Anding and OpenNeuro), even with population differences.
  • Robustness to Medication: The system showed only minimal differences in performance between drug-naive and treated MDD patients, indicating robustness.

Limitations and Future Work

  • Site Effects: While mitigation strategies were employed, site effects still impacted performance in multi-site scenarios. Future work will focus on developing more advanced harmonization and domain migration techniques.
  • Computational Demands: The global subject graph construction, while powerful, requires significant computing and storage resources. Future work will explore graph sampling methods (e.g., GraphSAGE, FastGCN) to improve scalability.
  • Structural Modality Interpretability: The interpretability of structural MRI features (T1) was less detailed. Further refinement of feature extraction methods for sMRI is needed.
  • Broader Applications: The LGMF-GNN framework is versatile and holds promise for diagnosing other brain diseases and neurological conditions.
  • Large-Scale Data Integration: The authors are involved in a large-scale mental health big data platform, aiming for more comprehensive validation and long-term progress.

Summary updated

BibTeX
@article{liu2024objective,
  author = {Shuyu Liu and Jingjing Zhou and Xuequan Zhu and Ya Zhang and Xinzhu Zhou and Shaoting Zhang and Zhi Yang and Ziji Wang and Ruoxi Wang and Yizhe Yuan and Xin Fang and Xiongying Chen and Yanfeng Wang and Ling Zhang and Gang Wang and Cheng Jin},
  title = {{An objective quantitative diagnosis of depression using a local-to-global multimodal fusion graph neural network}},
  journal = {Patterns},
  year = {2024},
  doi = {10.1016/j.patter.2024.101081},
}
Figure