KnowMDD: Knowledge-guided Cross Contrastive Learning for Major Depressive Disorder Diagnosis

Anchen Lin, Weikun Wang, Haijun Han, Fanwei Zhu, Qi Ma, Zengwei Zheng, Binbin ZhouIJCAI 2025 · DOI

TL;DR

This paper introduces KnowMDD, a novel framework for diagnosing Major Depressive Disorder (MDD) using fMRI data. By combining multi-atlas brain graph construction with knowledge-guided contrastive learning, KnowMDD effectively addresses data sparsity and leverages domain knowledge (specifically the Default Mode Network) to improve diagnostic accuracy and robustness.

What Problem Does This Paper Address?

Major Depressive Disorder (MDD) is a severe mental illness affecting millions worldwide. While Functional Magnetic Resonance Imaging (fMRI) shows promise for objective diagnosis by analyzing brain connectivity (FC), a significant challenge is the scarcity of fMRI data due to high acquisition costs. This small sample size hinders the effectiveness of traditional machine learning models in identifying subtle abnormal FC patterns. Furthermore, existing methods often fail to incorporate valuable domain knowledge about brain networks, such as the Default Mode Network (DMN), which is known to be implicated in MDD. This paper aims to bridge this gap by developing a method that can accurately diagnose MDD despite data limitations and by intelligently integrating domain-specific insights.

Methods

KnowMDD is built upon two core components:

  1. Multi-atlas-based Brain Functional Graph Construction: Instead of relying on a single brain atlas, KnowMDD utilizes multiple atlases (e.g., AAL, Harvard, Craddock) to segment the brain into Regions of Interest (ROIs). This multi-view approach generates richer and more complementary brain graph representations. Node features are constructed by combining Pearson correlation coefficients of FC between ROIs with additional pathological information (like age and gender) to capture individual heterogeneity.
  2. Knowledge-guided Contrastive Learning: This component is the heart of KnowMDD's novelty.
  3. Domain Knowledge Integration: The Default Mode Network (DMN), known to be associated with MDD, is specifically incorporated. The framework generates DMN-specific subgraphs to augment the data, guided by random walks starting from DMN nodes. This ensures that the learning process focuses on relevant brain regions.
  4. Attention Mechanism: An attention mechanism is employed to learn the global semantic relationships between brain regions, allowing the model to weigh the importance of different connections.
  5. Cross-Contrastive Learning: This strategy is used to learn robust representations. It treats brain graph representations from different atlases of the same patient as positive pairs and representations from different patients as negative pairs. This contrastive approach helps to capture consistent discriminative features across atlases, reduce noise, and mitigate the effects of data sparsity. The final loss function combines this contrastive loss with a binary classification loss.

Datasets

The primary dataset used for training and evaluation is the Rest-meta-MDD dataset, specifically Site 20, which contains 282 MDD and 251 Normal Control (NC) samples. The paper also reports results on data from Site 1, 21, and 25 to evaluate robustness and transferability.

Key Insights and Ideas

  • Multi-atlas fusion is crucial: Using multiple brain atlases provides complementary views of brain connectivity, leading to more robust graph representations than single-atlas approaches.
  • DMN knowledge is a powerful biomarker: Explicitly incorporating knowledge about the Default Mode Network significantly improves diagnostic performance, highlighting its critical role in MDD.
  • Contrastive learning combats sparsity: The cross-contrastive learning strategy effectively leverages limited data by learning to distinguish between positive (same patient, different views) and negative (different patients) pairs, enhancing generalization.
  • Attention refines importance: The attention mechanism helps the model focus on the most relevant brain regions and connections for diagnosis, further improving accuracy.

Key Results

KnowMDD demonstrates superior performance across various metrics (Accuracy, Sensitivity, Specificity, F1-score) compared to 9 baseline methods, including both self-supervised and graph learning approaches. For instance, on Site 20, KnowMDD achieved the best performance, with specific results highlighted in Table 1 (e.g., Accuracy of 78.87±7.319 when using AAL&Harvard&Craddock atlases). The ablation study confirms that each component of KnowMDD (domain knowledge, multi-atlas, attention) contributes significantly to its overall effectiveness. The model also shows strong transferability to datasets from different sites, indicating good generalizability.

Limitations and Future Work

  • Sensitivity vs. Specificity: While KnowMDD excels in sensitivity (correctly identifying MDD patients), its improvement in specificity (correctly identifying NC) is more modest. The authors suggest this might be due to the contrastive learning's focus on abnormal connectivity features.
  • Hyperparameter Sensitivity: The performance is sensitive to hyperparameters like the random walk step size (k) and learning rate (lr), requiring careful tuning for optimal results.
  • Future Work: The authors developed a demonstration system to showcase the practical application of KnowMDD, suggesting a path towards real-world deployment. Further research could explore incorporating more diverse domain knowledge or other neuroimaging modalities.

Summary updated

BibTeX
@inproceedings{lin2025knowmdd,
  author = {Anchen Lin and Weikun Wang and Haijun Han and Fanwei Zhu and Qi Ma and Zengwei Zheng and Binbin Zhou},
  title = {{KnowMDD: Knowledge-guided Cross Contrastive Learning for Major Depressive Disorder Diagnosis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/838},
}
Figure