🤖 machine learning

SA-HGNN: Sample-Adaptive Hyperbolic Graph Neural Network for EEG-Based Depression Recognition

This paper proposes SA-HGNN, a novel Sample-Adaptive Hyperbolic Graph Neural Network that leverages personalized graph construction, hyperbolic geometry, and attention pooling to accurately capture the inherent hierarchical structure of brain networks and improve EEG-based depression recognition.

Original authors: Yang Li, Pan Hu, Yan Zhang, Wenfan Yang, Tao Wu, Lianbo Guo

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Li, Pan Hu, Yan Zhang, Wenfan Yang, Tao Wu, Lianbo Guo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a massive, bustling city with thousands of neighborhoods (brain regions) constantly sending messages to one another. In a healthy city, these messages flow smoothly, but in a city affected by depression, the traffic patterns get weird. Some neighborhoods talk too much, others stop talking, and the whole map of connections becomes distorted.

For a long time, doctors and scientists have tried to read these "traffic maps" using EEG (electrodes on the scalp) to diagnose depression. They use computer models called Graph Neural Networks (GNNs) to understand these connections. However, the paper argues that the old way of looking at these maps is like trying to draw a complex, multi-level subway system on a flat piece of paper. It works okay for simple things, but it squishes and distorts the deep, hierarchical layers of the brain's structure.

Here is how the authors' new invention, SA-HGNN, fixes this problem, explained in simple terms:

1. The Problem: Flat Maps vs. 3D Hierarchies

The paper explains that the brain isn't just a flat web; it has a "hierarchical" structure. Think of it like a family tree or a corporate organization: you have a few big "boss" hubs at the top that connect to many smaller "employee" nodes at the bottom.

The Old Way (Euclidean Space): Imagine trying to fit a giant, branching tree onto a flat sheet of paper. As the tree gets bigger, the branches get squished together, and you lose the true shape of the tree. This is what happens when current AI models try to map the brain's complex hierarchy on a flat digital grid.
The New Way (Hyperbolic Space): The authors use a special kind of math called "Hyperbolic Geometry." Think of this as a saddle-shaped surface or a growing coral reef. On this surface, space expands exponentially as you move outward. This allows the AI to fit that giant, branching tree perfectly without squishing it. It captures the true "boss-to-employee" relationships in the brain much better.

2. The Three Magic Tools

The SA-HGNN model uses three specific tools to solve the depression detection puzzle:

A. The "Personalized Map Maker" (Sample-Adaptive Graph Construction)

The Problem: Every person's brain is slightly different. Using a standard, one-size-fits-all map doesn't work well.
The Solution: This module acts like a custom tailor. Instead of using a pre-made map, it looks at the specific EEG signals of one patient and dynamically builds a unique map just for them. It combines what we know about where electrodes are physically placed with what the brain is actually doing in that moment. It creates a "personalized traffic map" for every single patient.

B. The "3D Navigator" (Hyperbolic Graph Convolution)

The Problem: Once the map is made, the computer needs to understand the deep, hidden patterns inside it.
The Solution: This is the 3D Navigator. Instead of walking on a flat floor (Euclidean space), this tool walks on that curved, expanding surface (Hyperbolic space). It allows the computer to see the "rich-club" structure of the brain—where the most important hubs connect to the rest of the network—without getting lost or confused by the distortion. It finds the hidden hierarchy that other models miss.

C. The "Noise Filter" (Attention Pooling)

The Problem: EEG signals are noisy. It's like trying to hear a conversation in a crowded, loud stadium. Some brain signals are just static or irrelevant "noise."
The Solution: This module acts like a smart bouncer or a noise-canceling headphone. It looks at all the brain signals and asks, "Who is actually important?" It filters out the noisy, unimportant channels and keeps only the "informative" ones. This ensures the final diagnosis is based on the real brain patterns, not the static.

3. The Results: A Better Diagnosis

The authors tested this new system on a public dataset of EEG recordings from people with depression and healthy people. They tested it in two scenarios:

Resting State: When the person is just sitting quietly.
Task State: When the person is doing a specific mental task.

The Outcome:
The SA-HGNN model outperformed all the previous "flat map" models.

It achieved an accuracy of 95.24% for resting-state data and 94.26% for task-based data.
Crucially, it had a very high "Recall" (96.19% and 98.70%), meaning it was excellent at catching the people who actually had depression and rarely missed them.

4. What the Model "Saw"

When the researchers looked at the maps the AI created, they found patterns that matched real medical knowledge:

In depressed patients, the "middle" of the brain was talking to itself too much (hyperconnectivity), which is a known sign of depression.
During tasks, the "front-left" part of the brain showed strange changes, which links to the emotional and thinking difficulties seen in depression.

Summary

In short, the paper proposes a new AI tool that stops trying to flatten the brain's complex 3D structure onto a 2D sheet. Instead, it builds a custom, 3D map for each patient, navigates it using curved geometry to find hidden patterns, and filters out the noise to give a highly accurate diagnosis of depression. It's like upgrading from a flat paper map to a high-tech, 3D holographic navigation system for the human brain.

Technical Summary: SA-HGNN for EEG-Based Depression Recognition

Problem Statement

Major Depressive Disorder (MDD) is a leading global health crisis, necessitating objective, automated diagnostic methods. Electroencephalography (EEG) is widely used for this purpose due to its non-invasive nature and high temporal resolution. While Graph Neural Networks (GNNs) have emerged as a mainstream approach to model brain functional connectivity, existing methods face two critical limitations:

Inadequate Structural Modeling: Human brain functional connectivity possesses an inherent hierarchical structure (e.g., "rich-club" topology). Standard GNNs operate in Euclidean space, which struggles to represent exponentially growing hierarchical tree-like structures without distortion, leading to inaccurate modeling of the latent hierarchy in depression-affected networks.
Rigid Topology and Noise: Existing approaches often rely on static graph structures or fail to effectively filter redundant noise channels, which can obscure the authentic topological patterns of the brain network.

Methodology: Sample-Adaptive Hyperbolic Graph Neural Network (SA-HGNN)

To address these challenges, the authors propose SA-HGNN, a framework designed to extract the authentic hierarchical structure of depression-affected brain networks. The architecture consists of three core modules:

1. Sample-Adaptive Graph Construction (SAGC)

This module dynamically constructs personalized brain network topologies for each sample, moving beyond static graph definitions. It fuses two branches:

Physical Prior Branch: Constructs a baseline spatial topology ( $A_{prior}$ ) based on standard 10–20 electrode coordinates and Euclidean distances.
Sample-Specific Correlation Branch: Derives a data-driven similarity matrix ( $A_{sim}$ ) from extracted temporal features via L2 normalization and linear projection.
Fusion: The final adjacency matrix is formed by fusing the sample-specific correlations (filtered by a learnable mask) with the physical prior using a learnable scalar ( $\lambda$ ). This provides structural priors while adapting to individual variations.

2. Hyperbolic Graph Convolution (HGC)

To overcome the representation bottlenecks of Euclidean space, the model employs hyperbolic geometry (specifically the Poincaré ball model) to capture latent hierarchical relationships.

Mapping: Features are mapped between Euclidean and hyperbolic spaces using exponential ( $\text{exp}_0^c$ ) and logarithmic ( $\text{log}_0^c$ ) maps.
Convolution: The process involves projecting Euclidean features to the tangent space, transforming them into hyperbolic space, applying linear transformations, and aggregating neighboring structural information within the tangent space before mapping back to Euclidean space. This allows the network to accommodate the exponential growth of hierarchical structures inherent in brain networks.

3. Attention Pooling (AP)

This module adaptively filters out highly redundant noise channels to preserve the authentic hierarchical topology.

Scoring: A learnable global node interacts with all EEG channel nodes via multi-head self-attention to compute importance scores for each node.
Coarsening: The top $K$ nodes are selected to generate a pooled feature set and a coarsened adjacency matrix for subsequent layers.
Loss Constraint: An AP Uniformity Loss ( $L_{uni}$ ) is introduced to prevent over-smoothing and ensure representational diversity by constraining the features of discarded nodes.

Loss Function

The total loss function combines three terms:

Classification Loss ( $L_{cls}$ ): Standard cross-entropy loss for depression recognition.
Sparse Loss ( $L_{sparse}$ ): Encourages simpler graph topologies in the SAGC module.
Uniformity Loss ( $L_{uni}$ ): Ensures diversity in the attention pooling mechanism.

Key Contributions

The paper claims the following primary contributions:

Novel Architecture: It is the first work to adopt Hyperbolic Graph Neural Networks to model the hierarchical functional connectivity of MDD brain networks.
Sample-Adaptive Graph Construction: The introduction of the SAGC module to dynamically construct individualized brain network topologies, overcoming the rigidity of static graphs.
Integrated Framework: The development of an architecture combining Hyperbolic Graph Convolution and Attention Pooling to capture hierarchical structures in hyperbolic space while filtering noisy channels.
Empirical Validation: Demonstration of superior performance across both resting-state and task-related EEG paradigms on public datasets.

Experimental Results

The model was evaluated on the public HUSM dataset, containing EEG recordings from 34 MDD patients and 30 healthy controls (HCs) under two paradigms: Resting-state (HUSM-Rest) and Task-related (HUSM-Task).

Performance: SA-HGNN achieved state-of-the-art results, outperforming seven baseline graph-based models (including DCGNN, GraphSleepNet, and SDGCN).
- HUSM-Rest: Achieved 95.24% Accuracy and 95.77% F1 Score, surpassing the second-best method (GraphSleepNet) by 1.82% and 2.29%, respectively.
- HUSM-Task: Achieved 94.26% Accuracy and 94.69% F1 Score, outperforming the second-best method (GCBNet) by 0.39% and 0.42%.
Ablation Study: Experiments confirmed that the full model (combining SAGC, HGC, and AP) significantly outperforms variants using only single or paired modules, validating the synergistic effect of the proposed components.
Robustness: Sensitivity analysis showed the model maintains stable performance across a range of hyperparameter values for sparsification and uniform loss weights.

Significance and Claims

The authors claim that SA-HGNN effectively addresses the structural distortion caused by modeling hierarchical brain networks in Euclidean space. By leveraging hyperbolic geometry, the model more accurately captures the "rich-club" topology and latent hierarchical patterns disrupted in MDD.

Furthermore, the visualization of learned adjacency matrices revealed neurobiological biomarkers consistent with clinical findings:

Resting State: Hyperconnectivity in midline central regions (associated with default mode network overactivation).
Task State: Notable alterations in left frontal–parietal regions (linked to emotional–cognitive deficits).

The paper concludes that SA-HGNN provides a robust, noise-mitigated framework for capturing abnormal functional connectivity patterns, offering a significant improvement over traditional Euclidean-based GNNs for automated MDD diagnosis.