📄 medicine

A Dual-Stream Spatio-Temporal Network with Bidirectional Rectification for Differential Diagnosis of Dementia using EEG

This paper proposes a dual-stream spatio-temporal network with a bidirectional rectification fusion module and knowledge distillation to improve the differential diagnosis of Alzheimer's disease and frontotemporal dementia using resting-state EEG, achieving measurable performance gains over baseline methods on the ADFTD-RS dataset.

Original authors: Lingjia Yu, Ming Meng, Yunyuan Gao, Xugang Xi

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lingjia Yu, Ming Meng, Yunyuan Gao, Xugang Xi

Original paper licensed under CC BY 4.0 (https://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

The human brain is a vast, intricate network where billions of cells communicate in a constant, electrical hum. When this network begins to fray, as it does in neurodegenerative diseases, the pattern of that hum changes. Two of the most common and devastating conditions that cause this decline are Alzheimer's disease and frontotemporal dementia. While both lead to memory loss and personality changes, they damage different parts of the brain and require different treatments. Distinguishing between them early is a critical challenge for doctors, yet their symptoms often overlap so closely that even experienced clinicians struggle to tell them apart. To solve this, researchers have turned to the electroencephalogram, or EEG, a tool that records the brain's electrical activity from the scalp. It is a non-invasive method that captures the brain's rhythm with incredible speed, offering a window into the mind's inner workings. However, the signals recorded are notoriously messy, filled with static and noise, and the differences between the two diseases are often so subtle that they are easily lost in the clutter.

A team of researchers at Hangzhou Dianzi University has developed a new approach to cut through this noise, aiming to help computers make a more precise diagnosis. Their work focuses on a specific problem: existing methods usually look at the brain's electrical signals in just one way, either as a timeline of events or as a map of connections between different brain regions. The researchers found that relying on just one of these perspectives is like trying to understand a complex story by reading only the dates or only the character names; you miss the full picture. Furthermore, simply combining these two views often makes the problem worse, as the errors and noise from one view can amplify the errors in the other. To address this, the team built a dual-stream network, a computer system designed to look at the EEG data through two lenses simultaneously. One lens tracks the timing of the electrical waves, while the other maps how different parts of the brain are talking to each other.

The core of their innovation lies in how these two streams of information are brought together. Instead of forcing the data to merge immediately, which would mix the noise along with the signal, the system first acts as a filter. It uses a process called bidirectional rectification, where the timing information helps clean up the connection map, and the connection map helps clean up the timing data. Imagine two people trying to hear a faint conversation in a noisy room; if one person points out where the noise is coming from, the other can ignore it, and vice versa. This mutual correction allows the system to isolate the true patterns of the disease. Once the data is cleaned, the system fuses the two streams using a mechanism that weighs the importance of each piece of information, ensuring that the final decision is based on the strongest, most reliable evidence.

To make this system even more effective, the researchers had to overcome another hurdle: the lack of enough data. In medical studies, it is often difficult to find a large number of patients with frontotemporal dementia compared to those with Alzheimer's or healthy controls. This imbalance can trick a computer into guessing the majority class to get a high score, rather than learning to spot the rare disease. The team solved this by using a technique called knowledge distillation. They first trained a simpler version of their system on the more common conditions to learn the basics of the disease. Then, they used this expert system to guide the training of the more complex system, teaching it how to pay attention to the right features without getting confused by the lack of examples for the rarer condition.

When tested on a dataset of resting-state EEG recordings from patients with Alzheimer's, patients with frontotemporal dementia, and healthy individuals, the new system showed measurable improvements over previous methods. It achieved a subject-level accuracy of 60.00% and an F1-score of 57.44%. While these numbers might seem modest, they represent a significant step forward in a field where distinguishing between these two diseases is notoriously difficult. The system was able to identify specific patterns in the brain's electrical flow that correspond to the known biology of each disease. For instance, it detected that Alzheimer's patients showed changes in the back and center of the brain, consistent with the known degradation of memory networks, while frontotemporal dementia patients showed shifts in the front and sides of the brain, matching the areas responsible for behavior and language.

The researchers acknowledge that this is a starting point rather than a final solution. The system was tested on a single dataset, and its ability to work across different hospitals and patient groups still needs to be proven. Additionally, the computational power required to run such a complex model is currently too high for portable, real-time use in a doctor's office. However, the study demonstrates that by combining different ways of looking at the brain's signals and carefully filtering out the noise, it is possible to build a more reliable tool for diagnosis. This approach offers a promising path toward computer-aided screening that could one day help doctors identify these devastating diseases earlier and with greater confidence, potentially changing how neurodegenerative conditions are managed in the future.

Technical Summary: A Dual-Stream Spatio-Temporal Network with Bidirectional Rectification for Differential Diagnosis of Dementia using EEG

Problem Statement
Accurate differential diagnosis between Alzheimer's disease (AD) and frontotemporal dementia (FTD) is critical for early screening but remains challenging due to highly overlapping symptoms and EEG waveforms. While scalp electroencephalogram (EEG) offers excellent temporal resolution, its application is hindered by inherent low signal-to-noise ratios, non-stationarity, and subtle inter-class differences. Existing single-domain representation methods face specific limitations: pure temporal models (e.g., Transformers) often neglect brain network topology, while pure spatial models (e.g., Graph Neural Networks) struggle with spurious edges generated by estimating effective connectivity from short-term, noisy EEG data. Furthermore, direct fusion of these features often amplifies noise rather than mitigating it. Additionally, clinical datasets frequently suffer from severe sample imbalance, particularly regarding the scarcity of FTD cases, which biases standard data-driven models.

Methodology
The authors propose a Dual-Stream Spatio-Temporal Network designed to extract and integrate complementary features while dynamically suppressing noise. The framework consists of three primary components:

Dual-Stream Feature Extraction:
- Temporal Stream: Utilizes multiscale temporal patching to segment EEG signals into varying window lengths. These patches are projected into high-dimensional vectors, augmented with positional and learnable granularity embeddings, and processed through a multi-layer Transformer encoder to capture long-range transient dependencies and cross-channel relationships.
- Spatial Stream: Initializes node features using the logarithmic power spectral density (log-PSD) of each channel. It constructs an effective connectivity graph using a Multivariate Autoregressive (MVAR) model to calculate Granger causality. A random-walk symmetric normalization is applied to the adjacency matrix, and a Graph Convolutional Network (GCN) aggregates spatial information to capture whole-brain causal abnormalities.
Spatio-Temporal Feature Fusion Based on Bidirectional Rectification (STSBR):
To prevent the cross-contamination and amplification of noise inherent in direct feature concatenation, the authors introduce a bidirectional rectification module prior to fusion.
- Channel-Level Rectification: Global pooling representations from both streams are concatenated and processed by an MLP to generate channel-level suppression weights.
- Spatial-Level Synergistic Rectification: Local features of one modality are concatenated with the global context of the other modality to generate spatial cross-masks via an MLP.
- Cross-Attention Fusion: The rectified features ( $X'_t$ and $X'_G$ ) are fused using a linear cross-attention mechanism. This executes a bidirectional query process where temporal features query graph features and vice versa, achieving deep semantic alignment and efficient fusion.
Mask-Guided Knowledge Distillation:
To address class imbalance (specifically the scarcity of FTD samples), a two-stage training strategy is employed.
- A binary "Teacher" model is pre-trained on AD and Healthy Control (HC) data to learn baseline manifolds.
- A three-class "Student" model is initialized with the Teacher's encoder weights.
- During fine-tuning, a masked attention distillation loss ( $L_{distill}$ ) guides the Student model. A mask ( $V_i$ ) ensures the distillation only applies to AD and HC samples ( $V_i=1$ ), while FTD samples ( $V_i=0$ ) are excluded from the distillation loss, allowing the Student model to learn FTD features autonomously.
- A "warm-up" phase freezes the backbone to stabilize the decision boundary before full end-to-end optimization.

Key Contributions

Dual-Stream Architecture: A parallel extraction framework that overcomes the representation bottlenecks of single-domain models by capturing both multiscale temporal dynamics and effective connectivity graph spatial features.
Bidirectional Rectification Module: A novel fusion mechanism that performs dynamic noise suppression on both temporal and spatial streams before cross-attention integration, effectively preventing noise amplification and feature misalignment.
Mask-Guided Distillation Strategy: A technique utilizing a pre-trained binary teacher model to stabilize decision boundaries and handle three-class classification under severe data imbalance without forcing the model to learn FTD features from the teacher.

Experimental Results
The proposed framework was evaluated on the ADFTD-RS dataset (88 participants: 36 AD, 23 FTD, 29 HC) using a subject-independent split.

Performance: The model achieved a subject-level accuracy of 60.00% and an F1-score of 57.44%.
Comparison: These results outperformed five baseline methods, including handcrafted features, EEGNet, EEGInception, MedGNN, and a pure Transformer. Notably, the proposed method improved the F1-score by 6.87% over the best-performing baseline (Transformer).
Ablation Studies: Removing the bidirectional rectification (w/o STSBR) caused a significant drop in F1-score (to 42.34%), confirming the necessity of noise filtering. Removing the distillation strategy (w/o Distill) reduced the F1-score to 54.64%, highlighting the importance of handling class imbalance.
Interpretability: Topographic visualizations of causal outflow revealed distinct patterns: AD showed increased centroparietal outflow (linked to Default Mode Network impairment), while FTD exhibited anteriorly shifted outflow (frontocentral and left temporal), aligning with known neurophysiological mechanisms.

Significance and Claims
The paper claims to offer a feasible approach for the computer-aided differential diagnosis of dementia based on resting-state EEG. The authors assert that their framework addresses the critical limitations of existing methods by:

Providing a more comprehensive representation of neurodegenerative diseases through coupled spatio-temporal modeling.
Mitigating feature interference and noise amplification through dynamic bidirectional rectification.
Offering a practical regularization approach for imbalanced clinical data via attention-guided distillation.

The authors maintain a modest stance regarding clinical deployment, acknowledging that the current evaluation is limited to a single database of resting-state EEG. They note that further validation on multi-center datasets is required to assess generalization and that computational overhead remains a constraint for real-time execution on edge devices. The work is presented as a step toward non-invasive computational tools for screening neurodegenerative diseases, rather than a fully deployed clinical solution.

More like this