EEG-SSFormer: Towards a Robust Mamba-Based Architecture for Dementia Detection from Resting State Electroencephalography

This paper introduces EEG-SSFormer, a novel Mamba-based architecture that decouples channel-wise representation learning from cross-channel interactions to effectively model long-range dependencies in resting-state EEG, achieving a 57.65% accuracy in distinguishing between normal controls, mild cognitive impairment, and dementia on the large-scale CAUEEG dataset while offering interpretable clinical insights.

The Gist

Imagine your brain is a bustling city with thousands of traffic lights (electrodes) constantly flashing in different patterns. When the city is healthy, the traffic flows smoothly. But when the city starts to suffer from "dementia" (a foggy, confused state), the traffic lights start blinking in strange, chaotic rhythms.

For a long time, doctors have tried to diagnose this by taking expensive, heavy X-rays of the brain (like MRI scans) or using radioactive tracers (PET scans). These are like sending a helicopter to inspect every single street corner—it's accurate, but it's expensive, invasive, and hard to do for everyone, especially in remote areas.

A cheaper, easier alternative is EEG, which is like putting a helmet with sensors on your head to listen to the city's traffic lights from the outside. It's portable, cheap, and safe. However, the data it produces is a massive, chaotic river of noise. For years, computers have struggled to make sense of this river and find the specific patterns that signal dementia.

This paper introduces a new, super-smart computer brain called EEG-SSFormer that solves this problem. Here is how it works, broken down into simple concepts:

1. The Problem: The "Traffic Jam" of Data

Previous computer models tried to read the EEG data like a human reads a book, word by word.

• The Old Way (CNNs/Transformers): Imagine trying to understand a 10-minute speech by reading every single letter at once. It's overwhelming. The computer gets confused, forgets the beginning by the time it reaches the end, or gets stuck trying to remember too much at once.
• The Specific Challenge: In dementia, the "noise" isn't just random; it's a subtle change in how the traffic lights talk to each other over a long period. Old models were bad at listening to these long conversations.

2. The Solution: The "Mamba" Detective

The authors built a new model based on something called Mamba. Think of Mamba as a highly efficient detective who doesn't just read the whole book at once.

• Selective Attention: Instead of trying to remember every single letter, Mamba knows how to skim the text, remembering the important plot points and forgetting the boring filler. It can process a very long story (a long EEG recording) without getting tired or confused.
• The "Channel-Independent" Strategy: Imagine the EEG helmet has 19 different microphones (channels).
  • Old Approach: The computer mixed all 19 microphones together into one giant smoothie, hoping to taste the flavor of dementia.
  • EEG-SSFormer Approach: This model listens to each microphone separately first, like a detective interviewing 19 different witnesses one by one to get their unique story. Only after understanding each witness individually does it ask them to talk to each other to see if their stories match up. This prevents the "noise" from one microphone from drowning out the important clues from another.

3. The Training Ground: A Massive Library

To teach this detective, the researchers used the CAUEEG dataset, which is the largest library of brain waves from dementia patients in the world (over 1,100 people).

• They didn't just look at the data; they tested it rigorously. They made sure the computer didn't just "cheat" by memorizing specific people's voices. They tested it on completely new people it had never seen before (a "subject-wise split"), ensuring the model actually learned the disease, not just the person.

4. The Results: Smarter and Lighter

The results were impressive:

• Better Accuracy: The new model correctly identified dementia, mild cognitive impairment, and healthy brains better than the previous "heavyweight" champions (like ResNet and VGG).
• Efficiency: The old champions were like massive, fuel-guzzling trucks (millions of parameters). The new Mamba model is a sleek, electric sports car. It achieved better results with four times fewer parameters. It's faster, cheaper to run, and easier to deploy on portable devices.

5. The "Why": What Did the Model Learn?

The best part is that the researchers didn't just get a "black box" answer. They asked the model, "Why did you think this person has dementia?"

• The Map: The model pointed to specific areas of the head (like the back of the head and the temples) where the brain waves were most different. This matches what human doctors have seen in decades of research.
• The Frequency: The model realized that the "Theta" waves (a specific slow rhythm) are the biggest red flag for dementia, while "Delta" waves are tricky and can confuse the diagnosis if not handled carefully.

The Big Picture

This paper is a major step forward because it proves that we can use cheap, portable headsets combined with smart, efficient AI to detect dementia early.

Instead of waiting for a patient to get a massive, expensive MRI scan in a big city hospital, we might soon be able to use a simple headset in a rural clinic or a doctor's office. The new "Mamba" AI acts like a tireless, super-smart assistant that listens to the brain's traffic lights, spots the chaos of dementia early, and helps doctors intervene before the condition gets too severe. It's a lighter, faster, and more accessible way to protect our minds.

Technical Summary

1. Problem Statement

The paper addresses the critical challenge of early and accessible diagnosis of dementia (specifically Mild Cognitive Impairment [MCI] and Dementia) using Resting-State Electroencephalography (rs-EEG).

• Limitations of Current Methods: Traditional neuroimaging (MRI, PET) is expensive, invasive, and lacks portability. While rs-EEG is cost-effective and portable, analyzing it is difficult due to its long, multichannel nature.
• Deep Learning Challenges: Existing deep learning approaches (CNNs, RNNs, Transformers) face specific hurdles with rs-EEG:
  • RNNs/LSTMs: Struggle with long sequences due to memory constraints and non-parallelizable training.
  • Transformers: Suffer from quadratic computational complexity relative to sequence length and often require massive datasets to model complex features effectively.
  • Channel Mixing: Traditional models often mix all input channels simultaneously, assuming they share the same underlying process. However, EEG channels capture distinct neural processes; mixing them too early can obscure channel-specific patterns and introduce noise.
• Data Scarcity & Validation: Many existing studies use "trial-wise" validation (shuffling segments across subjects), leading to data leakage and overestimated performance. There is a lack of large-scale, subject-wise validated benchmarks for dementia detection using raw rs-EEG.

2. Methodology: EEG-SSFormer

The authors propose EEG-SSFormer, a novel architecture based on State Space Models (SSM), specifically utilizing the Mamba layer, designed to handle long sequences and decouple channel learning.

Core Architectural Components:

1. Channel-Independent Feature Learning:

   • Instead of mixing channels immediately, the model treats each electrode channel as a separate univariate time series.
   • Patching: Input signals are parsed into discrete segments (patches) of length $P$ using a 1D convolution.
   • Inverted LayerNorm: Unlike standard LayerNorm (which normalizes across features), the model uses Inverted LayerNorm to normalize across time steps for each feature. This preserves channel-specific variations and reduces distribution shift sensitivity.
   • Mamba SSM: Each channel is processed independently through Mamba layers. Mamba uses a selective scanning algorithm to filter irrelevant information and highlight important temporal dependencies, offering linear computational complexity ($O(L)$) compared to Transformers' quadratic complexity.

2. Decoupled Channel-and-Feature Mixing:

   • Temporal Mixing: Handled by the Mamba layers within each channel.
   • Channel Mixing: After temporal processing, a separate Channel Mixer captures interactions between the $C$ channels.
   • Spatial vs. Spectral: The authors implemented and compared two mixing strategies:
     • Spatial Domain: Uses point-wise convolutions (inverted bottleneck structure) to mix channels.
     • Spectral Domain: Uses a frequency-domain mixer (EinFFT) based on Fourier transforms.
   • Result: The Spatial Domain mixer was found to be superior for this specific application, likely because it preserves temporal localization better than the global frequency mixing of the spectral approach.

3. Input Integration:

   • The model takes raw rs-EEG signals (10-second windows, 19 channels, 200Hz).
   • Age Integration: The subject's age is normalized and concatenated with the average-pooled output of the network before the final classification layer, as age is a critical risk factor.

4. Interpretability:

   • Channel Occlusion: Systematically zeroing out specific electrodes to measure the drop in classification probability, generating topographic maps of importance.
   • Frequency Band Occlusion: Iteratively band-stop filtering canonical bands (Delta, Theta, Alpha, Beta, Gamma) to determine which frequencies drive the classification.

3. Key Contributions

• Novel Architecture: First application of a Mamba-based, channel-independent architecture specifically for the differential diagnosis of dementia (HC vs. MCI vs. Dementia) using raw rs-EEG.
• Robust Benchmarking: The study utilizes the Chung-Ang University EEG (CAUEEG) dataset, the largest public rs-EEG dataset for dementia (1,155 subjects). Crucially, it employs a strict subject-wise split for training, validation, and testing to prevent data leakage, a common flaw in prior literature.
• Efficiency: The proposed model achieves state-of-the-art performance with significantly fewer parameters (3.8M) compared to strong CNN baselines like 1D-VGG-19 (20.2M) and 1D-ResNet-18 (~11.3M).
• Physiological Insights: The model provides explainable results that align with clinical literature regarding brain regions and frequency bands associated with cognitive decline.

4. Results

The model was evaluated on the CAUEEG dataset with a 3-class classification task (Healthy Control, MCI, Dementia).

• Performance:
  • EEG-SSFormer-PW + Age achieved the highest accuracy on the test set: 58.37%.
  • EEG-SSFormer-PW (without age) achieved 57.65%.
  • Baselines: Outperformed 1D-ResNet-18 (51.88%) and 1D-VGG-19 (54.01%).
  • Ablation: The spatial channel mixer outperformed the spectral mixer by ~5.3% on the validation set. The use of Inverted LayerNorm improved accuracy by ~2.5% over standard LayerNorm.
• Class-Specific Performance:
  • The model performed best on Healthy Controls (HC).
  • MCI classification remained the most challenging (lowest F1 scores), consistent with the difficulty of distinguishing early-stage cognitive decline.
• Interpretability Findings:
  • Spatial Importance:
    • HC: Frontal, left-temporal, and central-parietal regions.
    • MCI: Central-parietal, right-temporal, and occipital regions.
    • Dementia: Occipital and left-temporal regions.
  • Frequency Importance:
    • Theta band (4-8 Hz): Critical for Dementia detection; removing it caused a 43% drop in accuracy for the Dementia class.
    • Delta band (0.5-4 Hz): Removing it significantly improved Dementia classification (suggesting delta noise might confuse the model for dementia vs. HC/MCI) but hurt HC accuracy.
    • Alpha band: Removing it hurt MCI classification significantly.

5. Significance

• Clinical Relevance: The study demonstrates that Mamba-based models are highly effective for long-sequence EEG data, offering a scalable, low-cost, and portable alternative to MRI/PET for dementia screening.
• Methodological Advancement: By decoupling channel learning from temporal modeling, the paper validates that treating EEG channels as independent univariate series before mixing yields better results than traditional joint embedding.
• Future Directions: The work highlights the need for larger, more homogeneous datasets and suggests that future models could benefit from incorporating specific disease subtypes (e.g., Vascular vs. Alzheimer's) to further refine diagnostic precision. The code is publicly available to facilitate further research.