I^2RiMA: Spectral Riemannian Representation with Temporal Attention for Mental Stress Detection based on EEG Signals
The paper proposes I^2RiMA, an Intra-Inter Riemannian Manifold Attention Network designed to address cross-subject mental stress detection challenges in EEG signals. It overcomes limitations of conventional Riemannian methods by constructing spatial covariance matrices at each frequency point and mapping them to the SPD tangent space to preserve frequency-specific discriminative cues. The model utilizes frequency cluster aggregation to select informative spectral components and reduce redundancy
Analysis
TL;DR
- The paper proposes I^2RiMA, an Intra-Inter Riemannian Manifold Attention Network designed to address cross-subject mental stress detection challenges in EEG signals.
- It overcomes limitations of conventional Riemannian methods by constructing spatial covariance matrices at each frequency point and mapping them to the SPD tangent space to preserve frequency-specific discriminative cues.
- The model utilizes frequency cluster aggregation to select informative spectral components and reduce redundancy, aligning clusters with specific EEG rhythms.
- An intra-inter slice attention module is introduced to adaptively integrate local slice-level spectral dynamics with global temporal context across EEG sequences.
- Experimental results on three datasets demonstrate that I^2RiMA outperforms five state-of-the-art baselines, achieving up to 82.78% balanced accuracy with high efficiency (1.60M parameters, 31.95M FLOPs).
Why It Matters
This research addresses a critical bottleneck in brain-computer interfaces and digital health: the variability of EEG signals across different subjects. By effectively handling subject-dependent and frequency-specific patterns, it offers a robust solution for personalized stress monitoring systems. For practitioners, this represents a significant step toward deploying efficient, accurate neuro-decoding models in real-world clinical or consumer applications without requiring excessive computational resources.
Technical Details
- Architecture: The core innovation is the I^2RiMA network, which combines Riemannian geometry with attention mechanisms specifically tailored for EEG data processing.
- Frequency-Specific Covariance: Unlike traditional methods that model spatial covariance primarily in the time domain, this approach constructs covariance matrices independently at each frequency point. These are mapped to the Symmetric Positive Definite (SPD) tangent space to maintain channel-wise geometry and capture neural oscillations crucial for cognitive state decoding.
- Frequency Cluster Aggregation: A mechanism is employed to form compact, data-driven frequency clusters aligned with known EEG rhythms. This selects informative spectral components while reducing redundancy, enhancing the signal-to-noise ratio for stress-related features.
- Intra-Inter Slice Attention: The model features a dual-level attention module. It captures local spectral dynamics within slices (intra-slice) and integrates global temporal context across the entire sequence (inter-slice), addressing the fragmentation issues common in standard temporal tokenization.
- Performance Metrics: Evaluated on three distinct datasets, the method achieved a maximum balanced accuracy of 82.78%, surpassing existing baselines while maintaining low computational complexity suitable for edge deployment.
Industry Insight
- Efficiency in Edge AI: The low parameter count (1.60M) and FLOPs (31.95M) suggest that such models can be deployed on wearable devices or mobile health platforms, enabling real-time stress monitoring without heavy cloud dependency.
- Focus on Cross-Subject Generalization: As EEG-based diagnostics move from lab settings to broader populations, methods that inherently handle subject variability through geometric representations and attention mechanisms will become standard. Researchers should prioritize architectures that decouple subject-specific noise from universal physiological markers.
- Integration of Domain Knowledge: The explicit alignment of frequency clusters with EEG rhythms highlights the importance of incorporating neurophysiological priors into deep learning models. Future developments in biomedical AI will likely see increased hybridization of signal processing techniques with neural networks to improve interpretability and performance.
Disclaimer: The above content is generated by AI and is for reference only.