⚡ electrical engineering

BandRouteNet: An Adaptive Band Routing Neural Network for EEG Artifact Removal

BandRouteNet is an adaptive, parameter-efficient neural network for EEG denoising that utilizes a frequency-aware routing mechanism and a full-band conditioner to effectively remove diverse artifacts like EOG and EMG.

Original authors: Phat Lam

Published 2026-04-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Phat Lam

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 you are trying to listen to a beautiful, delicate violin solo (the EEG signal), but you are sitting in a crowded, noisy cafe. Suddenly, someone drops a tray of silverware (a sudden EOG/eye blink artifact), and in the background, a loud espresso machine is constantly hissing (a steady EMG/muscle noise artifact).

If you just turn down the volume on everything, you’ll lose the violin entirely. If you try to use a standard noise-canceling headphone, it might struggle because the "noise" is changing constantly—sometimes it's a low rumble, sometimes it's a high-pitched screech.

This paper introduces BandRouteNet, a smart "digital ear" designed specifically to clean up these brain signals. Here is how it works using three simple ideas:

1. The "Frequency Specialist" (Band-specific Denoiser)

Instead of trying to fix the whole sound at once, BandRouteNet acts like a team of specialized audio engineers.

One engineer only listens to the bass (low frequencies), where eye blinks usually hide.
Another listens only to the treble (high frequencies), where muscle tension noise lives.

By splitting the signal into different "frequency bands," the network doesn't get confused. It knows exactly which "shelf" of the sound to clean without accidentally messing up the parts that are already clear.

2. The "Smart Volume Knob" (The Routing Mechanism)

This is the most clever part. In most systems, the "cleaning" is applied at the same strength all the time. This is like a janitor scrubbing a floor with equal force even when there is no dirt—it ends up wearing down the good parts of the floor.

BandRouteNet uses an Artifact Router. Think of this as a smart sensor that watches the signal in real-time. If it detects a sudden "clatter" (an artifact), it turns the cleaning power up to 100%. But if it hears the beautiful violin playing clearly, it turns the cleaning power down to 0%, letting the pure signal pass through untouched. It only "scrubs" exactly where and when the noise appears.

3. The "Big Picture Observer" (Full-band Conditioner)

While the specialists are busy with their specific frequencies, there is a "Manager" watching the entire scene. This is the Full-band Conditioner.

The Manager doesn't do the heavy scrubbing, but they provide context. They say to the Bass Engineer, "Hey, I see a huge crash coming in the high frequencies; get ready!" or "The overall volume is getting low, adjust your settings." This ensures that even though the engineers are working separately, they are all working in harmony with the overall rhythm of the music.

Why does this matter?

The researchers tested this against many other "digital ears," and BandRouteNet won. It was:

More Accurate: It cleaned the noise better without "smudging" the brain signal.
More Efficient: It is incredibly "lightweight." In computer terms, it’s like having a world-class conductor who fits inside a tiny pocket watch (only 0.2 million parameters).

The Bottom Line: Because it is so small and smart, this technology could eventually be put into wearable devices (like a headband or a tiny medical sensor) to give doctors and scientists a crystal-clear view of brain activity, even while a patient is moving or blinking.

Technical Summary: BandRouteNet: An Adaptive Band Routing Neural Network for EEG Artifact Removal

1. Problem Statement

Electroencephalography (EEG) signals are highly susceptible to contamination from physiological artifacts, primarily Electrooculographic (EOG) (eye movements) and Electromyographic (EMG) (muscle activity). These artifacts present significant challenges because:

Spectral Diversity: EOG artifacts are typically concentrated in low-frequency bands, whereas EMG artifacts are broadband and more prominent in high frequencies.
Temporal Non-stationarity: Artifact intensity varies significantly over time.
Modeling Limitations: Existing deep learning methods often apply uniform processing across the entire signal, leading to either under-correction of severe artifacts or over-smoothing of clean neural components. Furthermore, many models fail to integrate both temporal sequence modeling and frequency-domain characteristics effectively.

2. Methodology

The authors propose BandRouteNet, a dual-path framework designed to exploit both frequency-specific patterns and global temporal context. The architecture consists of two cooperative components:

A. Band-specific Denoiser (The "What" and "Where"):

DFT Decomposition: The input signal is decomposed into $K$ frequency bands (Delta, Theta, Alpha, Beta, Gamma, and a high-frequency $\epsilon$ band) using Discrete Fourier Transform (DFT).
Band Adapter & FiLM: Each band is processed in a shared latent space. To ensure the model knows which band it is processing, a learnable band identity embedding is added. To incorporate global context, Feature-wise Linear Modulation (FiLM) is used to scale and shift band features based on parameters from the full-band pathway.
Artifact Router (The Routing Mechanism): This is a key innovation. Instead of uniform denoising, the router predicts a soft, time-varying mask ( $g_k \in [0, 1]$ ). This mask adaptively interpolates between the original feature and a "denoised proposal," allowing the model to apply stronger correction only to corrupted temporal-spectral regions.
Cross-band Fusion: After individual band processing, a module using Multi-Head Self-Attention (MHSA) models the dependencies between different frequency bands to ensure spectral coherence.

B. Full-band Conditioner (The "Global Context"):

This branch processes the original noisy signal to capture long-range temporal dependencies using a combination of 1D Convolutions and GRUs.
It provides three outputs:
1. A coarse-grained refinement signal to supplement the final reconstruction.
2. A temporal gating mask ( $\lambda$ ) to control the contribution of the coarse signal.
3. Modulation parameters ( $\tau, \psi$ ) to guide the Band-specific Denoiser via FiLM.

C. Final Fusion:
The final output is a weighted combination of the aggregated band-wise denoised signals and the coarse-grained full-band refinement.

3. Key Contributions

Frequency-Aware Architecture: Unlike uniform denoisers, BandRouteNet explicitly models the spectral heterogeneity of EEG artifacts.
Adaptive Artifact Routing: The introduction of a routing mechanism allows for selective, data-dependent denoising strength, preserving clean signal morphology.
Dual-Path Integration: The model successfully bridges the gap between localized frequency-domain processing and global temporal-domain modeling.
Extreme Parameter Efficiency: The model achieves state-of-the-art results with only 0.2M trainable parameters, making it suitable for edge computing and resource-constrained medical devices.

4. Results

The model was evaluated on the EEGDenoiseNet benchmark across three scenarios: EOG, EMG, and Mixed EOG/EMG.

Superior Performance: BandRouteNet outperformed all baseline models (including FCNN, CNN, RNN, EEGDNet, and LRR-UNet) across all metrics: RRMSEt (temporal error), RRMSEs (spectral error), CC (correlation), and SNRimp (signal-to-noise improvement).
Robustness: It demonstrated high stability across various Signal-to-Noise Ratio (SNR) levels, particularly in low-SNR (high noise) environments.
Interpretability: Visualization of the Artifact Router confirmed that the model correctly identifies and targets artifacts in their respective frequency bands (e.g., low-frequency activation for EOG and high-frequency for EMG).
Ablation Success: Ablation studies proved that the Artifact Router and the Full-band Conditioner are the most critical components for achieving high-fidelity reconstruction.

5. Significance

BandRouteNet represents a significant advancement in EEG signal processing. By moving away from "one-size-fits-all" denoising toward an adaptive, frequency-aware approach, it provides a more nuanced way to clean neural data. Its combination of high performance and low computational footprint makes it a highly practical candidate for real-time applications in Brain-Computer Interfaces (BCIs), clinical neurological monitoring, and portable sleep analysis tools.