CoPrimeEEG: CRT-Guided Dual-Branch Reconstruction from Co-Prime Sub-Nyquist EEG

CoPrimeEEG is a neural reconstruction framework that utilizes co-prime sub-Nyquist sampling and a CRT-guided dual-branch architecture to efficiently reconstruct high-fidelity EEG signals while simultaneously predicting temporal masks and bandpower features.

The Gist

Imagine you are trying to record a beautiful, complex symphony, but your recording device is incredibly cheap and can only capture a tiny fraction of the sound. Usually, this would result in a garbled, noisy mess.

CoPrimeEEG is like a "super-intelligent musical restorer" that can take those tiny, broken fragments of sound and reconstruct the entire, crystal-clear symphony perfectly.

Here is how it works, broken down into simple ideas:

1. The "Co-Prime" Trick: Two Different Perspectives

Imagine you are trying to watch a fast-moving race through two different strobe lights.

• Strobe A flashes every 3 seconds.
• Strobe B flashes every 5 seconds.

Because 3 and 5 are "co-prime" (they don't share any common factors), the flashes don't sync up in a predictable pattern. This "mismatch" is actually a superpower. By looking at the gaps between the flashes of both lights, you can mathematically figure out exactly what happened in the moments the lights were off.

In this paper, instead of recording brainwaves (EEG) at a high, power-hungry speed, they use two different "low-speed" recording rhythms. Because these rhythms are mathematically offset, they capture a much richer "puzzle" of the brain's activity than a single slow recording ever could.

2. The Dual-Branch Brain: The Expert Reconstructors

The researchers built a digital brain (a neural network) with two "arms" (branches).

• One arm listens to the first rhythm.
• The other arm listens to the second rhythm.

These two arms then meet in the middle, combine their notes, and use a process called CRT (Chinese Remainder Theorem)—which is essentially a mathematical way of "filling in the blanks"—to paint a high-definition picture of the original brainwaves.

3. The Four-Part Quality Check (The "Critic")

To make sure the reconstruction isn't just a "guess," the system uses a strict four-part grading system to train itself:

1. The Mirror Test (Waveform Fidelity): Does the new wave look exactly like the original?
2. The Spotlight Test (Masking): It learns to identify which parts of the signal are actually important and which are just background noise.
3. The Rhythm Test (Bandpower): It ensures the "tempo" and "energy" of the brainwaves (like Alpha or Beta waves) are medically accurate.
4. The Reality Check (CRT-Consistency): It double-checks its work by downsampling its own reconstruction to see if it matches the original "broken" fragments it started with.

Why does this matter?

Currently, high-quality EEG machines require a lot of power and bulky equipment because they have to "listen" to the brain constantly at very high speeds. This is a problem for wearable devices like smart headbands or long-term medical monitors.

CoPrimeEEG offers a way to:

• Save Battery: You can record at a much slower, "low-power" rate.
• Keep Quality High: You still get the high-definition data needed for doctors to make diagnoses.
• Stay Small: It uses fewer computer "brain cells" (parameters) to do the job, making it perfect for small, wearable gadgets.

In short: It’s a way to get "High-Definition" brain data while using "Low-Definition" energy.

Technical Summary

Technical Summary: CoPrimeEEG

Title: CoPrimeEEG: CRT-Guided Dual-Branch Reconstruction from Co-Prime Sub-Nyquist EEG

1. Problem Statement

The primary challenge addressed by this paper is the high power consumption and data bandwidth requirements associated with high-rate Electroencephalogram (EEG) acquisition. Traditional Nyquist-rate sampling requires high sampling frequencies, which drains battery life in wearable devices and creates bottlenecks in wireless transmission. While sub-Nyquist sampling (sampling below the Nyquist rate) can reduce data volume, standard compressive sensing or uniform undersampling often leads to significant information loss and reconstruction artifacts, making it difficult to perform reliable downstream clinical or neurological analysis.

2. Methodology

The authors propose CoPrimeEEG, a deep learning framework that leverages Co-prime Sampling Theory and the Chinese Remainder Theorem (CRT) to reconstruct high-fidelity EEG signals from two low-rate, non-uniformly sampled streams.

• Sampling Strategy: Instead of a single low-rate stream, the system uses two streams obtained via co-prime decimations (sampling rates that are relatively prime to each other). This mathematical property ensures that the combined information from both streams covers the signal spectrum more effectively than a single stream.
• Architecture (Dual-Branch Encoder):
  • Dual-Branch Convolutional Encoder: The framework processes the two co-prime streams through separate convolutional branches to extract unique temporal features from each sampling pattern.
  • Fusion & Upsampling: The features from both branches are fused into a joint representation, which is then passed through an upsampling module to reconstruct the high-rate waveform.
  • Multi-Task Learning: Beyond simple waveform reconstruction, the model simultaneously predicts:
    1. A Temporal Usefulness Mask: Identifying which parts of the signal are most informative.
    2. Canonical Bandpower Features: Predicting frequency-domain characteristics (e.g., Alpha, Beta, Theta waves) to ensure physiological relevance.
• Principled Loss Function: The model is trained using a composite loss function consisting of four specialized terms:
  1. Waveform Fidelity: Minimizes the error between the reconstructed and original time-domain signals.
  2. Mask Sparsity and Smoothness: Regularizes the temporal mask to ensure it is both sparse (focusing on important segments) and temporally consistent.
  3. Log-Domain Bandpower Supervision: Ensures the reconstructed signal maintains accurate spectral power distributions, which is critical for EEG analysis.
  4. CRT-Consistency Term: A novel regularization term that enforces mathematical agreement between the reconstructed high-rate signal and its original co-prime downsampled versions, ensuring the reconstruction is physically consistent with the input.

3. Key Contributions

• Mathematical-Neural Integration: Unifies classical signal processing theory (Co-prime sampling and CRT) with modern deep learning architectures.
• Multi-Objective Reconstruction: Moves beyond simple signal recovery by incorporating physiological feature supervision (bandpower) and temporal importance (masking).
• CRT-Guided Learning: Introduces a consistency constraint that bridges the gap between the continuous reconstruction and the discrete, non-uniform input samples.
• Efficiency: Achieves high-fidelity results with a parameter-efficient architecture compared to existing methods.

4. Results

The framework was evaluated on real-world EEG datasets. The results demonstrate that CoPrimeEEG achieves state-of-the-art (SOTA) performance across multiple quantitative metrics:

• Error Metrics: Superior performance in Mean Squared Error (MSE) and Mean Absolute Error (MAE).
• Signal Quality Metrics: Higher Signal-to-Noise Ratio (SNR), Peak Signal-to-Noise Ratio (PSNR), and Pearson Correlation coefficients.
• Efficiency: The model achieves these superior results while utilizing fewer parameters than competing reconstruction models.

5. Significance

CoPrimeEEG represents a significant step toward ultra-low-power wearable EEG technology. By allowing hardware to sample at much lower rates without sacrificing the ability to perform high-fidelity downstream analysis (like frequency band estimation), this method enables:

• Longer battery life for continuous, long-term neurological monitoring.
• Reduced wireless transmission bandwidth, making it ideal for IoT-based healthcare.
• A robust mathematical foundation that ensures reconstructed signals are not just "visually" similar, but physiologically accurate for clinical use.