Impact of Kernel Dimensionality on the Generalizability and Efficiency of Convolutional Neural Networks to Decode Neural Drive from High-density Electromyography Signal

This study demonstrates that while 1D, 2D, and 3D convolutional neural networks all achieve high generalizability in decoding neural drive from high-density electromyography signals, increasing kernel dimensionality does not significantly improve performance but does reduce computational efficiency, suggesting that simpler architectures are preferable for balancing decoding accuracy and speed.

The Gist

Imagine you are trying to teach a computer to understand your thoughts by listening to the electrical "whispers" of your muscles. This is the goal of a Neural-Machine Interface (NMI). When you decide to move your hand, your brain sends a signal down your spine to your muscles. These signals are like a choir of tiny singers (motor units) all singing at once.

To hear this choir clearly, scientists use a high-tech microphone grid called HD-sEMG (High-Density Surface Electromyography) that sits on your skin. The problem is, the signal is messy, like a crowded room where everyone is talking at once.

For years, scientists used complex math tricks (called BSS algorithms) to separate the voices. But these tricks are slow and need constant recalibration, like tuning a radio every time you walk into a new room.

Recently, scientists started using Artificial Intelligence (AI), specifically a type called Convolutional Neural Networks (CNNs), to do the listening. Think of a CNN as a super-smart detective that learns to recognize patterns in the noise.

The Big Question: How "Big" Should the Detective's Eyes Be?

The authors of this paper asked a crucial question: Does the detective need a giant, 3D pair of eyes to see everything, or will a simple 1D pair of eyes work just fine?

In AI terms, this is about the dimensionality of the "kernel" (the filter the AI uses to scan the data):

• 1D CNN: Looks at the signal like a timeline. It only cares about when things happen (temporal). Imagine reading a book one word at a time, left to right.
• 2D CNN: Looks at the signal like a photo. It cares about where things happen on the skin (spatial). Imagine looking at a map to see where the noise is coming from.
• 3D CNN: Looks at the signal like a movie. It cares about both when and where (spatiotemporal). Imagine watching a video of the map over time.

The common assumption was: "The more complex the detective (3D), the better they will be at solving the case."

The Experiment: A Race Between Detectives

The researchers built three detectives (1D, 2D, and 3D) with identical brains, except for the "lens" they used to look at the muscle signals. They trained them on data from people doing knee extensions and ankle movements at different strengths (10%, 30%, 50% of their max effort).

They then tested these detectives on:

1. New Strengths: Did they work if the person pushed harder or softer than they did in training?
2. New Muscles: Did they work if the person used a different muscle?
3. Speed: How fast could they solve the case?

The Surprising Results

Here is what they found, translated into everyday terms:

1. The "Bigger is Better" Myth is Mostly False
The complex 3D detective (the movie watcher) did not consistently win.

• At low effort (whispering): The 3D detective was actually quite good at hearing the rhythm of the signal, but it got the volume wrong. It was like hearing a song perfectly but thinking it was being played twice as loud.
• At high effort (shouting): The simple 1D detective (the timeline reader) often performed just as well, or even better, than the complex ones.
• The Lesson: You don't need a 3D movie camera to understand a muscle's intent. A simple timeline or a 2D map is often enough.

2. The Cost of Complexity (The "Heavy Backpack" Problem)
While the 3D detective was slightly smarter in some specific situations, it was exhausting.

• On a standard computer (CPU): The 3D detective was 8 times slower than the 1D detective. It was like trying to run a marathon while carrying a heavy backpack full of bricks.
• On a super-computer (GPU): The speed gap closed, but the 3D detective was still the slowest.
• The Lesson: If you want to put this technology into a real prosthetic arm or a robot that runs on a small battery, the heavy 3D model is too slow and power-hungry. The lightweight 1D or 2D models are much more practical.

3. The "False Alarm" Issue
The complex 3D detective had a weird quirk: when the person was resting (sitting still), the 3D model sometimes thought they were moving! It saw "ghost signals." The simple 1D model was much better at knowing when to stay quiet.

The Final Verdict

The paper concludes that simpler is often better.

• Don't over-engineer: You don't need the most complex AI to decode muscle signals.
• Efficiency wins: A simple 1D or 2D model can decode your muscle intentions just as accurately as a complex 3D model, but it does it much faster and with less computing power.
• Real-world impact: This means we can build better, faster, and cheaper brain-controlled robots and prosthetics because we don't need massive supercomputers to make them work. We can run them on small, portable devices.

In short: The paper tells us that in the world of decoding muscle signals, a simple, focused flashlight (1D/2D) is often more useful than a giant, heavy, 3D floodlight (3D), especially when you need to move fast and save battery.

Technical Summary

1. Problem Statement

High-density surface electromyography (HD-sEMG) signals contain rich spatial and temporal information regarding motor unit action potentials (MUAPs). While Blind Source Separation (BSS) algorithms (e.g., Convolution Kernel Compensation) are the "gold standard" for decoding neural drive (motor intent), they suffer from poor generalizability across different contraction intensities, muscles, and sessions, often requiring frequent recalibration.

Deep Convolutional Neural Networks (CNNs) have emerged as a promising alternative. However, it remains unclear whether increasing architectural complexity by using higher-dimensional convolutional kernels (2D for spatial, 3D for spatiotemporal) significantly improves decoding generalizability compared to simpler 1D (temporal) kernels, or if this complexity merely incurs unnecessary computational costs. This study aims to systematically evaluate the trade-off between kernel dimensionality, generalizability (across intensities and muscles), and computational efficiency.

2. Methodology

Data Collection & Preprocessing

• Datasets: The study utilized three datasets:
  • Training Dataset: 5 healthy participants performing isometric knee extensions (Vastus Lateralis and Medialis) at 25% MVC across two sessions.
  • Evaluation Dataset A: 7 participants performing ankle plantarflexion (Gastrocnemius Medialis) at 10%, 30%, and 50% MVC to test cross-intensity generalizability.
  • Evaluation Dataset B: 6 participants performing ankle plantarflexion (Gastrocnemius Medialis, Lateralis, and Soleus) at 30% MVC to test cross-muscle generalizability.
• Signal Processing: HD-sEMG signals (64 channels) were decomposed into Motor Unit Spike Trains (MUSTs) using BSS algorithms (CKC/DEMUSE). The "Neural Drive" was quantified as Cumulative Spike Trains (CSTs).
• Input Formatting:
  • 1D CNN: Input shape $(64, 40)$ representing 64 channels $\times$ 40 time samples.
  • 2D CNN: Input reshaped to $(13, 5, 40)$ to preserve spatial electrode grid topology.
  • 3D CNN: Input reshaped to $(13, 5, 40, 1)$ to extract spatiotemporal features.

Model Architecture

Three CNN models were developed with identical network structures but differing only in kernel dimensionality:

• Structure: Two convolutional blocks (each with 2 conv layers, Batch Norm, Max Pooling, Dropout) followed by a multi-head dense layer (4 heads) to output binary spike probabilities.
• Hyperparameters:
  • 1D CNN: 128 filters in early layers; ReLU activation.
  • 2D/3D CNNs: Reduced filter counts (64/32) to manage parameter size; Leaky-ReLU activation.
  • Training: Trained using varying subset sizes ($n=1$ to $19$) from the training data. 3-fold cross-validation was used.
• Evaluation Metrics:
  • Generalizability: Correlation coefficient ($R$) and Root Mean Square Error (RMSE) between CNN-decoded CST and BSS-derived CST.
  • Efficiency: Inference time per sample on CPU and GPU.

3. Key Contributions

1. Systematic Comparison: First study to directly compare 1D, 2D, and 3D CNNs for HD-sEMG neural drive decoding using a unified architecture and independent evaluation datasets.
2. Generalizability Analysis: Quantified how kernel dimensionality affects performance across unseen subjects, contraction intensities, and muscle groups.
3. Efficiency Benchmarking: Provided concrete inference time data on both CPU and GPU, highlighting the computational burden of 3D models in resource-constrained environments.
4. Data Saturation Insights: Investigated the amount of training data required to saturate performance for different architectures, challenging the assumption that complex models always require significantly more data.

4. Key Results

Generalizability

• Cross-Intensity:
  • Low Intensity (10% MVC): 3D CNN showed the highest correlation ($R$) but the highest RMSE (amplitude error), likely due to noise sensitivity. 1D CNN had lower $R$ but smaller amplitude errors.
  • Medium Intensity (30% MVC): 2D and 3D CNNs outperformed 1D CNN in both $R$ and RMSE.
  • High Intensity (50% MVC): 1D CNN achieved the best overall performance (highest $R$, lowest RMSE), while 3D CNN performance degraded.
• Cross-Muscle:
  • Gastrocnemius Medialis (GM): All models performed best; 3D CNN had slightly higher $R$.
  • Gastrocnemius Lateralis (GL): All models performed worst with high inter-subject variability.
  • Soleus (SOL): Intermediate performance.
• Conclusion: Increasing kernel dimensionality did not consistently yield superior generalizability. The optimal architecture depended heavily on the specific contraction intensity and muscle.

Computational Efficiency

• CPU Performance: 3D CNN was significantly slower (4.1 ms/sample) compared to 1D (0.5 ms) and 2D (~0.6 ms).
• GPU Performance: GPU acceleration reduced inference times for all models to a similar range (0.8–1.2 ms). The 3D CNN saw the largest relative improvement (~70% reduction vs. CPU).
• Implication: While 3D CNNs are feasible on GPUs, they impose a heavy burden on CPU-based embedded systems common in prosthetics.

Training Data Saturation

• All architectures reached performance saturation with similar training set sizes (approx. 5–11 subsets) when measured by correlation ($R$).
• However, 2D and 3D CNNs required larger datasets to saturate RMSE (accuracy of magnitude) compared to 1D CNNs.

Artifacts

• False Positives: 2D and 3D CNNs occasionally produced spurious neural drive predictions during rest periods, whereas 1D CNNs remained clean. This is a critical safety concern for real-time control.

5. Significance

• Paradigm Shift: The study challenges the prevailing assumption that higher-dimensional (3D) CNNs are inherently superior for HD-sEMG decoding. It demonstrates that simpler 1D and 2D architectures can achieve comparable generalizability with significantly lower computational costs.
• Practical Deployment: For real-time Neural-Machine Interfaces (NMIs) operating on CPU-based hardware (e.g., wearable prosthetics), 1D or 2D CNNs are recommended due to their superior efficiency and robustness against false positives during rest.
• Design Guidance: The results suggest that for many NMI applications, preserving the temporal waveform structure (via 1D) or spatial topology (via 2D) is sufficient, and the added complexity of 3D convolutions may not justify the computational overhead unless specific high-SNR conditions and GPU resources are available.
• Future Directions: Highlights the need for better decomposition quality for specific muscles (like GL) and further investigation into low-SNR decoding mechanisms.