🤖 AI

How does downsampling affect needle electromyography signals? A generalisable workflow for understanding downsampling effects on high-frequency time series

This paper introduces a generalizable workflow that combines shape-based distortion metrics with machine learning classification to systematically evaluate downsampling strategies, demonstrating that shape-aware algorithms effectively preserve diagnostic information in needle electromyography signals while significantly reducing computational load for near real-time neuromuscular disease detection.

Original authors: Mathieu Cherpitel, Janne Luijten, Thomas Bäck, Camiel Verhamme, Martijn Tannemaat, Anna V. Kononova

Published 2026-07-30

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Mathieu Cherpitel, Janne Luijten, Thomas Bäck, Camiel Verhamme, Martijn Tannemaat, Anna V. Kononova

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 a detective trying to solve a mystery by listening to a very fast, high-pitched radio transmission. The signal is so fast and detailed that your recording device is overwhelmed, and your brain can't keep up with the sheer volume of data. This is the daily reality for doctors using needle electromyography (nEMG), a technique where a tiny needle listens to the electrical chatter of your muscles to diagnose diseases like ALS or myopathy. The signals are incredibly rich, but they are also massive, making it hard for computers to analyze them quickly enough to help a doctor in real-time. To fix this, scientists often try "downsampling," which is like taking a high-definition video and shrinking it to a lower resolution to make it easier to stream. But here's the catch: if you shrink the video too much, you might lose the crucial clues needed to solve the mystery. The big question is: how much can you shrink the signal before you lose the diagnosis?

A team of researchers set out to answer this by building a special "quality control" workflow to test different ways of shrinking these muscle signals. They treated the signals like a complex song and tried various methods to cut out notes without ruining the melody. They found that the most common way of shrinking data—simply skipping every few notes while filtering out the high-pitched noise—was actually the best at keeping the medical clues intact. Surprisingly, they discovered that they could shrink the data by a factor of 30 (making it 30 times smaller) without hurting the computer's ability to diagnose the disease, which sped up the analysis by about 60 times. However, they also found that some fancy methods designed to keep the "shape" of the signal looking perfect actually scrambled the important frequency details, making the diagnosis less accurate. The study suggests that for this specific type of muscle signal, we don't need the super-high-definition version to get the job done, opening the door for faster, real-time medical tools.

The Detective's Dilemma: Listening to Muscle Music

To understand what these researchers did, we first need to understand the "music" they are listening to. When you move a muscle, your nerves send electrical sparks to tell it to contract. A needle electromyography (nEMG) test listens to these sparks. It's like putting a microphone inside a muscle to hear the tiny electrical "pop" of a single motor unit (a nerve and the muscle fibers it controls). These sounds are incredibly fast and detailed, recorded at a rate of about 23,437.5 measurements per second.

The problem is that this amount of data is a heavy burden for computers. If a doctor wants to use a computer program to help diagnose a disease in real-time, the computer has to crunch through millions of these data points. It's like trying to read a library of books in a single second. To make this possible, scientists use downsampling. Think of downsampling like taking a high-resolution photo and resizing it to fit on a small phone screen. You throw away some pixels to make the file smaller and easier to load. In the world of signals, this means keeping only every k-th measurement and throwing the rest away.

But there's a risk. If you throw away too much, you might lose the specific "notes" that tell a doctor whether a muscle is healthy or diseased. Some diseases make the electrical signals look short and spiky, while others make them long and wavy. If your downsampling method smooths out those spikes or stretches out the waves, the computer might get confused and give the wrong diagnosis. The researchers wanted to find the "sweet spot": the point where the file is small enough to be fast, but still big enough to be accurate.

The Experiment: Testing the Shrink Rays

The researchers didn't just guess; they built a systematic workflow to test five different "shrink rays" (downsampling algorithms) on a dataset of real muscle recordings from 22 patients. These patients fell into three groups: healthy controls, people with myopathy (muscle disease), and people with ALS (a nerve disease).

They tested five different ways to shrink the data:

Decimate: The standard method. It picks every k-th point but uses a filter first to stop high-frequency noise from turning into fake low-frequency noise (a bit like using a sieve to catch the big rocks before pouring the sand).
MinMax: This method looks at chunks of data and keeps only the highest and lowest points, trying to preserve the "height" of the waves.
M4: A smarter version of MinMax that tries to connect the dots better so the shape doesn't look jagged.
LTTB (Largest Triangle Three Buckets): A method that tries to keep the overall "area" or shape of the signal by picking points that form the biggest triangles.
MinMaxLTTB: A hybrid that picks the highest and lowest points first, then uses the LTTB method on the rest.

For each method, they tried shrinking the data by different amounts (factors of 2, 5, 10, all the way up to 1000). Then, they ran a computer program (a Random Forest classifier) to see if it could still tell the difference between healthy, myopathic, and ALS signals after the data was shrunk.

The Results: Speed vs. Shape

The results were surprising and very practical. The researchers found that the Decimate method was the clear winner. It could shrink the data by a factor of 30 (keeping only 1 out of every 30 measurements) and still maintain the same level of diagnostic accuracy as the original, full-sized data. This resulted in a 60-fold speedup in feature extraction time. In plain English, the computer could analyze the muscle signals 60 times faster without losing any ability to spot the disease.

However, the other methods, which were designed to look "prettier" or preserve the visual shape of the signal, didn't do as well.

The LTTB and MinMaxLTTB methods could only shrink the data by a factor of 25 before the accuracy started to drop.
The M4 method failed even sooner, at a factor of 20.
The MinMax method also hit a wall at 25.

Why did the "ugly" standard method (Decimate) win? The researchers used a special tool called SHAP analysis to peek inside the computer's brain. They discovered that the computer didn't care as much about the exact "height" or "shape" of the spikes (which the fancy methods tried to preserve). Instead, the computer relied heavily on the frequency of the signal—how fast the waves were vibrating—and how often the signal crossed the zero line. The standard Decimate method, with its anti-aliasing filter, preserved these frequency details perfectly. The fancy shape-preserving methods, however, accidentally distorted these frequencies, confusing the computer.

The "Myopathy" Bottleneck

One interesting detail the study found is that the "Myopathic" (muscle disease) group was the most sensitive to shrinking. The computer could handle shrinking the healthy and ALS signals quite a bit more, but the myopathic signals started to get confused at a factor of 30. This makes sense because myopathic signals are often very short and spiky. If you shrink the data too much, those tiny spikes get lost or merged together. This suggests that for patients with muscle diseases, doctors need to be extra careful not to shrink the data too aggressively.

What This Means for the Future

The study concludes that we don't need to keep every single data point to diagnose these diseases. By using the standard Decimate method and shrinking the data by 30 times, we can make automated analysis 60 times faster. This is a huge step toward "near real-time" analysis, where a computer could help a doctor instantly while they are still holding the needle.

However, the authors are careful to note that this is based on one specific dataset (EMGLAB) and one specific type of muscle contraction (low-level voluntary movement). They suggest that while their workflow is a great tool for testing other signals, the specific "30x" rule might change if they tested different types of muscle activity, like resting muscles or maximum effort, which might have different high-frequency clues. They also emphasize that while the computer gets faster, the final decision should still involve a human doctor who can look at the signal and say, "Yes, this still looks like a muscle disease to me."

In short, the paper shows that sometimes, less is more. By throwing away 97% of the data points in a smart way, we can make medical AI faster and just as accurate, bringing us one step closer to instant, automated muscle diagnostics.

Technical Summary: A Generalisable Workflow for Understanding Downsampling Effects on Needle Electromyography Signals

Problem Statement
Automated analysis of needle electromyography (nEMG) signals is a promising tool for detecting neuromuscular diseases (NMDs). However, nEMG signals possess high temporal resolution (e.g., ~23 kHz) and heterogeneous sampling rates, creating substantial computational burdens for feature-based machine learning (ML) models. This hinders near real-time analysis and clinical implementation. While downsampling offers a solution to reduce computational load, its impact on diagnostic signal content and classification performance remains insufficiently understood. Previous studies in related biomedical fields (EEG, ECG, surface EMG) have shown heterogeneous results, ranging from performance degradation to enhancement, yet no standardized workflow exists to systematically evaluate acceptable levels of information loss across different downsampling algorithms and factors.

Methodology
The authors propose a five-step generalisable workflow to systematically evaluate information loss caused by downsampling in high-frequency time series. This workflow was applied to the EMGLAB dataset, containing 22 patients with three diagnostic classes: Control, Myopathic, and Amyotrophic Lateral Sclerosis (ALS).

Configuration Definition: Five downsampling algorithms were tested: Decimate (uniform subsampling with anti-aliasing filtering), MinMax, M4, Largest Triangle Three Buckets (LTTB), and MinMaxLTTB. These were applied across 26 downsampling factors ( $k$ ), ranging from 2 to 1000.
Distance Metric Computation: For each configuration, distance metrics were computed between the original and downsampled signals. These included time-domain metrics (RMSE, correlation coefficients, zero-crossing rate, peak counts, skewness, kurtosis), frequency-domain metrics (Log Spectral Distance), and information-theoretic metrics (Normalized Compression Distance, Jensen-Shannon Divergence).
Classification Pipeline: A supervised classification task was performed using a pipeline featuring tsfresh for feature extraction and a Random Forest classifier. Performance was evaluated using balanced accuracy, per-class F1-scores, specificity, sensitivity, and ROC AUC, utilizing 5 repetitions of 3-fold stratified group cross-validation at the patient level.
Ranking Model: A pairwise ranking model (XGBoost) was trained to correlate distance metrics with classification performance. This model predicted which configuration would yield better accuracy based solely on distance metrics. SHAP (SHapley Additive exPlanations) values were extracted to identify which signal characteristics (e.g., frequency content vs. amplitude distribution) were most critical for the classifier.
Feature Space and Cost Analysis: The study analyzed shifts in feature importance using k-means clustering and dimensionality reduction (MDS) to visualize how feature relevance evolved with downsampling. Additionally, feature extraction time was measured to quantify the computational speedup trade-off.

Key Results

Algorithm Performance: On the EMGLAB dataset, Decimate (with anti-aliasing filtering) demonstrated the highest resilience, maintaining classification performance up to a downsampling factor of 30. Shape-aware algorithms (LTTB, MinMaxLTTB, MinMax) reached their critical performance-degradation points at factors of 25, while M4 degraded significantly earlier (factor 20).
Critical Limiting Factor: The Myopathic class was the most sensitive to downsampling, acting as the limiting factor for all strategies. This is clinically plausible as myopathic motor unit action potentials (MUAPs) are often short-duration and polyphasic, requiring higher sampling frequencies to resolve.
Signal Characteristics: SHAP analysis revealed that classification performance relied heavily on preserving frequency domain characteristics (specifically Log Spectral Distance and Zero Crossing Rate) rather than strict preservation of amplitude distribution (skewness/kurtosis). This suggests that shape-aware algorithms, which prioritize amplitude extremes, may not retain the specific signal properties necessary for NMD diagnosis in this context.
Computational Trade-off: Downsampling provided significant computational gains. The optimal configuration (Decimate at factor 30) achieved a ~60-fold speedup in feature extraction time with no statistically significant loss in balanced accuracy compared to the original signal. Feature extraction time decreased exponentially up to a factor of 50, after which gains plateaued.
Feature Space Stability: At low downsampling factors (up to 5), feature importance trajectories were similar across algorithms. Beyond this, Decimate diverged significantly from shape-aware methods. High downsampling factors increased the variance of feature importance across cross-validation folds, indicating reduced model stability.

Significance and Claims
The paper claims to fill a critical gap in the literature by providing a standardised, reproducible workflow to balance data reduction with model performance in high-frequency time-series applications. The authors assert that their findings offer practical guidance for selecting downsampling configurations that enable near real-time nEMG analysis without sacrificing diagnostic utility.

Specifically, the study demonstrates that:

Oversampling Hypothesis: The results suggest that standard nEMG signals may be oversampled for the specific task of automated classification using tsfresh and Random Forest, as diagnostic performance remains stable even when the effective Nyquist frequency is reduced to ~390 Hz (from a factor of 30 downsampling).
Algorithm Selection: Contrary to the intuition that preserving visual shape (via LTTB/MinMax) is paramount, the study found that uniform decimation with anti-aliasing filtering better preserved the diagnostic information content for this specific dataset and feature set.
Clinical Implementation: The workflow provides a foundation for developing clinically deployable nEMG analysis pipelines that meet real-time processing requirements.

Limitations and Future Directions
The authors modestly acknowledge that their results are specific to the EMGLAB dataset, which contains only moderate voluntary contraction data under controlled conditions and lacks rest segments or real-world clinical noise. Consequently, the optimal factors identified may not generalize to all nEMG modalities or datasets with different sampling frequencies. The study calls for validation on independent datasets and expert clinical assessment of downsampled signals before broader clinical recommendations can be made. The workflow is proposed as a general tool that can be adapted to evaluate other signal morphing techniques (e.g., compression) and applied to other high-frequency biomedical time series.