⚡ electrical engineering

A methodology to rank importance of frequencies and channels in electromyography data with Decision Tree classifiers

This study proposes a transparent methodology using Decision Tree classifiers to identify the most informative frequencies and channels in vastus lateralis EMG data, demonstrating that a streamlined subset of features can effectively evaluate muscle recovery during squat exercises with varying rest intervals.

Original authors: Albert A. Nasybullin, Nursultan Abdullaev, Maksim A. Baranov, Viacheslav V. Koshman, Vitaly A. Mahonin

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Albert A. Nasybullin, Nursultan Abdullaev, Maksim A. Baranov, Viacheslav V. Koshman, Vitaly A. Mahonin

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 your muscles are like a busy orchestra. When you exercise, they don't just make noise; they play a complex, chaotic symphony of electrical signals. Scientists call this Electromyography (EMG).

The problem is, this symphony is incredibly loud and messy. It's like trying to hear a single violin in a stadium full of cheering fans. Researchers want to know: "Which specific notes (frequencies) and which specific musicians (channels/electrodes) are actually telling us if the muscle is tired or fully recovered?"

This paper is about building a smart detective to solve that mystery. Here is the story of how they did it, explained simply:

1. The Experiment: The Squat Test

The researchers asked 10 volunteers to do a simple exercise: squat down and stand up, holding each position for three seconds. They did this in cycles.

The twist? After every cycle, the volunteers had to rest for a different amount of time: 1 minute, 5 minutes, or 10 minutes.

The Goal: To see if the "musical signal" from the muscle changes depending on how long they rested. If the muscle is still tired (1-minute rest), the signal sounds different than if it's fully refreshed (10-minute rest).

2. The Detective: A Single Decision Tree

Usually, scientists use massive, complex computer models (like a super-computer brain) to analyze this data. But those models are like "black boxes"—you put data in, and an answer comes out, but you have no idea why the computer made that choice.

In a medical or sports setting, you need to know why. You need transparency.
So, the authors chose a Decision Tree.

The Analogy: Think of a Decision Tree as a simple "20 Questions" game.
- Question 1: "Is the signal at 50Hz loud?"
- If Yes: Go left.
- If No: Go right.
- Question 2: "Is the signal at Channel 3 quiet?"
- ...and so on, until the tree reaches a conclusion: "This is a 1-minute rest!" or "This is a 10-minute rest!"

Because it's just one tree, you can literally look at the branches and say, "Ah! The computer decided it was a 1-minute rest because the signal at 120Hz was weak." It's transparent and easy to understand.

3. The Challenge: Too Many Clues

The data they collected was huge. They measured the "volume" of the signal at hundreds of different frequencies (from 1Hz to 450Hz) across three different sensors on the leg. That's over 1,300 potential clues for every single squat!

Most of these clues are just noise (static). The researchers needed to find the Golden Nuggets—the tiny handful of frequencies that actually matter.

4. The Method: The "Rank Frequency" Game

To find the best clues, they didn't just look at one tree. They built 1,349 different Decision Trees, each time removing a few clues to see what happened.

They then created a special scoring system called "Rank Frequency."

The Analogy: Imagine a talent show with 1,349 judges (the trees).
Every time a specific frequency (like "120Hz on Channel 3") helps a judge make the right guess, it gets a point.
But it's not just about points. The system also asks:
- "How good was the judge overall?" (Did the tree perform well?)
- "How often did this clue appear?"
- "Was this clue the most important one for that specific judge?"

By combining all these scores, they could rank every single frequency and channel from "Most Important" to "Useless Noise."

5. The Big Discovery

Here is the surprising result: You don't need the whole orchestra.

The study found that a very small, specific group of frequencies and channels was enough to tell the difference between a tired muscle and a rested one.

The "Top 10": They identified the top 10 "super-clues." Interestingly, Channel 3 (a specific sensor on the leg) was the star of the show, holding six of the top spots.
The Lesson: You don't need a massive, complicated computer to analyze muscle recovery. You just need the right few notes.

Why Does This Matter?

Transparency: Doctors and coaches can trust the results because they can see exactly which signals led to the conclusion.
Simplicity: Future devices (like wearable fitness trackers or medical sensors) can be much smaller and cheaper. Instead of needing a supercomputer to process 1,300 data points, they might only need to listen to the top 10 "notes."
Efficiency: It proves that sometimes, less is more. By ignoring the noise and focusing on the "Golden Nuggets," we get accurate results faster.

In a nutshell: The researchers taught a simple, transparent computer to listen to muscle signals. They figured out that you don't need to hear the whole noisy stadium to know if a player is tired; you just need to listen to the right few instruments.

1. Problem Statement

Electromyography (EMG) is a critical tool for analyzing muscle activity and assessing recovery in biomedical and sports sciences. However, EMG signals are complex, noisy, and high-dimensional. While ensemble machine learning methods (e.g., Random Forests) often achieve high accuracy in EMG classification, they suffer from a lack of interpretability. In medical and sports contexts, understanding why a model makes a decision (transparency) and identifying which specific signal features (frequencies and channels) are most relevant is paramount.

The study addresses two main challenges:

Feature Selection: Identifying the specific frequencies and electrode channels that are most informative for classifying muscle recovery states.
Interpretability: Developing a model that balances classification performance with the ability to explain feature importance, avoiding the "black box" nature of complex ensembles.

2. Methodology

A. Data Collection and Experimental Design

Participants: The final dataset included 10 healthy volunteers (7 males, 3 females; mean age 24.5 years) after excluding 4 participants due to data quality issues or prior injuries.
Protocol: Participants performed repeated squat-and-stand exercises (holding each position for 3 seconds).
Rest Intervals: The experiment tested three distinct recovery periods (the target classes): 1, 5, and 10 minutes.
Blocks: The experiment consisted of three blocks separated by 10-minute breaks. Within each block, participants performed sequences of exercises with rest intervals matching the block's specific duration ( $k \in \{1, 5, 10\}$ ).
Hardware: A custom 3-channel EMG system was built using an STM32F401 microcontroller with external amplifiers.
- Sampling Rate: 1 kHz.
- Placement: Electrodes were placed on the vastus lateralis muscle (3 channels: positive, negative, and reference).

B. Data Preprocessing

Baseline Correction: Removal of DC offset using a 5-second segment average.
Filtering:
- Bandpass Filter: 20–500 Hz to isolate muscle activity.
- Notch Filter: 50 Hz to eliminate power line noise.
Feature Extraction: The Power Spectral Density (PSD) was calculated for frequencies ranging from 1 Hz to 450 Hz across all 3 channels.
- Dataset Structure: 252 observations (trials) $\times$ 1,351 features (3 channels $\times$ 450 frequencies + 1).
- Class Distribution: Slight imbalance (1-min: 34.5%, 5-min: 33.3%, 10-min: 32.1%).

C. Machine Learning Approach

Model: Single Decision Tree Classifier (CART algorithm via Scikit-learn).
- Rationale: Chosen over ensemble methods for superior interpretability and suitability for small datasets.
Validation Strategy: Stratified K-Fold Cross-Validation (K=3) to handle class imbalance.
Hyperparameter Tuning: A comprehensive Grid Search was performed for every tree, optimizing parameters such as criterion (gini, entropy, log_loss), max_depth, min_samples_split, and max_features.
Iterative Feature Reduction:
1. Trained 1,349 individual trees.
2. Identified the most important feature based on "mean decrease in impurity."
3. Excluded the top feature and retrained the next tree.
4. Repeated until only one feature remained.

D. The "Rank Frequency" Method

To evaluate feature stability across the 1,349 models, the authors proposed a novel ranking function:
$\text{FeatureRank} = \frac{I(f_i)}{4} \cdot \sum_{i=1}^{N} \left( \frac{\text{mean}(f1(cv)_i)}{\max_x f1(cv)_i} + \frac{|f(x \neq 0)_i|}{|f_i|} + \frac{1}{\text{dt\_rank}_i} + \frac{1}{\text{f\_rank}_i} \right)$

Components:
- $I(f_i)$ : Indicator function (1 if informative, 0 otherwise).
- CV Score: Normalized cross-validation performance of the specific tree.
- Feature Density: Ratio of informative features to total features in that tree.
- Tree Rank: Position of the tree among all trees based on performance.
- Feature Rank: Position of the feature within that specific tree.
Goal: To identify features that consistently appear in high-performing models with high importance.

3. Key Results

Classification Performance:
- The best-performing classifier achieved an F1-weighted score of 0.4974, significantly outperforming the baseline of 0.33.
- 50% of the models exceeded an F1 score of 0.42.
- Statistical tests (Kolmogorov-Smirnov, Shapiro-Wilk, Anderson-Darling) confirmed the data distribution was non-normal.
Feature vs. Performance Relationship:
- Increasing the total number of features generally improved classification quality.
- However, increasing the number of informative (non-zero importance) features did not correlate with performance improvements. This suggests that a small subset of highly relevant features is sufficient for accurate classification.
Top Informative Features:
- Out of 1,350 features, only 3 features achieved a "Rank Frequency" score greater than 3.
- Electrode 3 (the third channel) was dominant, accounting for 6 of the top 10 ranked features.
- The top features included both high and low-frequency components, indicating that muscle recovery signatures are distributed across the spectrum but localized to specific channels.

4. Key Contributions

Interpretable Framework: Demonstrated that single Decision Trees can effectively classify muscle recovery states while providing transparent feature importance, addressing the "black box" issue of ensemble models.
Rank Frequency Methodology: Introduced a novel analytical tool to rank EMG frequencies and channels based on their stability and contribution across a large ensemble of single-tree models.
Feature Reduction: Proved that a limited subset of highly informative features is sufficient for accurate classification, paving the way for compact, low-computational-cost models.
Hardware/Software Integration: Validated a custom, low-cost EMG acquisition system (STM32-based) for high-fidelity signal processing.

5. Significance and Future Work

Clinical & Sports Application: The ability to pinpoint specific frequencies and channels allows for the development of streamlined diagnostic tools and real-time muscle recovery monitoring systems.
Hardware Optimization: The findings support the design of specialized, low-power hardware (e.g., FPGAs) that only processes the most critical frequencies, reducing data redundancy and power consumption.
Future Directions: The authors plan to validate this method on larger, open-source EMG datasets and explore ensemble methods that retain interpretability. They also aim to refine the hardware design for field-deployable applications.

Conclusion: The study successfully establishes that interpretable Decision Tree models, combined with a rigorous feature ranking methodology, can effectively identify the critical spectral and spatial characteristics of EMG signals required for assessing muscle recovery, offering a path toward transparent and efficient biomedical diagnostics.