EMG fatigue detection

[west pike]

Hey everyone! 👋

I'm building a muscle fatigue classification system using the Ninapro DB2 dataset. I’ve extracted features like RMS, MAV, waveform length, mean/median frequency, then trained an XGBoost model with PCA-reduced features on sliding windows of EMG data.

I'm currently simulating fatigue labels by assuming:

Repetitions 1–2 = non-fatigued

Repetitions 3–5 = fatigued

How can I generate dynamic fatigue labels based on actual signal changes across repetitions?

Should I track percent change in RMS or median frequency? What threshold ranges make sense for true fatigue? and how can i improve accuracy and auc?

[west pike]

first of all thanks for response, your reply is good for llm or nlp type problem but mine is completely irrelevant I'm making EMG fatigue detection using xgboost with pca.

or is there is something i did not catch

[opaque gyro]

No, I think I submitted this to the wrong one 😂 It's what happens when I chat too late

[west pike]

do you have any experience in my problem

[opaque gyro]

Not directly, I'm afraid. Yeah, this is the thread I meant to answer: #1376453080845127803 message

I'll take a look at it, though, see if maybe I can crack it

So this dataset covers 49 hand movements, and fatigue status is apparently inferred rather than explicitly labeled, probably from repetition and signal count, right?

[west pike]

yes exactly

[opaque gyro]

Are you familiar with Root Mean Square as a concept?

"Root Mean Square (RMS) is a fundamental amplitude-domain feature that quantifies the overall power or strength of the EMG signal. It serves as an indicator of the total electrical activity within the muscle and is directly related to the number of active motor units and their firing rates. As muscle fatigue progresses during sustained contractions, the RMS value typically increases. This rise reflects the neuromuscular system's compensatory strategy to maintain force output by recruiting additional motor units or increasing the firing frequency of already active units. This increase in RMS is not a direct measure of fatigue itself, but rather a measure of the compensatory effort the neuromuscular system is making to counteract the effects of fatigue and maintain performance. It signifies the body's struggle to maintain force output. For dynamic labeling, a significant increase in RMS, particularly when observed in conjunction with other fatigue indicators, is a strong sign that the muscle is undergoing stress and working harder. However, relying solely on RMS might not capture the earliest signs of fatigue if the compensatory mechanisms are highly efficient or if the task is submaximal, as the muscle might be fatiguing metabolically without a strong compensatory amplitude increase."

"Median Frequency (MDF) and Mean Power Frequency (MNF) are critical frequency-domain features derived from the power spectral density of the EMG signal. They represent the central tendency of the signal's frequency content. MDF specifically is the frequency value that divides the power spectrum into two equal energy halves. A hallmark of muscle fatigue is a characteristic shift of the EMG power spectrum towards lower frequencies, which is quantitatively reflected as a decrease in both MDF and MNF. This spectral compression is primarily attributed to a reduction in muscle fiber conduction velocity and alterations in motor unit firing rates as fatigue accumulates. MDF is widely regarded as a highly sensitive indicator of muscle fatigue. It is often considered more reliable than RMS for detecting fatigue, as it more directly reflects intrinsic physiological changes within the muscle fibers, such as a decrease in muscle fiber conduction velocity and shifts in motor unit firing rates. Notably, MDF can show a continuous decrease even when the external force output remains constant, indicating ongoing metabolic fatigue that might not yet be reflected in amplitude changes. This implies that MDF tracks the intrinsic physiological changes occurring within the muscle fibers themselves (e.g., accumulation of metabolic byproducts, altered membrane excitability, reduced conduction velocity) that are the root cause of fatigue. In contrast, RMS primarily tracks the compensatory neural strategy. Therefore, MDF may offer a more direct and potentially earlier indicator of the fatigue process itself, independent of the external force output or the compensatory efforts made by the neuromuscular system. For dynamic labeling, MDF's superior sensitivity makes it a prime candidate for detecting the onset of fatigue. Its ability to show continuous changes even at constant force suggests it can track fatigue progression more granularly than RMS alone, providing a more nuanced view"

"Which Feature to Track: A Detailed Comparison (RMS vs. MDF efficacy and robustness)
The scientific literature consistently supports that a decrease in MDF/MNF coupled with a concurrent increase in RMS during sustained or repetitive contractions are the most reliable and commonly accepted indicators of muscle fatigue. Both parameters are "well known in the field and clinically used in movement and rehabilitation science". MDF is generally considered more sensitive to the underlying physiological changes of fatigue, reflecting the spectral shift due to altered muscle fiber conduction velocity. It can provide early indications of fatigue even before significant changes in force or compensatory amplitude are apparent. RMS provides crucial information about the overall signal strength and the level of motor unit recruitment. Its increase is a direct reflection of the body's compensatory mechanisms to maintain force in the face of fatiguing muscle fibers.
Rather than choosing one feature over the other, the most robust and comprehensive approach to dynamic fatigue labeling is to use both RMS and MDF (or MNF) as complementary indicators. Their combined analysis offers a more holistic picture of fatigue, capturing both the intrinsic physiological changes within the muscle and the dynamic neuromuscular control strategies employed by the body. The overwhelming evidence from the literature advocates for a "both/and" approach. The combined and often synergistic trend of these features provides a far more robust and physiologically accurate signature of fatigue than either feature analyzed in isolation. This implies that one should not choose between RMS and MDF, but rather track both and potentially analyze their inter-relationships or their respective rates of change. This understanding reinforces the value of sophisticated feature engineering beyond simple, raw metrics."

"Features that capture the dynamics and interplay of these primary indicators—such as ratios (e.g., MNF/ARV ratio ), instantaneous differences (e.g., Instantaneous Mean Amplitude Difference ), or even time-frequency representations (e.g., CWT )—will be significantly more powerful for accurate and nuanced fatigue classification.
It is important to note that the relative suitability or stability of MNF versus MDF might vary depending on the specific muscle group under investigation and the nature of the exercise protocol. For instance, some studies suggest MNF might exhibit lower variability than MDF in certain all-out cycling tests. For the Ninapro DB2 dataset, which involves a wide array of hand and finger movements, evaluating both and observing their combined trends is a prudent strategy."

I've been using Gemini to aggressively pull together some questions about how people are using the dataset you use, and this was the best response in terms of what should work, because it gives you two methods that compliment each other in terms of the data they collect.

References:
Fatigue detection in EMG signals.,
https://fenix.tecnico.ulisboa.pt/downloadFile/395137878684/artigo.pdf 2. Machine Learning
Physical Fatigue Estimation Approach Based on IMU and EMG Wearable Sensors - UCC:
CORA, https://cora.ucc.ie/bitstreams/d51037fa-d7f0-4b3e-af17-9c95141968f2/download 3.
Effects of Force Load, Muscle Fatigue, and Magnetic Stimulation on Surface Electromyography
during Side Arm Lateral Raise Task: A Preliminary Study with Healthy Subjects,
https://pmc.ncbi.nlm.nih.gov/articles/PMC5405568/ 4. Detecting muscle fatigue during lower
limb isometric contractions tasks: a machine learning approach - PubMed Central,
https://pmc.ncbi.nlm.nih.gov/articles/PMC11965937/ 5. Classifying static muscle fatigue: an
surface electromyography signal analysis approach - SPIE Digital Library,
https://www.spiedigitallibrary.org/conference-proceedings-of-spie/13208/1320812/Classifying-sta
tic-muscle-fatigue--an-surface-electromyography-signal-analysis/10.1117/12.3036611.full?webS
yncID=d8e8abee-aed6-ae05-c117-aac5a9199362&sessionGUID=b9eb3b36-c8a9-5604-e3dd-8
21ae3dced0d 6. Developing a Novel Muscle Fatigue Index for Wireless sEMG Sensors: Metrics
and Regression Models for Real-Time Monitoring - Preprints.org,
https://www.preprints.org/manuscript/202503.1389/v1 7. Analyzing fatigue in dynamic exercise
through electromyography signals and similarity metrics | Request PDF - ResearchGate,

8. Advanced Analysis and
   Interpretation of the EMG Signal,
   https://www.cometasystems.com/advanced-analysis-and-interpretation-of-the-emg-signal/ 9.
   Muscle Fatigue and Time-Dependent Parameters of the Surface EMG Signal,
   https://www.bu.edu/nmrc/files/2010/03/Muscles-Alive-Ch-8.pdf 10. Surface Electromyography
   Thresholds as a Measure for Performance Fatigability During Incremental Cycling in Patients
   With Neuromuscular Disorders - PubMed Central,
   https://pmc.ncbi.nlm.nih.gov/articles/PMC8969223/ 11. Developing a Novel Muscle Fatigue
   Index for Wireless sEMG Sensors: Metrics and Regression Models for Real-Time Monitoring -
   MDPI, https://www.mdpi.com/2079-9292/14/11/2097 12. Evaluation of Electromyographic
   Frequency Domain Changes during a Three-Minute Maximal Effort Cycling Test - PMC -
   PubMed Central, https://pmc.ncbi.nlm.nih.gov/articles/PMC4424476/ 13. Detecting fatigue
   thresholds from electromyographic signals: A ...,
   https://www.researchgate.net/publication/305954156_Detecting_fatigue_thresholds_from_electr
   omyographic_signals_A_systematic_review_on_approaches_and_methodologies 14. DB2 -
   Ninapro, https://ninapro.hevs.ch/instructions/DB2.html 15. sEMG-based gesture recognition
   using multi-stream ... - Frontiers,
   https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.202
   5.1487020/epub 16. Muscle Fatigue Detection in EMG Using Time–Frequency Methods, ICA
   and Neural Networks - ResearchGate,
   https://www.researchgate.net/publication/45630828_Muscle_Fatigue_Detection_in_EMG_Using
   _Time-Frequency_Methods_ICA_and_Neural_Networks 17. Advances in Muscle Fatigue
   Detection: A Comprehensive Review,
   http://ir.kdu.ac.lk/bitstream/handle/345/7488/Advances in Muscle Fatigue Detecti
   on%20A%20Comprehensive%20Review.pdf?sequence=1&isAllowed=y

These are references I used to help answer your question

[west pike]

yes i know RMS, and that is quite a work thanks a lot ill try to implement and will share.

[opaque gyro]

Yeah, based on what I understood from it, I'm glad you're on implementation and not me! 😂 I wanted to make sure you got the sources, too, so you could work further. Hope it helps

I may not have domain expertise on it, but I'm fantastic at researching

[opaque gyro]

Just wanted to check in and see how things are progressing

[west pike]

i think its quite gd now