Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven

arXiv:2605.20599 · cs.LG · Submitted 2026-05-20 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Unsupervised clustering and classification of upper limb EMG signals during functional movements".

Jane: Here's the gist of the paper. The gist The study presents a comprehensive approach for clustering and classification of upper-limb surface electromyography (sEMG) signals during functional reach and grasp movements.

Jane: Signal Preprocessing: The preprocessing involved a fourth-order low-pass filter (0.6 Hz) and Hilbert envelope transformation, which effectively reduced noise and enhanced signal clarity.

Title and authors: Tom: To start, let's just talk about the title itself. "Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven." It tells you they are grouping similar movements together without being told beforehand what those movements are.

Jane: So, if you’re driving or cooking, you might move your hand in a few different ways to reach for something. This paper is trying to automatically figure out those distinct muscle activation patterns and categorize them based on how the signals look.

Lu: The core of the methodology involves signal preprocessing using a fourth-order low-pass filter at zero point six Hz, followed by a Hilbert envelope transformation to clean up the noise and make the signal clearer for feature extraction > <ref:2605.20599#pg1,a fourth-order low-pass filter>

Meng: Cleaning up that noise is always necessary when you're dealing with real-world sensors. What's next in their process? I need to know what kind of measurements they are pulling out of that cleaned signal.

Tom: After preprocessing, they extract twenty-six temporal and frequency-domain metrics <ref:2605.20599#pg1,26 temporal and frequency-domain metrics>. These are the raw numbers they start with before any fancy selection happens >

Jane: Twenty-six metrics sounds like a lot of data points for every single movement recorded on those fifty-two gestures, right <ref:2605.20599#pg1>? That's where the complexity starts to build up.

Lu: And those initial twenty-six attributes are then refined using several techniques, including visual analysis, mutual information, principal component analysis, and decision tree importance scores to filter out what’s actually useful > <ref:2605.20599#pg1>

Meng: Filtering is key for deployment. If you're running this on a device, you don't want all twenty-six metrics; you want the ones that matter most for accurate classification.

Lalam: They also selected the gestures themselves using hierarchical clustering based on Mahalanobis distance to find six representative movements, which they say balances biomechanical diversity with computational efficiency >

The paper's summary: Tom: So, summarizing what we just covered: this paper takes raw EMG data from upper limbs during functional reach and grasp movements and uses a four-stage pipeline to cluster and classify those signals. It’s moving from raw signal to a set of recognizable gestures >

Jane: Imagine you have a bunch of recordings, and instead of labeling every single one by hand, this system groups them into six distinct classes based on the underlying muscle activation patterns. That's the unsupervised clustering part >

Lu: The paper focuses specifically on DB4 from the NINAPRO dataset, which is designed for classification and regression tasks in gesture recognition, meaning their focus is very targeted here >

Meng: It’s interesting that they are focusing on a specific subset of data. When I try to deploy something at scale, I worry about generalization; does this clustering work well if the person using it has slightly different muscle patterns?

Tom: The paper addresses that by selecting six representative movements through hierarchical clustering and then using a two hundred ms window for temporal segmentation, which they say is optimal because the central nervous system organizes those patterns in timeframes close to two hundred milliseconds > <ref:2605.20599#pg2>

Jane: So, that segmentation window is important. It means they aren't just looking at a single burst of muscle activity; they are looking at a small sequence of movement >

Lu: And they chose two hundred ms because it offers a balance between resolution and stability, which helps in inter-class discrimination later on > <ref:2605.20599#pg2>

Meng: If you were running this on my production system, I'd ask about the computational cost of processing those segments. How much does that two hundred ms window add to the latency <ref:2605.20599#pg2>?

Lalam: The paper notes that they also used a two hundred ms window for temporal segmentation because it was deemed the most suitable option based on stability and physiological plausibility > <ref:2605.20599#pg2>

The paper's improvements: Tom: Now, let's look at what the authors suggest improving about their initial approach. They didn't just stop at clustering; they refined feature selection significantly to reduce noise and redundancy.

Jane: So, they took those twenty-six metrics and used a hybrid scheme involving visual analysis, mutual information, principal component analysis, and decision tree importance scores to narrow down what’s actually important >

Lu: They made some hard cuts there too; they eliminated all standard deviation features and six additional low-relevance features that didn't contribute much to the final classification >

Meng: Cutting stuff out is smart. It reduces the feature set size, which means faster processing on any hardware. But what about those channels? They dropped channels eleven and twelve because they had consistently low importance > <ref:2605.20599#pg3>

Tom: That leaves a final, refined feature set consisting of IAV, MAV, MAVS, RMS, and mDWT applied across the ten most relevant channels >

Jane: It’s like they distilled all that initial complexity down to a much smaller group of metrics that are still highly discriminative for the six selected gestures >

Lu: That final set of five features across ten channels is what allows them to achieve classification with high accuracy, close to ninety-five percent, using models like Extra Trees and Artificial Neural Networks >

Conclusion: Tom: So, wrapping up the paper "Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven," they’ve shown a comprehensive approach that moves from raw signal to six representative gestures using a careful selection of features and segmentation windows >

Jane: The main implication is that we can build robust systems for recognizing these functional movements without needing manual labeling for every single movement first. It's driven by clustering and data-driven refinement >

Lu: The work on gesture selection using Mahalanobis distance to select those six clusters, aiming for maximum intra-cluster homogeneity, provides a solid framework that we can adapt when dealing with different datasets or new movement types >

Meng: For deployment, the focus on an efficient feature set and strong classifiers like Extra Trees and Artificial Neural Networks gives us something concrete. The system is designed to be fast enough for real-world applications >

Lalam: And from my side, I confirm that the combination of the two hundred ms window for temporal segmentation and that final feature set makes the classification quite stable across different subjects in their test data > <ref:2605.20599#pg2>

Tom: It’s a solid piece of work, focusing on building a reliable recognition system. We'll be keeping an eye on how these results translate into actual myoelectric control strategies for prostheses >

Jane: Definitely. So that’s the rundown of this paper, "Unsupervised clustering and classification of upper limb EMG signals during functional movements: a data-driven." Next time we talk, we'll look at something completely different in the world of AI research.

Salazar Ávez L. F., Escobar-Saltarຑn D., Salazar Sch༼hez M. B., Henao-Aguirre S. C.

Universidad de Antioquia · Tecnológico de Monterrey

cs.LG

Submitted: 2026-05-20

Updated: 2026-05-20

Comments: 19 Congreso Colombiano de Computación (19CCC)

Journal ref: Salazar Alvarez et al., Advances in Computing, CCC 2025, CCIS 2692, pp. 158-171 (2027)

DOI: 10.1007/978-3-032-28447-1_11

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: The gist The study presents a comprehensive approach for clustering and classification of upper-limb surface electromyography (sEMG) signals during functional reach and grasp movements.

Key concepts

Signal Preprocessing: This step cleans the raw EMG data to remove noise and improve signal quality. It involves filtering out electrical interference (like 60 Hz mains noise) using a low-pass filter and then transforming the signal using the Hilbert envelope. This process ensures that only relevant muscle activation information remains for accurate feature extraction.
Hierarchical Clustering with Mahalanobis Distance: This technique was used to group similar movements together based on their EMG patterns. Mahalanobis distance is a specific mathematical measure that considers the variance and correlation between different features, making it better at finding distinct movement groups even when data is complex or redundant. This helped select six representative gestures.
Temporal Segmentation Window: This refers to the specific time frame (200 ms) chosen to analyze each segment of muscle activity. This window was selected because it balances the need for detailed temporal resolution with stability, aligning with how the central nervous system organizes posture-specific muscle activation patterns during movement.

Terminology

Clustering — 4× in this paper · explained in 2 episodes
Artificial Neural Networks — 1× in this paper
Clarity — 1× in this paper · explained in 3 episodes
Extra Trees — 1× in this paper
Feature Extraction — 1× in this paper · explained in 3 episodes
Generalization — 1× in this paper · explained in 13 episodes
Mutual Information — 1× in this paper · explained in 2 episodes

Summary

The gist The study presents a comprehensive approach for clustering and classification of upper-limb surface electromyography (sEMG) signals during functional reach and grasp movements.

How it works

Signal Preprocessing: The preprocessing involved a fourth-order low-pass filter (0.6 Hz) and Hilbert envelope transformation, which effectively reduced noise and enhanced signal clarity.
Feature Extraction: Feature extraction yielded 26 temporal and frequency-domain metrics, which were later refined using visual analysis, mutual information, principal component analysis, and decision tree importance scores.
Gesture Selection: Gesture selection was performed through hierarchical clustering using Mahalanobis distance, resulting in six representative movements that balanced biomechanical diversity and computational efficiency.
Temporal Segmentation: A 200 ms window was identified as optimal for temporal segmentation based on stability and physiological plausibility.

Methodology Details

The methodology was organized into four main stages: (A) signal preprocessing and transformation, (B) feature extraction, (C) movement selection by cluster analysis, and (D) classifier selection and validation.

Signal preprocessing involved several steps to attenuate noise and ensure inter-subject comparability:

a) Electrical-noise filtering: mains interference (60 Hz) was removed.

b) Butterworth low-pass filter: a fourth-order digital filter with a 0.6 Hz cut-off frequency smoothed abrupt variations without affecting the low-frequency content relevant to muscle activation.

c) Envelope calculation (Hilbert): the Hilbert transform was used to obtain the analytic envelope of each channel, providing a clearer and more continuous representation of the signal.

d) Subject- and channel-wise normalization: each signal was scaled to reduce intrinsic variability between individuals.

Feature extraction involved extracting statistical features from the processed signals in both the time and frequency domains, including mean absolute value (MAV), root mean square (RMS), integrated absolute value (IAV), and wavelet-based mDWT. Initial metrics included 26 attributes per signal segment.

Movement Selection and Feature Refinement

The selection of a representative, non-redundant subset of functional movements was a key phase in the development of this project. Gesture similarity was assessed via hierarchical cluster analysis, with Mahalanobis distance chosen as the proximity metric. The clustering strategy employed complete linkage. This resulted in six major clusters that satisfied maximum intra-cluster homogeneity and minimum inter-cluster similarity. Within each macro-cluster, the most representative movement was selected based on the highest energy (based on RMS analysis) and greatest potential for replicability.

Feature selection utilized a hybrid scheme combining exploratory inspection, statistical indicators, and model-derived criteria. This process led to the elimination of all standard deviation features, six additional low-relevance features, and channels 11 and 12 due to their consistently low importance. The final feature set consisted of IAV, MAV, MAVS, RMS, and mDWT applied across the 10 most relevant channels.

Classifier Models Evaluation

The recognition of the six selected gestures was carried out in two complementary phases to combine speed and precision.

a) PyCaret: The automated comparison using PyCaret identified Extra Trees (ET) and Artificial Neural Networks (ANN) as top performers.

b) Independent Testing: Subsequent independent training confirmed their stability and generalization capacity, with ANN showing progressive learning and ET maintaining robust, consistent results. The results showed that ANN and ET models outperformed the others, achieving accuracy close to 95%.

Optimal Parameters

The analysis of temporal segmentation evaluated three non-overlapping rectangular window lengths: 100 ms, 200 ms, and 300 ms. The 18th page indicated that the 200 ms window emerged as the most suitable option. This choice was supported by physiological literature showing that the central nervous system can organize posture-specific muscle activation patterns within time frames close to 200 ms.

Improvements for AI systems

Preprocessing Strategy for Noise Reduction: Implement a two-step preprocessing procedure involving a fourth-order low-pass filter (0.6 Hz) to remove mains interference and Hilbert envelope transformation to enhance signal clarity, which effectively reduc[es] noise and enhancing signal clarity. This results in cleaner input data for subsequent feature extraction.
Movement Selection via Mahalanobis Distance: Use hierarchical clustering with Mahalanobis distance as the proximity metric to select representative movements, as this technique accounts for the variance–covariance structure of the features, penalizing highly correlated dimensions and thus providing a more robust criterion in the presence of internal redundancies. This ensures a selection that minimizes overlap in feature space.
Optimal Temporal Segmentation Window: Employ a 200 ms window for temporal segmentation, as this size emerged as the most suitable option, providing a balance between resolution and stability and is supported by evidence that the central nervous system can organize posture-specific muscle activation patterns within time frames close to 200 ms. This leads to better inter-class discrimination in later processing stages.
Hybrid Feature Selection Pipeline: Utilize a hybrid scheme for feature selection combining visual analysis, mutual information, PCA, and decision tree importance scores to ensure the final feature set captures muscle activity without unnecessary redundancies. The resulting efficient feature set (IAV, MAV, MAVS, RMS, and mDWT) allows for reduced computational load while maintaining high discriminative power.
Robust Classifier Selection: Employ Extra Trees (ET) and Artificial Neural Networks (ANN) as top performers based on PyCaret evaluation and independent testing to achieve high accuracy near 95%. This system can accurately classify functional reach and grasp movements with stability, supporting the implementation of adaptive, low-latency control strategies for myoelectric prostheses.

Abstract

This study presents a comprehensive approach for the clustering and classification of upper-limb surface electromyography (sEMG) signals during functional reach and grasp movements. The methodology was applied to the NINAPRO DB4 dataset, which provides multichannel EMG recordings of 52 gestures. A four-stage pipeline was designed, including signal preprocessing, fea-ture extraction, gesture selection via hierarchical clustering, and comparative model evaluation. Preprocessing involved a fourth-order low-pass filter (0.6 Hz) and Hilbert envelope transformation, effectively reducing noise and enhancing signal clarity. Feature extraction yielded 26 temporal and frequency-domain met-rics, which were later refined using visual analysis, mutual information, principal component analysis, and decision tree importance scores. A final subset of five key features was selected for classification tasks. Gesture selection was per-formed through hierarchical clustering using Mahalanobis distance, resulting in six representative movements that balanced biomechanical diversity and compu-tational efficiency. A 200 ms window was identified as optimal for temporal seg-mentation based on stability and physiological plausibility. Classifier models were evaluated in two stages. Automated comparison using PyCaret identified Extra Trees (ET) and Artificial Neural Networks (ANN) as top performers. Sub-sequent independent training confirmed their stability and generalization capac-ity, with ANN showing progressive learning and ET maintaining robust, con-sistent results. The findings support the implementation of adaptive, low-latency control strategies for myoelectric prostheses and provide a scalable pipeline for future real-time applications.