pith. sign in

arxiv: 2605.20599 · v1 · pith:2U5SP55Dnew · submitted 2026-05-20 · 💻 cs.LG

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

L. F. Salazar \'Alvarez , D. Escobar-Saltar\'en , M. B. Salazar S\'anchez , S. C. Henao-Aguirre This is my paper

Pith reviewed 2026-05-21 06:57 UTC · model grok-4.3

classification 💻 cs.LG
keywords sEMG classificationupper limb movementshierarchical clusteringfeature selectionmyoelectric prosthesesNINAPRO DB4Extra Treesneural networks
0
0 comments X

The pith

A pipeline selects five EMG features and six movements to classify upper-limb gestures for prosthetic control.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper develops a data-driven method to cluster and classify surface electromyography signals from upper-limb movements using the NINAPRO DB4 dataset. It preprocesses the signals, extracts and refines features, uses hierarchical clustering to pick six representative gestures, and evaluates classifiers to find stable performers. The approach aims to enable more practical control of myoelectric prostheses by reducing complexity while maintaining accuracy. A reader might care because it offers a concrete path toward real-time, adaptive prosthetic systems that could improve daily function for users.

Core claim

Through a four-stage pipeline applied to multichannel EMG recordings of 52 gestures, the study identifies a subset of five key features and six representative movements via hierarchical clustering with Mahalanobis distance that support effective classification, with Extra Trees and artificial neural networks showing robust performance suitable for low-latency prosthetic control.

What carries the argument

Hierarchical clustering with Mahalanobis distance applied to refined temporal and frequency features to select six biomechanically diverse yet computationally efficient movements from the full set of 52 gestures.

If this is right

  • Adaptive low-latency control strategies for myoelectric prostheses can be implemented based on the identified features and movements.
  • The pipeline provides a scalable foundation for real-time applications in prosthetic devices.
  • Extra Trees classifier maintains consistent results while artificial neural networks demonstrate progressive learning on the selected data.
  • A 200 ms analysis window offers an optimal balance of stability and physiological relevance for segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If these features hold across users, prosthetic calibration time could decrease substantially.
  • The method could extend to other types of biosignals or movement types beyond upper limbs.
  • Real-world testing on actual prosthetic hardware would reveal any gaps between lab data and practical deployment.

Load-bearing premise

The five features and six movements chosen from this particular dataset will work well for new users, different conditions, or real prosthetic hardware without needing much re-tuning.

What would settle it

Apply the same five features and six movements to EMG data from a separate group of subjects using different recording equipment and measure whether the classification accuracy stays above the levels reported here.

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a data-driven four-stage pipeline for preprocessing, feature extraction and selection, gesture clustering, and classification of upper-limb sEMG signals from the NINAPRO DB4 dataset. Preprocessing uses a 0.6 Hz fourth-order low-pass filter and Hilbert envelope; 26 features are reduced to five via visual analysis, mutual information, PCA, and decision-tree importance; six representative gestures are chosen by Mahalanobis-distance hierarchical clustering; and Extra Trees and ANN classifiers are evaluated on 200 ms windows. The central claim is that this pipeline supports adaptive, low-latency myoelectric prosthesis control and provides a scalable approach for real-time applications.

Significance. If the selected five-feature, six-movement subset generalizes, the work supplies a concrete, computationally efficient pipeline for prosthetic control that leverages a public dataset and automated model comparison. The explicit description of preprocessing choices and the two-stage classifier evaluation are strengths that aid reproducibility. However, the absence of reported quantitative metrics weakens the immediate applicability assessment.

major comments (3)
  1. Abstract: the claim that 'the findings support the implementation of adaptive, low-latency control strategies' is load-bearing for the prosthetic application but is unsupported by any accuracy, confusion-matrix, or cross-validation numbers; the abstract only names the models without performance figures.
  2. Gesture selection via hierarchical clustering: the reduction to six movements is performed entirely on NINAPRO DB4 recordings; because this step is data-dependent, the manuscript must demonstrate that the same six movements remain representative under inter-subject variability or electrode shift to justify the scalability claim.
  3. Feature refinement step: the iterative selection of the final five features combines visual inspection, mutual information, PCA, and decision-tree importance on the same dataset; this risks circularity and requires an explicit subject-independent hold-out evaluation to confirm the features are not overfit to the particular recordings.
minor comments (2)
  1. Abstract: hyphenation artifacts ('fea-ture', 'met-rics', 'compu-tational') should be removed for readability.
  2. The statement that a 200 ms window is 'optimal' based on 'stability and physiological plausibility' would benefit from a short quantitative justification or reference to the stability metric used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript. The comments have identified important areas for improvement, particularly regarding the support for our claims and the robustness of our data-driven selections. We provide point-by-point responses to the major comments below.

read point-by-point responses

  1. Referee: Abstract: the claim that 'the findings support the implementation of adaptive, low-latency control strategies' is load-bearing for the prosthetic application but is unsupported by any accuracy, confusion-matrix, or cross-validation numbers; the abstract only names the models without performance figures.

    Authors: We agree that the abstract should include quantitative evidence to support the claims. In the revised version, we will update the abstract to report the classification accuracies for the Extra Trees and ANN classifiers, as well as the cross-validation results, thereby strengthening the justification for adaptive, low-latency control strategies. revision: yes

  2. Referee: Gesture selection via hierarchical clustering: the reduction to six movements is performed entirely on NINAPRO DB4 recordings; because this step is data-dependent, the manuscript must demonstrate that the same six movements remain representative under inter-subject variability or electrode shift to justify the scalability claim.

    Authors: While the clustering was conducted on the entire dataset, which includes multiple subjects, we will enhance the manuscript by adding a subject-wise stability analysis of the selected gestures using leave-one-subject-out clustering to address inter-subject variability. For electrode shift, as this is not directly simulatable from the provided data without additional assumptions, we will acknowledge this as a limitation and propose it for future investigation. revision: partial

  3. Referee: Feature refinement step: the iterative selection of the final five features combines visual inspection, mutual information, PCA, and decision-tree importance on the same dataset; this risks circularity and requires an explicit subject-independent hold-out evaluation to confirm the features are not overfit to the particular recordings.

    Authors: We concur that feature selection on the full dataset could introduce bias. The revised manuscript will incorporate a subject-independent evaluation protocol, where feature selection is performed on a training subset of subjects and validated on a held-out set of subjects, to confirm the generalizability of the five selected features. revision: yes

Circularity Check

0 steps flagged

No significant circularity; pipeline is data-driven on external public dataset

full rationale

The paper describes a standard four-stage empirical pipeline (preprocessing with low-pass filter and Hilbert transform, extraction of 26 features refined via mutual information/PCA/decision-tree importance to a 5-feature subset, Mahalanobis hierarchical clustering to select 6 gestures, 200 ms windowing, and ET/ANN classification) applied to the public NINAPRO DB4 dataset. No equations are presented that define a quantity in terms of itself or rename a fitted parameter as a first-principles prediction. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked as load-bearing justifications. Feature/gesture selection and model evaluation are performed on the same external recordings with reported independent training steps, but this constitutes ordinary empirical methodology rather than a derivation that reduces to its inputs by construction. The central claims rest on observable results from the dataset and standard libraries, not on self-referential reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 2 axioms · 0 invented entities

The central claims depend on several data-driven parameter choices and domain assumptions that are not derived from first principles or external benchmarks.

free parameters (4)
  • low-pass filter cutoff = 0.6 Hz
    Set at 0.6 Hz to reduce noise; chosen during preprocessing without stated theoretical justification.
  • analysis window length = 200 ms
    Selected as 200 ms based on stability and physiological plausibility after testing.
  • number of representative gestures = 6
    Result of hierarchical clustering chosen to balance diversity and computational cost.
  • number of retained features = 5
    Reduced from 26 to 5 via successive visual, information-theoretic, and importance-based filters.
axioms (2)
  • domain assumption Mahalanobis distance appropriately captures similarity among EMG feature vectors
    Invoked for hierarchical clustering without comparison to alternatives or validation of the covariance assumption.
  • domain assumption The NINAPRO DB4 recordings are sufficiently representative of real-world functional movements for prosthetic control
    Used to justify generalization from the dataset to practical low-latency applications.

pith-pipeline@v0.9.0 · 5830 in / 1914 out tokens · 71696 ms · 2026-05-21T06:57:44.264917+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

  1. [1]

    PREVALENCIA DE ALTERACIONES SENSITIVAS Y FACTORES,

    A. F. Díaz Charrupi, M. C. Mejía Vélez, and P. M. García López, “PREVALENCIA DE ALTERACIONES SENSITIVAS Y FACTORES,” Medellín, 201AD. Accessed: Dec. 09, 2024. [Online]. Available: https://repository.ces.edu.co/bitstream/han- dle/10946/2656/Prevalencia_alteraciones_sensitivas.pdf;jses- sionid=901C3AFAC4DB03EADC95E63954914E30?sequence=2

    work page 2024

  2. [2]

    A quan- titative taxonomy of human hand grasps,

    F. Stival, S. Michieletto, M. Cognolato, E. Pagello, H. Müller, and M. Atzori, “A quan- titative taxonomy of human hand grasps,” J Neuroeng Rehabil, vol. 16, no. 1, Feb. 2019, doi: 10.1186/s12984-019-0488-x

    work page doi:10.1186/s12984-019-0488-x 2019

  3. [3]

    NINAPRO

    University of Applied Sciences Western Switzerland - Valais, “NINAPRO.” [Online]. Available: https://ninapro.hevs.ch/

    work page

  4. [4]

    El Concepto de Distancia y su Aplicación en Estadística Multivariada

    F. M. Hernández Arellano Millward Brown México, “El Concepto de Distancia y su Aplicación en Estadística Multivariada.”

    work page

  5. [5]

    Humans need only 200 ms to generate posture - specific muscle activation patterns for successful vertical jumps in reaction to an auditory trigger,

    M. F. Bobbert and A. S. Koopman, “Humans need only 200 ms to generate posture - specific muscle activation patterns for successful vertical jumps in reaction to an auditory trigger,” Front Sports Act Living, vol. 5, 2023, doi: 10.3389/fspor.2023.1123335

    work page doi:10.3389/fspor.2023.1123335 2023