Exoskeleton: IMU/EMG Dataset Analysis and Motion Intention Modeling

Analysis and modeling of the public IMU + EMG dataset using ML models (CNN, LSTM, SVM) for lower-limb motion intention recognition.

Dataset (DOI) • Peer‑reviewed article

Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals
Silas Ruhrberg Estévez, Josée Mallah, Dominika Kazieczko, Chenyu Tang & Luigi G. Occhipinti
Scientific Reports, 2025 • Volume 15 • Article number: 38242

Overview

This repository contains:

• MATLAB scripts to preprocess the public dataset, train neural and classical models, and reproduce figures/analyses.
• Analysis utilities to generate confusion matrices, performance tables, and plots.

The dataset used in this work is publicly available via the University of Cambridge Apollo repository: https://doi.org/10.17863/CAM.113504. The full study is published in Nature Scientific Reports: https://www.nature.com/articles/s41598-025-22103-1.

Note: This repository focuses solely on analysis of the provided dataset; firmware and live data-collection utilities are out of scope.

Table of contents

• Quick start
• Repository structure
• Dataset Dataset: University of Cambridge Apollo Repository. DOI: https://doi.org/10.17863/CAM.113504 Article: Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals. Scientific Reports (2025), Volume 15, Article 38242. https://www.nature.com/articles/s41598-025-22103-1

BibTeX:

@article{RuhrbergEstevez2025,
    title = {Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals},
    volume = {15},
    ISSN = {2045-2322},
    url = {http://dx.doi.org/10.1038/s41598-025-22103-1},
    DOI = {10.1038/s41598-025-22103-1},
    number = {1},
    journal = {Scientific Reports},
    publisher = {Springer Science and Business Media LLC},
    author = {Ruhrberg Estévez, Silas and Mallah, Josée and Kazieczko, Dominika and Tang, Chenyu and Occhipinti, Luigi G.},
    year = {2025},
    month = oct
}

Maintainer: Silas Ruhrberg Estévez

• LSTM: run src/c_run_lstm_model.m
• SVM baseline: run src/d_run_svm_baseline.m
• Random baseline: run src/e_random_baseline.m

5. Reproduce analyses/plots

• Use scripts in analysis/ to generate confusion matrices, performance tables, time-series figures, and channel ablation analyses.

Repository structure

analysis/                  % Reproducible analysis and figure scripts (MATLAB)
	a_accuracy_cm.m
	b_confusion_matrix.m
	b_transfer_learning.m
	c_performance_table.m
	d_broken_channel_accuracy.m
	e_timeseries_classification.m
	f_processing_plot.m
	g_input_signals.m

src/                       % MATLAB modeling pipeline
	a_preprocess_data.m      % Point to raw data path; outputs processed features
	b_run_cnn_model.m        % Train/evaluate CNN
	c_run_lstm_model.m       % Train/evaluate LSTM
	d_run_svm_baseline.m     % SVM baseline
	e_random_baseline.m      % Random baseline sanity check
	exoskeleton_library.m    % Shared utility functions
	f_run_transfer_learning.m% Transfer learning experiments
	g_cnn_broken_channels.m  % CNN with channel ablation
	i_timeseries_classification.m % End-to-end time-series classification

README.md

Dataset

• DOI: https://doi.org/10.17863/CAM.113504
• Please review the usage terms on the dataset page. Place the downloaded files under a folder such as data/raw/.
• In src/a_preprocess_data.m, set the input path to your local raw dataset directory. The script produces processed outputs under data/processed/ (created if missing).

MATLAB environment and prerequisites

Recommended toolboxes (depending on which scripts you run):

• Signal Processing Toolbox
• Statistics and Machine Learning Toolbox
• Deep Learning Toolbox

General notes:

• The scripts are standard MATLAB .m files and should run on recent MATLAB releases. If you encounter version-specific issues, please open an issue with your MATLAB version and error message.
• Some analyses may take longer on CPU; a GPU (with Deep Learning Toolbox support) will speed up CNN/LSTM training.

Reproduce results (end-to-end)

1. Preprocess data

• Edit paths in src/a_preprocess_data.m to your dataset location.
• Run to generate train/validation/test splits and any derived features used downstream.

2. Train and evaluate models

• CNN: src/b_run_cnn_model.m produces training curves and evaluation metrics.
• LSTM: src/c_run_lstm_model.m for sequence modeling.
• SVM baseline: src/d_run_svm_baseline.m as a classical baseline.
• Random baseline: src/e_random_baseline.m sanity check.

3. Analyses and figures

• Confusion matrices and accuracy: analysis/a_accuracy_cm.m, analysis/b_confusion_matrix.m
• Transfer learning experiments: analysis/b_transfer_learning.m, src/f_run_transfer_learning.m
• Performance tables: analysis/c_performance_table.m
• Broken channel robustness: analysis/d_broken_channel_accuracy.m, src/g_cnn_broken_channels.m
• Time-series classification figures: analysis/e_timeseries_classification.m, src/i_timeseries_classification.m
• Processing and input signal plots: analysis/f_processing_plot.m, analysis/g_input_signals.m

Outputs are written to logical subfolders next to the scripts or into a results directory created by the scripts; check the printed paths in MATLAB’s Command Window.

Scripts guide

• src/a_preprocess_data.m

  • Contract: reads raw dataset, outputs standardized/segmented representations and splits.
  • Edge cases: missing channels, variable sampling rates, empty trials. The script includes checks and will warn if inputs are incomplete.

• src/b_run_cnn_model.m

  • Convolutional model for motion intention classification. Accepts the preprocessed dataset; logs metrics and confusion matrices.

• src/c_run_lstm_model.m

  • Sequence model for temporal dependencies; good for longer windows.

• src/d_run_svm_baseline.m

  • Classical baseline using summary features or flattened windows.

• src/f_run_transfer_learning.m

  • Reuses pretrained representations; see comments in the file for the specific source/target configuration.

• analysis/*

  • Standalone figure and table generators. They assume model outputs exist; re-run after training to update figures.

Results and figures

The analysis/ scripts reproduce key tables and figures, including:

• Accuracy and confusion matrices by class
• Transfer learning performance
• Robustness to broken/missing channels
• Example time-series segments, processing and input signal visualizations

Figures are saved to disk by each script; see the script comments for output locations.

Citation

If you use this repository, the dataset, or the results in your research, please cite:

• Dataset: University of Cambridge Apollo Repository. DOI: https://doi.org/10.17863/CAM.113504
• Article: Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals. Scientific Reports (2025), Volume 15, Article 38242. https://www.nature.com/articles/s41598-025-22103-1

Ruhrberg Estévez, S., Mallah, J., Kazieczko, D. et al. Deep learning for motion classification in ankle exoskeletons using surface EMG and IMU signals. Sci Rep 15, 38242 (2025). https://doi.org/10.1038/s41598-025-22103-1

License and acknowledgements

• License: MIT — see LICENSE in the repository root.
• Thanks to all contributors and participants involved in data collection and validation.

---

Maintainer: Silas Ruhrberg Estevez