Decoding Attempted Hand Movements in Stroke Patients Using Surface Electromyography

Mads Jochumsen, Imran Khan Niazi, Muhammad Zia ur Rehman, Imran Amjad, Muhammad Shafique, Syed Omer Gilani, Asim WarisSensors 2020 · DOI

TL;DR

This study demonstrates that surface electromyography (sEMG) can reliably decode nine different attempted hand and forearm movements in stroke patients, even with a simple electrode setup. Autoencoders (AE) achieved the highest classification accuracy, paving the way for more personalized and accessible EMG-controlled rehabilitation devices.

What Problem Does This Paper Address?

Stroke survivors often experience motor disabilities, necessitating effective rehabilitation strategies. While brain-computer interfaces (BCIs) using EEG can detect movement intentions, their accuracy decreases with more complex tasks. Surface electromyography (sEMG), which measures muscle electrical activity, offers a promising alternative for decoding movement intentions, especially in patients who have regained some muscle function. However, previous studies often used complex electrode setups and were limited to single-session evaluations, leaving questions about reliability and generalizability. This paper aims to address these gaps by investigating the feasibility of decoding multiple hand and forearm movements using a simple sEMG setup, evaluating different classification algorithms, and assessing the test-retest reliability of the decoding process.

Methods

The researchers employed a straightforward approach to capture muscle activity. Six surface EMG electrodes were placed on three forearm muscles: Extensor Carpi Radialis, Flexor Carpi Radialis, and Flexor Carpi Ulnaris. This setup was chosen for its simplicity and potential for low-cost embedded systems.

For data analysis, they explored three classification techniques:

  • Linear Discriminant Analysis (LDA): A classical, computationally efficient linear classifier.
  • Autoencoders (AE): A type of artificial neural network used for feature learning, which can then be used for classification.
  • Convolutional Neural Networks (CNN): A deep learning model that can directly process raw EMG data without explicit feature extraction.

The study involved two recording sessions on separate days to assess test-retest reliability. Participants attempted nine different motion classes: Hand Close, Hand Open, Wrist Extension, Wrist Flexion, Supination, Pronation, Lateral Grasp, Pinch Grasp, and Rest. For classification, they used a "within-session" approach (training and testing on data from the same day) and a "between-session" approach (training on one day and testing on the other).

Datasets

The study utilized data collected from 15 stroke patients. No pre-existing public datasets were used; instead, a new dataset was generated specifically for this research. The data consisted of surface EMG recordings from the forearm muscles during the performance of nine distinct hand and forearm movements.

Key Insights and Ideas

  • Simple sEMG is Powerful: Even with a minimal electrode setup (three channels on three muscles), it's possible to decode a variety of hand and forearm movements in stroke patients. This suggests that complex electrode arrays might not always be necessary for effective myoelectric control.
  • Autoencoders Shine: Autoencoders outperformed both LDA and CNNs in within-session classification accuracy, highlighting their effectiveness in learning relevant features from sEMG data for this task.
  • Reliability is Key: The study demonstrated good test-retest reliability for decoding across two different days, a crucial factor for practical applications like home-based rehabilitation.
  • Motor Impairment Doesn't Necessarily Hinder: Surprisingly, the level of motor impairment (as measured by the Fugl-Meyer score) did not correlate with classification accuracy. This implies that even individuals with significant motor deficits might still be able to benefit from sEMG-based control systems.
  • "Rest" is Easiest, "Lateral Grasp" is Toughest: The "Rest" condition was the most accurately classified, while "Lateral Grasp" proved to be the most challenging, likely due to its similarity to other grasp movements.

Key Results

  • High Classification Accuracy: On average, autoencoders achieved approximately 79% and 80% classification accuracy for the nine movement classes on day 1 and day 2, respectively.
  • Good Test-Retest Reliability: The intraclass correlation coefficient (ICC) for classification accuracy was 0.88, indicating strong agreement between the two recording sessions.
  • Classifier Performance: Within-session, AE achieved significantly higher accuracies than LDA and CNN. Between-session, the differences between classifiers were less pronounced.
  • Computational Efficiency: LDA was the fastest classifier for training and testing, while CNNs were the slowest but still achieved reasonable test times.
  • No Correlation with Impairment: The Spearman correlation between the upper limb Fugl-Meyer score and classification accuracy was low (0.24), indicating no significant association.

Limitations and Future Work

The authors acknowledge several limitations:

  • Sample Size: The study included only 15 stroke patients, and they suggest that including more patients with a wider range of impairments would strengthen the findings.
  • Generalizability: While the electrode placement was standardized, individual patient variability in muscle activation and potential compensatory strategies could influence results. Future work could explore more personalized electrode placements or advanced feature extraction techniques.
  • Limited Movement Repertoire: The study focused on nine specific hand and forearm movements. Expanding to a larger set of movements or more complex functional tasks would be beneficial.
  • Offline Analysis: The classification was performed offline. Future research should focus on validating these approaches in real-time control scenarios with exoskeletons or robotic devices, incorporating afferent feedback.

The researchers suggest investigating different feature types (e.g., spectral, wavelet) and electrode setups to further improve classification accuracy. They also highlight the need for online validation and the potential use of low-cost hardware platforms like Raspberry Pi for practical implementation.

Summary updated

BibTeX
@article{jochumsen2020decoding,
  author = {Mads Jochumsen and Imran Khan Niazi and Muhammad Zia ur Rehman and Imran Amjad and Muhammad Shafique and Syed Omer Gilani and Asim Waris},
  title = {{Decoding Attempted Hand Movements in Stroke Patients Using Surface Electromyography}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20236763},
}
Figure