Automated Channel Selection in High-Density sEMG for Improved Force Estimation

Gelareh Hajian, Ali Etemad, Evelyn Morin

TL;DR

This paper tackles the challenge of reducing the dimensionality of high-density surface electromyography (HD-sEMG) data for accurate muscle force estimation. They introduce two novel channel selection methods, PCA in the frequency domain (PCAfreq) and a Power-Correlation Ratio (PCR), which significantly improve force estimation accuracy while drastically reducing the number of EMG channels needed.

What Problem Does This Paper Address?

Estimating muscle force from surface electromyography (sEMG) signals is crucial for applications like prosthetics, rehabilitation, and human-machine interfaces. High-density sEMG (HD-sEMG) offers richer spatial information, potentially leading to more accurate force estimation. However, HD-sEMG generates a massive amount of data, leading to high computational costs, increased processing time, and potential redundancy. This makes real-time applications challenging. The core problem is how to effectively select a smaller, informative subset of HD-sEMG channels that maintains or even improves force estimation accuracy while significantly reducing data dimensionality.

Methods

The researchers developed and compared two novel channel selection techniques:

  • Principal Component Analysis in the Frequency Domain (PCAfreq): Instead of applying PCA to the raw time-domain sEMG signals, this method first computes the Fast Fourier Transform (FFT) of the sEMG signals for each channel. PCA is then applied to these frequency-domain representations (specifically, the magnitudes, discarding phase information). Channels with the largest contribution to the first principal component (PC) are selected.
  • Power-Correlation Ratio (PCR): This is a novel index designed to identify channels that contain strong force-related information while minimizing redundancy with other channels. It involves:
  • Calculating the power spectral density (PSD) for each channel, as PSD is known to correlate with generated force.
  • Normalizing these PSDs.
  • Calculating the Pearson correlation coefficient between each channel and all other channels to quantify redundancy.
  • Combining the normalized spectral power and the inverse of the average correlation to create the PCR index. Channels with the highest PCR values are selected.

These methods were compared against a baseline: PCA in the Time Domain (PCAtime), which is a more standard approach. The force estimation itself was performed using the Fast Orthogonal Search (FOS) algorithm, a non-linear system identification method.

Datasets

The study utilized a newly recorded dataset from 13 healthy participants performing isometric elbow flexion contractions. Data was collected using a specialized experimental setup (Queen's University Arm - QARM) that allowed for precise control of joint angles and force measurement. The sEMG data was acquired from three linear HD-sEMG arrays placed on the biceps brachii (long and short heads) and brachioradialis muscles. The dataset included variations in elbow joint angles (60°, 90°, 120°) and forearm postures (neutral, supinated), creating a comprehensive benchmark for evaluating channel selection methods.

Key Insights and Ideas

  • Frequency Domain PCA is Key: Applying PCA in the frequency domain (PCAfreq) proved more effective than in the time domain (PCAtime) for channel selection in this context. This suggests that the spectral characteristics of sEMG signals hold crucial information for force estimation that might be obscured in the raw time-domain data.
  • PCR Offers a Powerful Combination: The novel PCR index effectively balances the need for high spectral power (indicating force-related information) with low inter-channel correlation (reducing redundancy). This dual objective is critical for efficient dimensionality reduction without sacrificing accuracy.
  • First PC is Often Sufficient: The research strongly suggests that the first principal component (PC) captures the vast majority of the signal's variance (over 90% in some cases). Therefore, using only the first PC for channel selection is sufficient to improve force estimation accuracy, simplifying the process.
  • Subject-Specific Channel Selection: The study highlights that optimal channel subsets can vary significantly between individuals, reinforcing the need for subject-dependent channel selection techniques.

Key Results

  • Significant Accuracy Improvement: The best-performing methods, PCAfreq and PCR, achieved an approximate 30% improvement in force estimation accuracy (reduction in %NMSE) when using a subset of three channels per muscle, compared to using all seven differential channels.
  • Drastic Dimensionality Reduction: These methods reduced the number of channels by 57% (from seven to three channels per muscle).
  • PCR Outperforms PCAtime: For subsets of one and two channels, PCR performed better than PCAfreq and significantly better than PCAtime. While PCAfreq and PCR showed similar performance for three-channel subsets, PCR was slightly better overall.
  • PCAtime Deteriorates Performance: Notably, PCAtime, when used for channel selection, often increased the error in force estimation, especially with fewer channels.
  • No Significant Difference for Single Channels: Selecting only one channel, while achieving maximum dimensionality reduction, led to a significant decrease in accuracy.

Limitations and Future Work

  • Not Necessarily Global Optimum: The authors acknowledge that their selected channel subsets might not be the absolute global optimum. A brute-force search could potentially find even smaller subsets with higher accuracy, but this is computationally infeasible.
  • User-Dependency: The FOS algorithm and the proposed channel selection methods are user-dependent, requiring a calibration/learning phase.
  • Technique Dependency: The optimal channel subset might change if different force estimation techniques (e.g., neural networks) are used.
  • Artifact Susceptibility: HD-sEMG electrodes can be more prone to artifacts due to their larger contact area and potential for movement during muscle flexion.
  • Limited Scope: The study focused on isometric contractions and linear electrode arrays. Future work should explore dynamic contractions and different electrode configurations (e.g., 2D HD-sEMG).

The authors propose future work including exploring machine learning and deep learning for channel selection, investigating the temporal evolution of FOS models, and incorporating joint dynamics from wearable sensors for even more accurate force estimation and channel selection.

Summary updated

Abstract

Accurate and real-time estimation of force from surface electromyogram (EMG) signals enables a variety of applications. We developed and validated new approaches for selecting subsets of high-density (HD) EMG channels for improved and lower-dimensionality force estimation. First, a large dataset was recorded from a number of participants performing isometric contractions in different postures, while simultaneously recording HD-EMG channels and ground-truth force. The EMG signals were acquired from three linear surface electrode arrays, each with eight monopolar channels, and were placed on the long head and short head of the biceps brachii and brachioradialis. After data collection and pre-processing, fast orthogonal search (FOS) was employed for force estimation. To select a subset of channels, principal component analysis (PCA) in the frequency domain and a novel index called the power-correlation ratio (PCR), which maximizes the spectral power while minimizing similarity to other channels, were used. These approaches were compared to channel selection using time-domain PCA. We selected one, two, and three channels per muscle from the original seven differential channels to reduce the redundancy and correlation in the dataset. In the best case, we achieved an approximate improvement of 30% for force estimation while reducing the dimensionality by 57% for a subset of three channels.
BibTeX
@inproceedings{hajian2020automated,
  author = {Gelareh Hajian and Ali Etemad and Evelyn Morin},
  title = {{Automated Channel Selection in High-Density sEMG for Improved Force Estimation}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=Inj5oYvVe7},
}
Figure