TL;DR
This paper proposes a novel method for estimating knee extension force using mechanomyography (MMG) signals, combining Gray Relational Analysis (GRA) for feature selection with an Improved Cuckoo Search optimized Support Vector Regression (ICS-SVR) for accurate prediction. This approach offers a more robust and accurate non-invasive method for monitoring muscle activity, crucial for advancing rehabilitation devices and prosthetics.
What Problem Does This Paper Address?
Accurate monitoring of muscle force is essential for effective rehabilitation and for developing advanced assistive devices like powered prostheses. Traditional methods for measuring muscle force can be invasive or cumbersome. While surface biosignals like electromyography (sEMG) and mechanomyography (MMG) offer non-invasive alternatives, sEMG signals can be susceptible to environmental noise and skin impedance variations. MMG signals, generated by muscle vibrations, are less affected by these issues and can provide rich information about muscle activity. However, MMG signals are characterized by weak signal strength, low frequency, and strong randomness, making it challenging to decode them for accurate muscle force estimation. Existing methods often rely on a limited set of features or struggle with computational complexity and model overfitting. This paper aims to address these challenges by developing a more accurate and efficient MMG-based muscle force estimation technique.
Methods
The proposed method integrates two key components:
- Gray Relational Analysis (GRA): To tackle the challenge of selecting effective MMG features, GRA is employed. This technique analyzes the correlation between various MMG features (time-domain, frequency-domain, time-frequency-domain, and nonlinear dynamics features) and the actual knee joint extension force. By calculating a "relational degree," GRA helps identify features that are highly correlated with muscle force, thereby reducing the dimensionality of the input data and preventing overfitting.
- Improved Cuckoo Search Support Vector Regression (ICS-SVR): Support Vector Regression (SVR) is chosen for its robust performance in handling nonlinear and complex data. However, SVR's accuracy is highly dependent on its critical parameters (C and σ). To optimize these parameters, the authors introduce an Improved Cuckoo Search (ICS) algorithm. ICS enhances the standard Cuckoo Search (CS) by incorporating:
- Chaotic Tent Initialization: This method uses chaotic mapping to generate a more diverse initial population of potential solutions, helping to avoid local optima.
- Adaptive Control Parameters: The algorithm dynamically adjusts key parameters (like step size and discovery probability) during the search process, improving convergence speed and accuracy.
The ICS algorithm is used to find the optimal SVR parameters, which are then used to train the SVR model for estimating knee extension force based on the selected MMG features.
Datasets
The study utilized MMG and force signals collected from five healthy male subjects. The data collection involved:
- Signal Acquisition: MMG signals were captured using three accelerometers placed on the rectus femoris (RF), vastus lateralis (VL), and vastus medialis (VM) muscles. Force signals were recorded using a force sensor. Both signals were sampled at 1 kHz.
- Experimental Protocol: Subjects performed maximal voluntary contractions (MVC) and forces ranging from 10% to 80% of MVC.
- Data Preprocessing: Signals were filtered, and stable muscle force data segments were selected.
- Feature Extraction: A sliding window approach was used to extract 25 features per channel (time-domain, frequency-domain, time-frequency-domain, and nonlinear dynamics), resulting in a total of 75 features.
- Data Splitting: 90% of the data was used for training and 10% for testing.
- Normalization: MMG features and muscle force were normalized using the Z-score method.
No new benchmark datasets were introduced; the study used data collected from human subjects.
Key Insights and Ideas
- MMG as a Viable Alternative to sEMG: The paper strongly advocates for MMG as a robust, non-invasive signal for muscle force estimation, highlighting its reduced susceptibility to skin impedance and environmental noise compared to sEMG.
- The Power of Feature Selection: The study demonstrates that not all extracted MMG features are equally informative. GRA effectively identifies a subset of features that significantly improve estimation accuracy, preventing the computational burden and overfitting associated with using all features.
- ICS Outperforms Standard Optimization: The proposed Improved Cuckoo Search (ICS) algorithm showed superior performance in optimizing SVR parameters compared to standard CS, PSO, and GWO on benchmark functions, leading to better overall estimation accuracy.
- Synergy of GRA and ICS-SVR: The combination of GRA for intelligent feature selection and ICS-SVR for accurate modeling proved to be a powerful approach for estimating knee extension force, outperforming other classical machine learning models.
Key Results
- Feature Combination D is Optimal: Feature combination D, which included MMG features with a relational degree greater than 0.85, yielded the best estimation results, achieving the smallest RMSE and MAPE, and the largest R value.
- Superior Estimation Accuracy: The proposed ICS-SVR model, using feature combination D, achieved an RMSE of 0.1319, MAPE of 0.0349, and R of 0.9966 for subject S1.
- Model Performance: Across five subjects, the ICS-SVR model consistently outperformed BPNN, ELM, and CS-SVR in terms of RMSE, MAPE, and R. For instance, the average RMSE for ICS-SVR was 0.2205 ± 0.0840, significantly lower than BPNN (0.4706 ± 0.1299) and CS-SVR (0.2673 ± 0.1061).
- Stability: The ICS-SVR model demonstrated less fluctuation in its estimation results compared to other models, indicating greater stability.
Limitations and Future Work
- Limitations:
- The study focused on isometric muscle contraction cases, and its applicability to dynamic force generation needs further investigation.
- The study had a limited sample size and involved subjects with a single physical health status, which might limit the generalizability of the findings.
- Future Work:
- Extend the research to dynamic force generation during isometric muscle contraction and force generation during joint movement.
- Refine key interaction techniques for specific rehabilitation and powered prostheses.
- Apply the proposed scheme to estimate movement parameters and muscle force using other time-series signals like sEMG and EEG.
- Investigate the application of MMG in other fields such as muscle function assessment, powered prostheses, and medical rehabilitation.
Summary updated