The proposed method significantly mitigated the combined impact of such factors on the performance of the EMG-PR system. They also reported that MCFV can have a negative impact on overall prosthesis control performance if neglected [34]. To overcome this research gap, this study systematically investigated the co-existing influence of two dynamic factors (MoS and MCFV) on the performance of the myoelectric pattern recognition system when used to decode multiple classes of target limb movements.
Next, the sensitivity of EMG signal patterns to the combined effect of MoS-MCFV in the context of limb movement goal decoding was examined, and a new feature extraction method that attempted to mitigate the coexisting influence of both factors was proposed. dynamics. .
Data preprocessing
It is worth noting that due to the complexity of the experiments, a rest period of approximately 5 seconds was introduced between successive limb movement classes to avoid muscle or mental fatigue. One way to deal with this problem would therefore be to ensure the use of the same equipment and sampling parameters throughout the experimental sessions across subjects.
Feature extraction and classification
However, in a situation where the EMG recordings are made using different equipment and sampling parameters, the features of the extracted features may be somewhat inconsistent, resulting in performance degradation of the pattern recognition system. Here, the parameter T represents the observation time of the signal per window segment (typically 150 ms, as in Figure 3), and N indicates the number of samples per segment used in the calculation. In particular, this is consistent with Parseval's theorem, which states that the energy content of a signal can be obtained by using time- or frequency-based representation of the signal [25, 27].
Meanwhile, a version of the descriptors in Equations (2) and (5) was used to solve a different multi-factor problem in the field of myoelectric pattern recognition. Furthermore, to properly represent the underlined signal patterns in different scenarios, the signal amplitude along the first non-uniform dimension is obtained by calculating the difference between the maximum and minimum values [], using the second-order approximate derivative, and then normalized as shown in Equation (9) and (10). Common mode information in mobile scenarios per motion class was extracted using the mean logarithm of the signal and normalized using Equation (11) and (12), respectively.
Using the two strategies described in Section 2.6, we examined the extent to which the proposed invTDD method would minimize the dynamic coexistence impact of MoS and MCFV on PR-based systems, and further compared its performance with the commonly used feature extraction methods. described as follows. a) The time-dependent spectral domain (TD-PSD) feature set, recently proposed by Al-Timemy et al. Meanwhile, to reduce human intervention and also find an optimal data combination, a quintuple cross-validation technique [46] was used for splitting the extracted feature matrix in training and test sets. Therefore, we built a radial basis function-driven SVM classifier and compared its classification performance with that of the LDA classifiers.
Data Analysis and performance evaluation
Results and Discussion 3.1 Intra-Scenario analysis
Considering EMG recordings from the static scenario (S1)
Considering EMG recordings from the non-static scenarios (S2, S3, and S4) The training and testing procedures employed in scenario S1 was also applied to the
Furthermore, the significant different test between the proposed methods and the other feature extraction methods was performed using single factor analysis of variance (ANOVA) with a confidence level set at p<0.05. This result indicates that the invTDD method was able to mitigate the combined effect of MoS-MCFV by achieving a significant reduction in CE of medium at p<0.05) in scenario S2 compared to the previously proposed methods. Analyzing the plots in Figures 4c and 4d, comparable performance trends were observed when the feature matrices constructed from the data from scenarios S3 and S4 (ascent and descent of stair terrain) were used in the training and testing of the built PR-based motion-intent classifier .
Interestingly, training the PR-based system with data from moderate force levels seemed to generalize better than those from low or high muscle contraction force levels, as shown in Figure 4. Therefore, we decided to we trained PR-based limb movement goal classifiers with feature matrices extracted from mid-muscle contraction force level data and tested the classifiers using features pooled across muscle contraction force levels for each scenario to get the MCC values. Note that the data used for training are extracted from a specific muscle contraction force level while the test data include data from all muscle contraction force levels in scenario S1 (a), S2 (b), S3(c) and S4(d) .
Note: Train-L, Train-M, and Train-H represent training with a data set of low, medium, and high muscle contraction force levels, respectively.
Mathew correlation coefficient analysis
It should be noted that each classifier is trained with features from data of average contraction force level and tested with features extracted from data of all force levels, and the obtained results are shown in Figure 5a-5d. Also, in the other three scenarios (Figures 5b, 5c and 5d, using data from non-static scenarios for training and testing), the invTDD method performed significantly better in terms of minimizing the combined effect of MoS-MCFV than the previous methods especially in scenarios S2 (Figure 5b) and S4 (Figure 5d). Likewise, analysis based on F1_Score was also performed and the obtained results revealed that the coexisting influence of MoS-MCFV was significantly mitigated by the proposed invTDD method with an overall F1_Score score of 0.93% which is much higher than those of the methods others.
Therefore, we can conclude from these results that the invTDD method may mitigate the inherent coexisting impact of MoS-MCFV on PR-based systems compared to the other previous methods in the context of Intra-Scenario analysis. To provide more concise evaluation results, the RMS feature was excluded in the subsequent analyzes because it achieved the least performance among the compared methods. Note that the training data was drawn from the medium muscle contraction force level while the test data was obtained from all the contraction levels in scenario S1 (a), S2 (b), S3 (c) and S4 (d).
Inter-Scenario analysis
The average results obtained across subjects and limb movement classes are analyzed using boxplots, as shown in Figure 6. The training data were obtained from a particular contraction force level across scenarios and test data from the other two contraction levels across scenarios. Training with data from FL2 and testing with data from FL1+FL3 across all scenarios, H-ML: Training with data from FL3 and testing with data from FL1+FL2 across all scenarios.
Based on the bar graph analysis ( Figure 6 ), the PR-based limb movement intent classifier achieved the lowest CE for the proposed invTDD method compared to the other methods for all three cases. Meanwhile, invTDD recorded a standard error value of 0.91%, which is also much lower than those of the compared methods, indicating its relatively high level of inter-subject stability. Furthermore, when the PR-based movement intention classifier was trained using data from the average muscle contraction force level across all scenarios and tested using the correlation of data from the other two contraction levels of muscles across scenarios (M-LH), the invTDD method again achieved a much lower CE of 11.60%.
Also, the invTDD method substantially minimized the simultaneous impact of MoS-MCFV when the classifier was trained using data from a high contraction force level and tested using data from the other two force levels in scenarios (H-ML), as shown in figure 6. .
Impact of MoS-MCFV on Decoding Individual Limb Movement-Intent
Based on the within-scenario analysis results presented in Figure 4a (for static scenario: S1), it was found that training a PR-based movement intention classifier with a dataset from a certain level of muscle contraction force (for example: low) and testing the trained model by chaining the data set from the other two unseen force levels (for example: medium + high) would generally degrade the performance of the system. One possible reason why the proposed invTDD achieved significantly better results compared to other methods even in the presence of invisible force levels would be that. Consolidating the above observations are the results of MCC (Figure 5) and F1_Score (0.96% and 0.93%, respectively), which further showed a clear advantage of invTDD in terms of reducing the overall effect of MoS-MCFV on the performance of the PR-based system.
Nevertheless, with the introduction of the invTDD method, this effect was substantially minimized compared to the previously proposed methods, as shown in Figure 6. Remarkably, we also found that the performance of the invTDD and the other features generalized better when the classifier was trained with using data from average contraction force levels in different scenarios and tested with a combination of data from the low and high contraction force levels, which is consistent with the results presented in Sections 3.1.1 and 3.1.2, as well as those reported in previous studies. Extensive analyzes of the experimental results showed that the coexistence of MoS-MCFV would significantly impair the overall performance of the PR-based limb movement intention classification.
Therefore, it is believed that the results of this study would allow prosthetic developers to understand how to effectively improve the clinical robustness of state-of-the-art myoelectric control systems, especially against the coexisting effect of MoS and MCFV. In the future, we hope to further examine the accuracy and robustness of the proposed invTDD framework in terms of its ability to address other important issues affecting the functioning of the myoelectric pattern recognition system, particularly in the context of highly dexterous forearm/fine finger movements that typically require complex neuromuscular coordination . In addition, we hope to consider online processing of myoelectric data via a cloud platform [51] to study the performance of the proposed invTDD feature set with respect to the coexisting influences of MoS-MCFV.
Englehart, Electromyogram pattern recognition for control of motorized upper limb prostheses: State of the art and challenges for clinical use, Journal of Rehabilitation Research and Development. Phinyomark et al., Feature extraction of the first difference in EMG time series for EMG pattern recognition, Computer Methods and Programs in Biomedicine.
