Table 1.
Demographic profile of the 15 subjects who participated in EMG data collection.
Fig 1.
Block diagram from recording to prosthetic control (a) muscles selection (b) EMG sensors placement on muscles sites (c) collection of raw EMG data (d) bandpass and notch filters implementation to remove noises (e) time domain, frequency domain or time-frequency domain features extraction(f) classifier implementation using input features, classes or labels show specific movements (g) controller takes classes as input and provide specific signals to protype (h) prosthetic limbs are actuator that perform specific actions on the basis of input from controller.
Fig 2.
Complete detail of the control study design (a) subject: source of EMG data collection (b) data refining and initially preprocessing of the raw EMG data (c) segmentation: eight individual time domain windowing techniques used for the segmentation of refined preprocessed data (d) feature set:two feature sets involved in this study, 40 time domain features and 6 frequency domain features (e) classifiers: seven individual classifiers are implemented and compared (f) protocol:fifteen finger movements comprises of individual, dual and triple fingers movements are the final output of classifiers that may be provided to the control system for the prosthetic control.
Table 2.
Time-domain feature set: forty features with descriptions.
Table 3.
Frequency-domain feature set: six features with descriptions.
Table 4.
Time-domain windowing techniques for EMG data preprocessing: mathematical formulas and descriptions.
Table 5.
Classifiers used in recognition of EMG data: mathematical formulations and descriptions.
Fig 3.
Classification accuracy of time-domain windowing techniques combined with (a) time-domain features and (b) frequency-domain features for 15 subjects performing 15 movements.
Fig 4.
Classification error rates of time-domain windowing techniques combined with (a) time-domain features and (b) frequency-domain features for 15 subjects performing 15 movements.
Fig 5.
Standard deviations of classification accuracies using time-domain windowing techniques in preprocessing for 15 subjects performing 15 movements, evaluated with (a) time-domain feature sets and (b) frequency-domain feature sets.
Fig 6.
Variance in classification accuracies using time-domain windowing techniques in preprocessing for 15 subjects performing 15 movements, evaluated with (a) time-domain feature sets and (b) frequency-domain feature sets.
Fig 7.
Range of classification accuracies achieved by classifiers using time-domain windowing techniques in preprocessing for 15 subjects performing 15 movements, evaluated with (a) time-domain feature sets and (b) frequency-domain feature sets.
Fig 8.
Coefficient of variations of accuracies of classifiers with time domain windows technique in preprocessing for 15 subjects performing 15 movements (a) using time domain features set (b) using frequency domain features set.
Fig 9.
Maximum of classification accuracies achieved with time domain windows technique in preprocessing for 15 subjects performing 15 movements (a) using time domain features set (b) using frequency domain features set.
Fig 10.
Minimum of accuracies of classifiers with time domain windows technique in preprocessing for 15 subjects performing 15 movements (a) using time domain features set (b) using frequency domain features set.
Fig 11.
Median of accuracies of classifiers with time domain windows technique in preprocessing for 15 subjects performing 15 movements (a) using time domain features set (b) using frequency domain features set.
Table 6.
Classification accuracy comparison of 7 classifiers using 40 time-domain features with time-domain windowing techniques for preprocessing.
Table 7.
Classification accuracy comparison of machine learning models using 6 frequency-domain features extracted with time-domain windowing techniques in preprocessing.
Table 8.
Classification performance metrics of machine learning models using 6 frequency-domain features extracted with rectangular window preprocessing.
Table 9.
Comparative performance evaluation of classifiers using 40 time-domain features with rectangular window preprocessing technique.
Fig 12.
Confusion matrix illustrating the worst-case classification performance of the Logistic Regression (LR) classifier for Subject 5, using 6 frequency-domain features extracted with rectangular windowing.
Fig 13.
Confusion matrix illustrating the optimal classification performance of the BNN classifier for Subject 12, using 6 frequency- domain features extracted with rectangular windowing.
Fig 14.
Confusion matrix illustrating the worst-case classification performance of the LD (LD) classifier for Subject 5, using 40 time-domain features extracted with rectangular windowing.
Fig 15.
Confusion matrix illustrating the optimal classification performance of the Support Vector Machine (SVM) classifier for Subject 5, using 40 time-domain features extracted with rectangular windowing.