Identification and classification of upper limb motions using PCA

Introduction

An electromyogram (EMG) is the measurement of the action potentials produced from the neural stimulation of skeletal muscles. Like the electroencephalogram (EEG), an EMG signal can be measured invasively or non-invasively depending on the application [1, 4, 6, 7, 11]. However, surface-detected signals [i.e. surface electromyogram (SEMG)] are preferably used to obtain information about the muscle force relationship of superficial muscle activity [2, 13, 17, 19, 23]. The knowledge of SEMG as an input to powered prosthesis control systems has increased considerably over the last two decades. The simplest way utilizes the mapping of amplitudes of SEMG, measured over controlled muscle [8, 14, 16, 20, 21].

The focus of this study will be on upper arm limb prosthetic control. According to a survey conducted in 2007, there were approximately 1.7 million people living with limb loss in the United States. It is estimated that one out of every 200 people have an amputation. The number of amputees worldwide is approximately 4 million and there is an increase of 150,000–200,000 people every year. Upper limb amputations account for about 68% of amputees. Out of this, elderly amputees account for about 30%, 60% are between 21 and 60 years and the remaining 10% are under the age of 21 years. It is estimated that the number of amputees with no prosthesis increases annually by about 17,000. It can further be correlated that the risk of traumatic amputations increases steadily with age, reaching its highest level among people 85 years or older for both males and females. Males are at higher risk for trauma-related amputation compared to females.

Extracting independent information from SEMG signals might be useful in understanding the intrinsic nature of signals via investigating the muscle-force relationship. It has already been shown that using one-way analysis of variance (ANOVA) can improve SEMG-based muscle force estimation [5, 10, 15, 18, 22].

Hence, the present study investigates whether using PCA may yield further improvements of muscle force estimation from upper arm SEMG signals. The results of this present study can be widely applied in many SEMG classification studies, especially in biomedical and prosthetic applications.

Methods

Results and discussion

The pictorial representations of four upper arm activities being selected are as shown in Figure 1. As spectral analysis is an important tool for the evaluation of SEMG data before its interpretation, a LabVIEW-based program was prepared to plot the waveforms of raw SEMG for all four activities.

Figure 1:

Upper arm muscle activities.

(A) Elbow extension, (B) adduction, (C) abduction and (D) flexion elbow.

Figure 2 depicts recorded plots for Activity A1 – elbow extension, Activity A2 – abduction, Activity A3 – abduction and Activity A4 – elbow flexion. A total of eight parameters were evaluated for the analysis of the SEMG signals, namely RMS value (c1), variance (c2), simple square integral (c3), standard deviation (c4), power of signal (c5), integrated EMG (c6), mean absolute value (c7) and total sum of power spectrum (c8).

Figure 2:

SEMG plots for all preformed activities (A1–A4).

(A1) Elbow extension, (A2) adduction, (A3) abduction and (A4) flexion elbow.

To further extend the SEMG-based muscle force estimate study, the interpretation of recorded data was done using PCA at various upper arm activities. PCA is a data reduction technique used to reduce a large number of variables to a smaller set of underlying factors that summarize the essential information contained in the variables.

There are a number of assumptions underlying the applications of PCA; however, normality and linearity are the important ones. In this experimental work, both these assumptions have effective results or recorded data have passed both assumptions effectively. In a best way to correlate the information provided by the first two components, the four or five variables are required to be retained, as suggested by Cadima and Jolliffe [3].

The analysis was conducted using StatistiXL. Based on percentage of variance, both first and second components (PCA 1 and PCA 2) were taken into consideration. Within each component, variables have specific PCA plots. Loading in the pattern matrix represents the unique relationship between the factor and the variables, while the score represents the structure for a given component.

PCA was applied to the data set of all seven subjects and final plot results are shown in Figure 3. Loading component score coefficients are shown in Table 1, as four components PCA1, PCA2, PCA3 and PCA4 were selected. PCA correlations between variables and factors are shown in Table 2.

Figure 3:

PCA plots (before and after rotation) for all seven subjects.

The different variables are compared using component-loading scores tabulated in Table 1. From Table 1, for combination 1, biceps flexion elbow contributes 5.045, −0.482, 0.117 and 0.0034 to PCA1, PCA2, PCA3 and PCA4 compared to triceps flexion elbow values −1.051, −0.037, −0.098 and 0.070, respectively. Similarly, for combination 2, triceps adduction contributes 2.47, 0.287, 0.100 and −0.032 to PCA1, PCA2, PCA3 and PCA4 compared to biceps adduction values −2.10, −0.06, 0.06 and −0.02, respectively.

From the muscle force estimate results, one can also infer that a proportion of variation of 0.89 is explained by the best combination of two variables (c1 and c2). For a three-variable combination (c1, c2−c3), a proportion of variation of 0.85 is explained. The combination of four variables would be c1−c2, and c3−c4, explaining a variation of 0.92; a variation of 0.91 is best for five best combinations (c1−c5). For six different variable combinations (c1−c6), a proportion of 0.93 variation is explained and finally, proportions of variation of 0.89 and 0.91 are explained for the force estimated in subjects for the best seven and eight different variable combinations, respectively. Any value closer to 1 indicates that those correlation patterns are best for a particular combination.

To investigate a correlation among different parameters, the Bartlett test, showing significant difference, and the Kaiser-Meyer-Olkin (KMO) test, showing the measure of sampling adequacy, were performed. Firstly, data were subjected to the Bartlett test (p-value was found to be 0.00) to investigate the statistical significance. A p-value of less than 0.05 was found to indicate that analysis is significant (as the computed p-value is lower than the significance level alpha=0.05, one should reject the null hypothesis H0 and accept the alternative hypothesis Ha), i.e. the R-matrix is not an identity matrix; hence, a relationship exists between the variables and is appropriate on the given data set.

The significant differences of initial eigenvalues of the extracted PCA factors for the different extracted features were compared; and the results are summarized in Figure 4. These differences of linear increase can be understood because PCA is used to convert a set of highly correlated variables to a set of independent variables using linear transformations. Finally, when the dependent features for a regression are specified, the PCA technique is more efficient for dimension reduction.

Figure 4:

Eigenvalues for increasing number of parameters.

Lastly, the proportion of the total variation explained by the seven factors is 98.17%. The individual communalities (92.11 for PCA 1, 95.84 for PCA 2, 98.53 for PCA 3 and 99.26 for PCA 4) imply how well the model is working for the individual variables and the total communality gives an overall assessment of performance. The final conclusion on the factorial structure of the analyzed variable tells that for all extracted variables, c1−c8 has a better correlation for PCA1.

Conclusion and future work

The objective of this experimental work was to investigate the usefulness of SEMG in estimating muscle force relationships for upper limbs. The research demonstrated the use of a bipolar electrode combination for independent muscle activities (A1–A4). The findings based on PCA also provide evidence in improving the quality of muscle force estimation. It is concluded that SEMG is an important tool in predicting muscle force estimates; however, the results depend upon the way the data have been recorded and processed.

Further, it is always difficult to compare previous results due to non-uniformity in protocols and also due to different ways of analyzing, interpreting and presenting the results [9]. However, a classification accuracy of 98% is achieved for the ongoing research work and is very appreciable for classifying predefined upper arm activities compared to other related works [2, 3, 18, 22]. Future work will address the application of independent component analysis (ICA) and support vector machine (SVM) in more detail for assessing muscle force estimates. Finally, in view of previous and present results, PCA improvement in muscle force estimates is very comprehensive for upper limb applications.