Medical Engineering & Physics 36 (2014) 754–760
Contents lists available at ScienceDirect
Medical Engineering & Physics journal homepage: www.elsevier.com/locate/medengphy
Technical note
Several practical issues toward implementing myoelectric pattern recognition for stroke rehabilitation Yun Li a,b,c , Xiang Chen a,∗ , Xu Zhang a,b , Ping Zhou a,b,c a b c
Institute of Biomedical Engineering, University of Science and Technology of China, Hefei, Anhui, China Sensory Motor Performance Program, Rehabilitation Institute of Chicago, IL, USA Department of Physical Medicine & Rehabilitation, Northwestern University, Chicago, IL, USA
a r t i c l e
i n f o
Article history: Received 27 November 2012 Received in revised form 28 December 2013 Accepted 12 January 2014 Keywords: Surface electromyography Pattern recognition Myoelectric control Stroke rehabilitation
a b s t r a c t High density surface electromyogram (sEMG) recording and pattern recognition techniques have demonstrated that substantial motor control information can be extracted from neurologically impaired muscles. In this study, a series of pattern recognition parameters were investigated in classification of 20 different movements involving the affected limb of 12 chronic stroke subjects. The experimental results showed that classification performance could be improved with spatial filtering and be maintained with a limited number of electrodes. It was also found that appropriate adjustment of analysis window length, sampling rate, and high-pass cut-off frequency in sEMG conditioning and processing would be potentially useful in reducing computational cost and meanwhile ensuring classification performance. The quantitative analyses are useful for practical myoelectric control toward improved stroke rehabilitation. © 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Surface electromyography (sEMG) contains rich motor control information from which the user’s intention can be identified with appropriate signal processing methods [1]. sEMG signals recorded from residual muscles of amputee subjects have been used for prosthesis control for many years [1]. In recent years, myoelectric control was also used in robot-aided therapy for stroke rehabilitation. Compared with involuntary exercise, voluntary task implementation (triggered or regulated by sEMG signals from stroke subject’s paretic limb [2–4]) is a more useful intervention for enhanced therapeutic effect [2–10]. For conventional myoelectric control, sEMG amplitude from a pair of agonist–antagonist muscles is used to control a one-degree of freedom (DOF) movement [1]. In pattern-recognition-based myoelectric control, multiple sEMG electrodes or even high-density electrode arrays have been utilized to ensure recording of sufficient myoelectric control information regarding muscle co-activations [11–13,18,19,22–25]. Recently, myoelectric pattern recognition techniques have also been applied to individuals with neurologic injuries. In a previous study [17], we demonstrated that applying pattern recognition techniques to high density sEMG recordings
∗ Corresponding author at: Institute of Biomedical Engineering, University of Science and Technology of China, Hefei, Anhui, China. Tel.: +86 551 360 1175. E-mail address: [email protected] (X. Chen). http://dx.doi.org/10.1016/j.medengphy.2014.01.005 1350-4533/© 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
achieved high accuracies in classification of 20 different movements involving the affected limb of stroke subjects. In this study, several practical issues related to sEMG signal recording and processing were examined. First, the high density sEMG recording provides a convenient approach to examine the effect of various electrode configurations and number of sEMG channels on the classification performance. After then, the effects of sEMG analysis windows length, sampling rate and filter settings were further examined. Appropriate selection of these parameters may depend on a tradeoff between the classification accuracy and the requirement for implementing a practical system. For example, a longer analysis window would result in lower statistical variance of features and higher classification accuracy, but meanwhile it leads to a larger system delay [11]. Using a low sampling rate may reduce signal resolution and compromise classification accuracy, but dramatically reduce computational burden [16]. Although previous efforts regarding the above-mentioned issues have been made for sEMG classification in amputees or able-bodied subjects [16,19], these issues have not been examined in partially paralyzed muscles after stroke. It is necessary to quantitatively analyze the effect of sEMG signal recording and processing parameters on classification accuracy in individuals with a different nature of injury from amputation (i.e. stroke). The findings from such analyses will be used to determine the optimal parameter values that will facilitate the implementation of patternrecognition-based myoelectric control for stroke rehabilitation.
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
2. Methods
755
comprised of several neighboring monopolar electrodes with different weights.
2.1. Dataset description The dataset was recorded from 12 chronic stroke subjects (8 males, 4 females, 61 ± 10 years), as reported in [17], where the demographic and clinical assessment information for all subjects can be found. The study was approved by Institutional Review Board of Northwestern University (Chicago, IL, USA). Written informed consent was obtained from all subjects prior to the experiment. During the experiment, each subject was comfortably seated and asked to follow a video demonstration and perform 20 functional movements using the affected limb. The 20 movements included wrist flexion/extension, wrist supination/pronation, elbow flexion/extension, hand open/close, thumb flexion/extension, index finger flexion/extension, finger 3–5 flexion/extension, fine pinch, lateral pinch, tip pinch, gun posture and ulnar wrist down/up. Each subject was asked to complete 20 experimental trials, with 5 repetitions of about 3-s muscle contraction of a same movement in each trial. Between the consecutive trials, the subjects were allowed to take a sufficient rest to avoid muscular and mental fatigue. 89 electrodes were used to record high-density sEMG signals in a “monopolar” manner (there was indeed a subtraction of common feedback, namely the mean of all the recording channels, provided to each channel) from the affected arm and hand (Fig. 1) using a Refa EMG recording system (TMS International BV, Enschede, Netherlands). The details of the electrodes placement can be found in [17]. The sEMG signals were collected with a sampling rate of 2 kHz per channel with a band pass filter between 20 and 500 Hz. 2.2. Data processing and pattern recognition The collected sEMG signals were processed offline with Matlab (ver. 2012a, the Mathworks, Natick, MA). The onset and offset of each movement repetition were manually determined from the sEMG signal stream. Electromyogram feature extraction was performed for each 256 ms window, incremented in 64 ms segments throughout each movement repetition. The feature set consisted of four time domain (TD) statistics: number of zero crossings, waveform length, number of slope sign changes, and mean absolute value. These measures were calculated for each sEMG signal from the 89 channels [11–15]. Finally, a linear discriminant classifier (LDC) [20] based on the maximum a-posteriori probability rule and Bayesian principle was chosen to get the classification results in a user-specific manner. For each subject, the data from first 4 repetitions were assigned as the training dataset and the data from the last repetition (i.e. the fifth) were referred to as the testing dataset. The classification accuracy was defined as the percentage of correct decisions to total number of testing samples for each subject. 2.3. Practical issues With the attempt to build a robust myoelectric control system toward improved stroke rehabilitation, the following practical issues were examined sequentially in this study. 2.3.1. Spatial filters The effect of different electrode configurations on classification performance was examined by implementing 5 spatial filters including single differential filter in transverse direction (SDT), single differential filter in longitudinal direction (SDL), double differential filter in transverse direction (DDT), double differential filter in longitudinal direction (DDL), and Laplace filter (LapD). Each spatial filter corresponded to an electrode configuration (Table 1)
2.3.2. Channel reduction To select a clinically applicable small number of sEMG channels, an electrode selection algorithm based on the sequential forward searching (SFS) method [14] was used, which iteratively added the most informative channels till the classification performance was regarded to be acceptable or comparable with that of all the 89 channels. 2.3.3. Window length adjustment Each pattern decision was produced from an analysis window within the continuous sEMG signal stream. In this study the analysis window length was originally set at 256 ms [17]. It was then adjusted from 64 to 512 ms with a 64 ms increment, and the dependence of the classification performance on the analysis window length was examined. 2.3.4. Re-sampling The sEMG signals were acquired with a sampling rate of 2 kHz per channel. In this study, the sEMG signals were down-sampled from 2000 to 200 Hz in 50 Hz decrements. The dependence of the classification performance on the reduced sampling rate was examined. In signal down-sampling process, Matlab used a proper anti-aliasing (low-pass) finite-impulse-response (FIR) filter with a cut-off frequency that was half of the corresponding downsampling rate to the original sEMG data. 2.3.5. Re-filtering The original sEMG signals were filtered by a system band-pass filter at 20–500 Hz to remove motion artifacts and high frequency noise. To examine the dependence of cut-off frequencies of a highpass filter on the classification performance, the original sEMG data were digitally re-filtered by a high-pass filter (6th order digital Butterworth filter) with a gradually increasing cut-off frequency from 20 to 120 Hz with a 5-Hz increment, and then the classification accuracy was re-calculated for each new cut-off frequency. 2.4. Statistical analysis The one-way repeated-measure analysis of variance (ANOVA) was applied on the classification accuracy. The level of statistical significance was set to p < 0.05 for all analyses. When necessary, post hoc pairwise multiple comparisons with Bonferroni correction were used. All statistical analyses were completed using SPSS software (ver. 16.0, SPSS Inc. Chicago, IL). 3. Results 3.1. Spatial filtering results The classification accuracies of different spatially filtered sEMG signals were averaged over 12 stroke subjects (Fig. 2). It was observed that the mean accuracy of 94% was achieved by the unfiltered monopolar (MN) sEMG signals while the accuracy slightly improved with different spatial filters. However, the ANOVA showed that there was no significant difference in classification accuracy among spatial filters (main effect: p = 0.074). Relying on the high-density sEMG recordings, the unfiltered MN-sEMG signals still yielded very high-classification accuracy (close to 100%) for some subjects, making it difficult to examine the performance improvement due to the spatial filtering. With this concern, we divided 12 stroke subjects into two groups: one was termed as group A including 7 subjects with unfiltered MN accuracy higher
756
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
Fig. 1. Schematic description of the electrode placement for 89-channel sEMG signal recordings.
than 95% and the other one was named as group B consisting of 5 subjects with unfiltered MN accuracy lower than 95%. For group B, the ANOVA showed a significant effect for the spatial filter factor on performance improvement (p = 0.031). The pairwise comparisons further revealed that both single differential filters (SDT and SDL) yielded the highest accuracies, which were significantly higher than the unfiltered MN accuracy (p = 0.016 for SDT vs. MN and p = 0.047 for SDL vs. MN). However, no significant difference in classification accuracy for group A could be observed among MN and all spatial filters (main effect: p = 0.55). Although there was no significant difference between SDT and SDL in classification accuracy for both groups (p > 0.05), considering SDL achieved the highest mean accuracy for all 12 stroke subjects, we adopted SDL as the optimized spatial filtering method.
3.2. Channel reduction results Fig. 3 shows the classification results averaged over 12 stroke subjects for different spatial filters when the number of SFS-selected channels increased from 1 to 20. Take selected 10 electrode channels as an example, there were roughly 3 channels selected from the upper arm, 5–6 channels selected from the forearm and 1–2 channels selected from the hand. This electrode placement result reflected the sEMG signals collected from different muscles contributing to the classification differently. The accuracy increased dramatically at the beginning of the curve and it almost climbed beyond 94% when 8 spatially filtered channels were selected. The pairwise comparisons in the ANOVA showed that for the spatial filter SDL, the classification performance with the channel number 20 (at which the classification accuracy was
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
757
Table 1 Applied spatial filters. Name
Description
Electrode configuration
SDT
Single differential filter in transverse direction
SDL
Single differential filter in longitudinal direction
DDT
Double differential filter in transverse direction
DDL
Double differential filter in longitudinal direction
LapD
Laplace filter
regarded to be quite high and comparable with 89 channels) was significantly superior to those with channel number from 1 to 7, but appeared insignificant difference in classification accuracy with those more than 7 (p = 0.042, p = 0.165 and p = 0.514 for channel number 7, 8 and 9 vs. 20, respectively). Thus, for a better signal processing capability, it was practical to use 8 channels. From the detailed sub-plot located in the center of Fig. 3, it was observed that the spatial filter SDL applied to a reduced number of channels yielded the best classification performance. This again demonstrated that SDL was the optimal spatial filtering method.
Therefore, the SDL and corresponding 8 spatially filtered channels for each subject were used in the following data analysis. 3.3. Window length adjustment results Fig. 4(a) shows the influence of adjusting window length on the classification accuracy with 8 selected spatially filtered SDLsEMG channels. When the window length increased from 256 ms to 512 ms, the classification accuracy was only slightly improved to approximately 2.1% but without significant difference (p = 0.24 for 256 ms vs. 512 ms). When the window length was reduced from 256 ms to 64 ms, the classification accuracy began to drop rapidly
100 100
98 90
MN SDT SDL DDT DDL LapD
94 92 90 88 86 GroupB GroupA All
84 82 80
80
100 Classification Accuracy (%)
Classification Accuracy (%)
Classification Accuracy (%)
96
70
60
50
90
40
MN
SDT
SDL DDT Spatial Filters
DDL
8
LapD
Fig. 2. Effect of spatial filtering on the classification accuracy for group B subjects, group A subjects and all the subjects using 89-channel high-density sEMG recordings.
1
2
3
4
5
9
6
7
10 11 12 Channel Number 8 9 10 11 12 13 14 15 16 17 18 19 20 Channel Number
Fig. 3. Effect of channel number on the classification accuracy for monopolar electrode sEMG recordings and 5 kinds of spatially filtered sEMG recordings.
758
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
(a)
and was found to be significantly lower (p = 0.047, p = 0.002 and p < 0.001 for 192 ms, 128 ms and 64 ms vs. 512 ms, respectively). It was revealed that the window length of 256 ms was a practical choice.
100
Classification Accuracy (%)
95 90
3.4. Re-sampling results 85
Fig. 4(b) further shows the effect of choosing different sampling rates on the classification results. It can be seen that there was a pronounced decrease in the average classification accuracy when the sampling rate went down from 2 kHz to 200 Hz. When comparing the accuracy achieved at the original 2 kHz with those at decreased sampling rates, the pairwise comparisons in ANOVA showed that significant differences in classification accuracy emerged when the sampling rate was set below 600 Hz (p = 0.065, p = 0.046 and p = 0.023 for 600 Hz, 550 Hz and 500 Hz vs. 2 kHz, respectively). Furthermore, the accuracy decreased more rapidly when the sampling rate was lower than 600 Hz (p < 0.05 for any sampling rate below 500 Hz vs. 2 kHz). However, choosing 600 Hz sampling rate led to a mean accuracy following below 90% (88.45%). It was decided that a 1 kHz sampling rate reached the practical point of diminishing returns as the mean accuracy was 90.88%, which only dropped 2.4% compared with that at 2 kHz sampling rate (p = 0.234).
80 75 70 65 60
(b)
64
128
192
256 320 384 Window Length (ms)
448
512
100
Classification Accuracy (%)
95 90 85
3.5. Re-filtering results
80 75 70 65 60 2000
1800
1600
1400
1200 1000 800 Sampling Rate (Hz)
600
400
200
95
Classification Accuracy (%)
(c)
90
Based on the above optimized settings, we also examined the choice of low cut-off frequency for a band-pass filter applied to the sEMG signals. Fig. 4(c) reports the classification results with a large range of low cut-off frequencies. Classification accuracy exhibited a decrease when a higher cut-off frequency was used. The pairwise comparisons following the ANOVA showed that compared with the accuracy achieved at the original 20 Hz cut-off frequency, the classification accuracy dropped down significantly (p = 0.065, p = 0.031 and p = 0.016 for 65 Hz, 70 Hz and 75 Hz vs. 20 Hz, respectively). This suggested that a cut-off frequency no higher than 65 Hz would be appropriate to avoid much compromise in classification performance. Surprisingly, it was also found that using a 60-Hz cut-off frequency had the average accuracy take a local maximum value over 90.37% and only decreased it by 0.3%, as compared to the original 20 Hz cut-off frequency (p = 0.076). Consequently, it was indicated that the cut-off frequency of 60 Hz was preferred. 4. Discussion
85
80 20
40
60 80 High-pass Cutoff Frequency (Hz)
100
120
Fig. 4. (a) Effect of window length on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used; (b) effect of sEMG sampling rate on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used. The window length was 256 ms; (c) effect of sEMG high-pass cut-off frequency on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used. The window length was 256 ms and sampling rate was 1 kHz.
In this study, some technical issues have been examined for the sEMG classification of stroke survivors. Considering the unique sEMG signals obtained from people with neurologic injuries, the results found in this study were significant. sEMG can be used to interpret the user’s intention of volitional muscle activities, and control externally powered devices for rehabilitation. We demonstrated the feasibility of identifying multiple intended movements involving the affected limb of stroke subjects by applying pattern recognition techniques to high-density sEMG recordings [17]. The identified movement intentions can serve as rich control inputs to assistive devices in sEMG-driven robot-aided therapy and facilitate motor recovery after stroke [7–10]. In a previous study [17], the subject characteristics were reported as a table, where the subjects’ clinical information was presented. Some subjects with lower clinical functionality assessment scores tended to yield lower classification accuracy, but this was not always the case. The relationship between the classification performance and subjects’ clinical assessment scores remains unclear, presumably due to the fact that the mechanisms underlying stroke are complex and that the clinical assessment is relatively subjective. This study examined
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
several practical issues for development of a myoelectric pattern recognition control system for stroke rehabilitation. The spatial filtering tended to increase the classification accuracy. This may be because that the spatial filters are able to remove common-mode signal components in monopolar sEMG signals with lower spatial resolution. This outcome for stroke subjects is consistent with the results of previous studies for amputees [14,15]. The results showed that the SDL filter produced the highest classification accuracy, although statistical analysis exhibited no significant difference in classification performance among all examined spatial filters. Using the single differential spatial filter is a practical way for improving classification accuracy in clinical practice. Consistent with the previous studies [12,14,15], we demonstrated that 20 or less sEMG electrodes can maintain high level of classification accuracy. When 8 localized channels were used, the highest classification accuracy was achieved by SDL filter, and there was not a significant compromise when it was compared with the results for high-density sEMG signals. Considering that the single differential sEMG recording is indeed more clinically feasible, using a relatively small number (i.e., 8 in this study) of single differential sEMG sensors proves to be sufficient for implementing myoelectric control system in practice. It is noteworthy that the above channel selection procedure was conducted in a user-specific condition. The between-subject variability of selected electrode positions was also observed from a previous study [14]. Furthermore, it is possible to maintain high levels of classification accuracy by selecting a limited number of electrodes. In this regard, high-density sEMG provides important guidance for optimizing electrode number and location to detect the intended movements. After the number of electrodes has been selected with the appropriate configuration, data window length, sampling rate, and high-pass cut-off frequency for sEMG signal filtering are three factors that have a significant impact on myoelectric control and practical implementation. It is acknowledged that a larger window length or a larger sampling rate may carry more useful information in each analysis window that would help producing more accurate decisions [11]. In these cases, there would be demand for more computational power and memory beyond the limitation of currently available microprocessor-based controllers. Especially, a larger window length would also introduce a longer signal processing delay [1,14]. The low-frequency components of sEMG recordings do not contribute much to classification performance. Thus, a higher high-pass cut-off frequency would be chosen to enhance the system stability and reduce interference and ECG artifacts. The results suggest that a combination of a window length of 256 ms, a sampling rate of 1 kHz and a high-pass cut-off frequency of 60 Hz seem to be a proper setting for conditioning and preprocessing sEMG signals for pattern classification of movement intentions. The combined adjustments of both the sampling rate decreased from original 2 kHz to 1 kHz and the high-pass cut-off frequency increased from 20 Hz to 60 Hz only led to a slight decrease of the classification accuracy (below 3.7%), where no significant difference could be found. However, the window length used in this study might not be lower than 256 ms, otherwise the classification accuracy would be compromised. It is interesting to find from Fig. 4(c) that with the 60-Hz cut-off frequency, the overall classification accuracy even reaches a local maximum, which is not much less than the accuracy achieved with the original 20-Hz cut-off frequency. The explanation for this is that using the 60-Hz high-pass cut-off frequency does eliminate most low-frequency motion artifacts (
Contents lists available at ScienceDirect
Medical Engineering & Physics journal homepage: www.elsevier.com/locate/medengphy
Technical note
Several practical issues toward implementing myoelectric pattern recognition for stroke rehabilitation Yun Li a,b,c , Xiang Chen a,∗ , Xu Zhang a,b , Ping Zhou a,b,c a b c
Institute of Biomedical Engineering, University of Science and Technology of China, Hefei, Anhui, China Sensory Motor Performance Program, Rehabilitation Institute of Chicago, IL, USA Department of Physical Medicine & Rehabilitation, Northwestern University, Chicago, IL, USA
a r t i c l e
i n f o
Article history: Received 27 November 2012 Received in revised form 28 December 2013 Accepted 12 January 2014 Keywords: Surface electromyography Pattern recognition Myoelectric control Stroke rehabilitation
a b s t r a c t High density surface electromyogram (sEMG) recording and pattern recognition techniques have demonstrated that substantial motor control information can be extracted from neurologically impaired muscles. In this study, a series of pattern recognition parameters were investigated in classification of 20 different movements involving the affected limb of 12 chronic stroke subjects. The experimental results showed that classification performance could be improved with spatial filtering and be maintained with a limited number of electrodes. It was also found that appropriate adjustment of analysis window length, sampling rate, and high-pass cut-off frequency in sEMG conditioning and processing would be potentially useful in reducing computational cost and meanwhile ensuring classification performance. The quantitative analyses are useful for practical myoelectric control toward improved stroke rehabilitation. © 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
1. Introduction Surface electromyography (sEMG) contains rich motor control information from which the user’s intention can be identified with appropriate signal processing methods [1]. sEMG signals recorded from residual muscles of amputee subjects have been used for prosthesis control for many years [1]. In recent years, myoelectric control was also used in robot-aided therapy for stroke rehabilitation. Compared with involuntary exercise, voluntary task implementation (triggered or regulated by sEMG signals from stroke subject’s paretic limb [2–4]) is a more useful intervention for enhanced therapeutic effect [2–10]. For conventional myoelectric control, sEMG amplitude from a pair of agonist–antagonist muscles is used to control a one-degree of freedom (DOF) movement [1]. In pattern-recognition-based myoelectric control, multiple sEMG electrodes or even high-density electrode arrays have been utilized to ensure recording of sufficient myoelectric control information regarding muscle co-activations [11–13,18,19,22–25]. Recently, myoelectric pattern recognition techniques have also been applied to individuals with neurologic injuries. In a previous study [17], we demonstrated that applying pattern recognition techniques to high density sEMG recordings
∗ Corresponding author at: Institute of Biomedical Engineering, University of Science and Technology of China, Hefei, Anhui, China. Tel.: +86 551 360 1175. E-mail address: [email protected] (X. Chen). http://dx.doi.org/10.1016/j.medengphy.2014.01.005 1350-4533/© 2014 IPEM. Published by Elsevier Ltd. All rights reserved.
achieved high accuracies in classification of 20 different movements involving the affected limb of stroke subjects. In this study, several practical issues related to sEMG signal recording and processing were examined. First, the high density sEMG recording provides a convenient approach to examine the effect of various electrode configurations and number of sEMG channels on the classification performance. After then, the effects of sEMG analysis windows length, sampling rate and filter settings were further examined. Appropriate selection of these parameters may depend on a tradeoff between the classification accuracy and the requirement for implementing a practical system. For example, a longer analysis window would result in lower statistical variance of features and higher classification accuracy, but meanwhile it leads to a larger system delay [11]. Using a low sampling rate may reduce signal resolution and compromise classification accuracy, but dramatically reduce computational burden [16]. Although previous efforts regarding the above-mentioned issues have been made for sEMG classification in amputees or able-bodied subjects [16,19], these issues have not been examined in partially paralyzed muscles after stroke. It is necessary to quantitatively analyze the effect of sEMG signal recording and processing parameters on classification accuracy in individuals with a different nature of injury from amputation (i.e. stroke). The findings from such analyses will be used to determine the optimal parameter values that will facilitate the implementation of patternrecognition-based myoelectric control for stroke rehabilitation.
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
2. Methods
755
comprised of several neighboring monopolar electrodes with different weights.
2.1. Dataset description The dataset was recorded from 12 chronic stroke subjects (8 males, 4 females, 61 ± 10 years), as reported in [17], where the demographic and clinical assessment information for all subjects can be found. The study was approved by Institutional Review Board of Northwestern University (Chicago, IL, USA). Written informed consent was obtained from all subjects prior to the experiment. During the experiment, each subject was comfortably seated and asked to follow a video demonstration and perform 20 functional movements using the affected limb. The 20 movements included wrist flexion/extension, wrist supination/pronation, elbow flexion/extension, hand open/close, thumb flexion/extension, index finger flexion/extension, finger 3–5 flexion/extension, fine pinch, lateral pinch, tip pinch, gun posture and ulnar wrist down/up. Each subject was asked to complete 20 experimental trials, with 5 repetitions of about 3-s muscle contraction of a same movement in each trial. Between the consecutive trials, the subjects were allowed to take a sufficient rest to avoid muscular and mental fatigue. 89 electrodes were used to record high-density sEMG signals in a “monopolar” manner (there was indeed a subtraction of common feedback, namely the mean of all the recording channels, provided to each channel) from the affected arm and hand (Fig. 1) using a Refa EMG recording system (TMS International BV, Enschede, Netherlands). The details of the electrodes placement can be found in [17]. The sEMG signals were collected with a sampling rate of 2 kHz per channel with a band pass filter between 20 and 500 Hz. 2.2. Data processing and pattern recognition The collected sEMG signals were processed offline with Matlab (ver. 2012a, the Mathworks, Natick, MA). The onset and offset of each movement repetition were manually determined from the sEMG signal stream. Electromyogram feature extraction was performed for each 256 ms window, incremented in 64 ms segments throughout each movement repetition. The feature set consisted of four time domain (TD) statistics: number of zero crossings, waveform length, number of slope sign changes, and mean absolute value. These measures were calculated for each sEMG signal from the 89 channels [11–15]. Finally, a linear discriminant classifier (LDC) [20] based on the maximum a-posteriori probability rule and Bayesian principle was chosen to get the classification results in a user-specific manner. For each subject, the data from first 4 repetitions were assigned as the training dataset and the data from the last repetition (i.e. the fifth) were referred to as the testing dataset. The classification accuracy was defined as the percentage of correct decisions to total number of testing samples for each subject. 2.3. Practical issues With the attempt to build a robust myoelectric control system toward improved stroke rehabilitation, the following practical issues were examined sequentially in this study. 2.3.1. Spatial filters The effect of different electrode configurations on classification performance was examined by implementing 5 spatial filters including single differential filter in transverse direction (SDT), single differential filter in longitudinal direction (SDL), double differential filter in transverse direction (DDT), double differential filter in longitudinal direction (DDL), and Laplace filter (LapD). Each spatial filter corresponded to an electrode configuration (Table 1)
2.3.2. Channel reduction To select a clinically applicable small number of sEMG channels, an electrode selection algorithm based on the sequential forward searching (SFS) method [14] was used, which iteratively added the most informative channels till the classification performance was regarded to be acceptable or comparable with that of all the 89 channels. 2.3.3. Window length adjustment Each pattern decision was produced from an analysis window within the continuous sEMG signal stream. In this study the analysis window length was originally set at 256 ms [17]. It was then adjusted from 64 to 512 ms with a 64 ms increment, and the dependence of the classification performance on the analysis window length was examined. 2.3.4. Re-sampling The sEMG signals were acquired with a sampling rate of 2 kHz per channel. In this study, the sEMG signals were down-sampled from 2000 to 200 Hz in 50 Hz decrements. The dependence of the classification performance on the reduced sampling rate was examined. In signal down-sampling process, Matlab used a proper anti-aliasing (low-pass) finite-impulse-response (FIR) filter with a cut-off frequency that was half of the corresponding downsampling rate to the original sEMG data. 2.3.5. Re-filtering The original sEMG signals were filtered by a system band-pass filter at 20–500 Hz to remove motion artifacts and high frequency noise. To examine the dependence of cut-off frequencies of a highpass filter on the classification performance, the original sEMG data were digitally re-filtered by a high-pass filter (6th order digital Butterworth filter) with a gradually increasing cut-off frequency from 20 to 120 Hz with a 5-Hz increment, and then the classification accuracy was re-calculated for each new cut-off frequency. 2.4. Statistical analysis The one-way repeated-measure analysis of variance (ANOVA) was applied on the classification accuracy. The level of statistical significance was set to p < 0.05 for all analyses. When necessary, post hoc pairwise multiple comparisons with Bonferroni correction were used. All statistical analyses were completed using SPSS software (ver. 16.0, SPSS Inc. Chicago, IL). 3. Results 3.1. Spatial filtering results The classification accuracies of different spatially filtered sEMG signals were averaged over 12 stroke subjects (Fig. 2). It was observed that the mean accuracy of 94% was achieved by the unfiltered monopolar (MN) sEMG signals while the accuracy slightly improved with different spatial filters. However, the ANOVA showed that there was no significant difference in classification accuracy among spatial filters (main effect: p = 0.074). Relying on the high-density sEMG recordings, the unfiltered MN-sEMG signals still yielded very high-classification accuracy (close to 100%) for some subjects, making it difficult to examine the performance improvement due to the spatial filtering. With this concern, we divided 12 stroke subjects into two groups: one was termed as group A including 7 subjects with unfiltered MN accuracy higher
756
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
Fig. 1. Schematic description of the electrode placement for 89-channel sEMG signal recordings.
than 95% and the other one was named as group B consisting of 5 subjects with unfiltered MN accuracy lower than 95%. For group B, the ANOVA showed a significant effect for the spatial filter factor on performance improvement (p = 0.031). The pairwise comparisons further revealed that both single differential filters (SDT and SDL) yielded the highest accuracies, which were significantly higher than the unfiltered MN accuracy (p = 0.016 for SDT vs. MN and p = 0.047 for SDL vs. MN). However, no significant difference in classification accuracy for group A could be observed among MN and all spatial filters (main effect: p = 0.55). Although there was no significant difference between SDT and SDL in classification accuracy for both groups (p > 0.05), considering SDL achieved the highest mean accuracy for all 12 stroke subjects, we adopted SDL as the optimized spatial filtering method.
3.2. Channel reduction results Fig. 3 shows the classification results averaged over 12 stroke subjects for different spatial filters when the number of SFS-selected channels increased from 1 to 20. Take selected 10 electrode channels as an example, there were roughly 3 channels selected from the upper arm, 5–6 channels selected from the forearm and 1–2 channels selected from the hand. This electrode placement result reflected the sEMG signals collected from different muscles contributing to the classification differently. The accuracy increased dramatically at the beginning of the curve and it almost climbed beyond 94% when 8 spatially filtered channels were selected. The pairwise comparisons in the ANOVA showed that for the spatial filter SDL, the classification performance with the channel number 20 (at which the classification accuracy was
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
757
Table 1 Applied spatial filters. Name
Description
Electrode configuration
SDT
Single differential filter in transverse direction
SDL
Single differential filter in longitudinal direction
DDT
Double differential filter in transverse direction
DDL
Double differential filter in longitudinal direction
LapD
Laplace filter
regarded to be quite high and comparable with 89 channels) was significantly superior to those with channel number from 1 to 7, but appeared insignificant difference in classification accuracy with those more than 7 (p = 0.042, p = 0.165 and p = 0.514 for channel number 7, 8 and 9 vs. 20, respectively). Thus, for a better signal processing capability, it was practical to use 8 channels. From the detailed sub-plot located in the center of Fig. 3, it was observed that the spatial filter SDL applied to a reduced number of channels yielded the best classification performance. This again demonstrated that SDL was the optimal spatial filtering method.
Therefore, the SDL and corresponding 8 spatially filtered channels for each subject were used in the following data analysis. 3.3. Window length adjustment results Fig. 4(a) shows the influence of adjusting window length on the classification accuracy with 8 selected spatially filtered SDLsEMG channels. When the window length increased from 256 ms to 512 ms, the classification accuracy was only slightly improved to approximately 2.1% but without significant difference (p = 0.24 for 256 ms vs. 512 ms). When the window length was reduced from 256 ms to 64 ms, the classification accuracy began to drop rapidly
100 100
98 90
MN SDT SDL DDT DDL LapD
94 92 90 88 86 GroupB GroupA All
84 82 80
80
100 Classification Accuracy (%)
Classification Accuracy (%)
Classification Accuracy (%)
96
70
60
50
90
40
MN
SDT
SDL DDT Spatial Filters
DDL
8
LapD
Fig. 2. Effect of spatial filtering on the classification accuracy for group B subjects, group A subjects and all the subjects using 89-channel high-density sEMG recordings.
1
2
3
4
5
9
6
7
10 11 12 Channel Number 8 9 10 11 12 13 14 15 16 17 18 19 20 Channel Number
Fig. 3. Effect of channel number on the classification accuracy for monopolar electrode sEMG recordings and 5 kinds of spatially filtered sEMG recordings.
758
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
(a)
and was found to be significantly lower (p = 0.047, p = 0.002 and p < 0.001 for 192 ms, 128 ms and 64 ms vs. 512 ms, respectively). It was revealed that the window length of 256 ms was a practical choice.
100
Classification Accuracy (%)
95 90
3.4. Re-sampling results 85
Fig. 4(b) further shows the effect of choosing different sampling rates on the classification results. It can be seen that there was a pronounced decrease in the average classification accuracy when the sampling rate went down from 2 kHz to 200 Hz. When comparing the accuracy achieved at the original 2 kHz with those at decreased sampling rates, the pairwise comparisons in ANOVA showed that significant differences in classification accuracy emerged when the sampling rate was set below 600 Hz (p = 0.065, p = 0.046 and p = 0.023 for 600 Hz, 550 Hz and 500 Hz vs. 2 kHz, respectively). Furthermore, the accuracy decreased more rapidly when the sampling rate was lower than 600 Hz (p < 0.05 for any sampling rate below 500 Hz vs. 2 kHz). However, choosing 600 Hz sampling rate led to a mean accuracy following below 90% (88.45%). It was decided that a 1 kHz sampling rate reached the practical point of diminishing returns as the mean accuracy was 90.88%, which only dropped 2.4% compared with that at 2 kHz sampling rate (p = 0.234).
80 75 70 65 60
(b)
64
128
192
256 320 384 Window Length (ms)
448
512
100
Classification Accuracy (%)
95 90 85
3.5. Re-filtering results
80 75 70 65 60 2000
1800
1600
1400
1200 1000 800 Sampling Rate (Hz)
600
400
200
95
Classification Accuracy (%)
(c)
90
Based on the above optimized settings, we also examined the choice of low cut-off frequency for a band-pass filter applied to the sEMG signals. Fig. 4(c) reports the classification results with a large range of low cut-off frequencies. Classification accuracy exhibited a decrease when a higher cut-off frequency was used. The pairwise comparisons following the ANOVA showed that compared with the accuracy achieved at the original 20 Hz cut-off frequency, the classification accuracy dropped down significantly (p = 0.065, p = 0.031 and p = 0.016 for 65 Hz, 70 Hz and 75 Hz vs. 20 Hz, respectively). This suggested that a cut-off frequency no higher than 65 Hz would be appropriate to avoid much compromise in classification performance. Surprisingly, it was also found that using a 60-Hz cut-off frequency had the average accuracy take a local maximum value over 90.37% and only decreased it by 0.3%, as compared to the original 20 Hz cut-off frequency (p = 0.076). Consequently, it was indicated that the cut-off frequency of 60 Hz was preferred. 4. Discussion
85
80 20
40
60 80 High-pass Cutoff Frequency (Hz)
100
120
Fig. 4. (a) Effect of window length on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used; (b) effect of sEMG sampling rate on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used. The window length was 256 ms; (c) effect of sEMG high-pass cut-off frequency on the classification accuracy for all the subjects. 8 selected spatially filtered SDL-sEMG channels were used. The window length was 256 ms and sampling rate was 1 kHz.
In this study, some technical issues have been examined for the sEMG classification of stroke survivors. Considering the unique sEMG signals obtained from people with neurologic injuries, the results found in this study were significant. sEMG can be used to interpret the user’s intention of volitional muscle activities, and control externally powered devices for rehabilitation. We demonstrated the feasibility of identifying multiple intended movements involving the affected limb of stroke subjects by applying pattern recognition techniques to high-density sEMG recordings [17]. The identified movement intentions can serve as rich control inputs to assistive devices in sEMG-driven robot-aided therapy and facilitate motor recovery after stroke [7–10]. In a previous study [17], the subject characteristics were reported as a table, where the subjects’ clinical information was presented. Some subjects with lower clinical functionality assessment scores tended to yield lower classification accuracy, but this was not always the case. The relationship between the classification performance and subjects’ clinical assessment scores remains unclear, presumably due to the fact that the mechanisms underlying stroke are complex and that the clinical assessment is relatively subjective. This study examined
Y. Li et al. / Medical Engineering & Physics 36 (2014) 754–760
several practical issues for development of a myoelectric pattern recognition control system for stroke rehabilitation. The spatial filtering tended to increase the classification accuracy. This may be because that the spatial filters are able to remove common-mode signal components in monopolar sEMG signals with lower spatial resolution. This outcome for stroke subjects is consistent with the results of previous studies for amputees [14,15]. The results showed that the SDL filter produced the highest classification accuracy, although statistical analysis exhibited no significant difference in classification performance among all examined spatial filters. Using the single differential spatial filter is a practical way for improving classification accuracy in clinical practice. Consistent with the previous studies [12,14,15], we demonstrated that 20 or less sEMG electrodes can maintain high level of classification accuracy. When 8 localized channels were used, the highest classification accuracy was achieved by SDL filter, and there was not a significant compromise when it was compared with the results for high-density sEMG signals. Considering that the single differential sEMG recording is indeed more clinically feasible, using a relatively small number (i.e., 8 in this study) of single differential sEMG sensors proves to be sufficient for implementing myoelectric control system in practice. It is noteworthy that the above channel selection procedure was conducted in a user-specific condition. The between-subject variability of selected electrode positions was also observed from a previous study [14]. Furthermore, it is possible to maintain high levels of classification accuracy by selecting a limited number of electrodes. In this regard, high-density sEMG provides important guidance for optimizing electrode number and location to detect the intended movements. After the number of electrodes has been selected with the appropriate configuration, data window length, sampling rate, and high-pass cut-off frequency for sEMG signal filtering are three factors that have a significant impact on myoelectric control and practical implementation. It is acknowledged that a larger window length or a larger sampling rate may carry more useful information in each analysis window that would help producing more accurate decisions [11]. In these cases, there would be demand for more computational power and memory beyond the limitation of currently available microprocessor-based controllers. Especially, a larger window length would also introduce a longer signal processing delay [1,14]. The low-frequency components of sEMG recordings do not contribute much to classification performance. Thus, a higher high-pass cut-off frequency would be chosen to enhance the system stability and reduce interference and ECG artifacts. The results suggest that a combination of a window length of 256 ms, a sampling rate of 1 kHz and a high-pass cut-off frequency of 60 Hz seem to be a proper setting for conditioning and preprocessing sEMG signals for pattern classification of movement intentions. The combined adjustments of both the sampling rate decreased from original 2 kHz to 1 kHz and the high-pass cut-off frequency increased from 20 Hz to 60 Hz only led to a slight decrease of the classification accuracy (below 3.7%), where no significant difference could be found. However, the window length used in this study might not be lower than 256 ms, otherwise the classification accuracy would be compromised. It is interesting to find from Fig. 4(c) that with the 60-Hz cut-off frequency, the overall classification accuracy even reaches a local maximum, which is not much less than the accuracy achieved with the original 20-Hz cut-off frequency. The explanation for this is that using the 60-Hz high-pass cut-off frequency does eliminate most low-frequency motion artifacts (
