652
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 39, NO. 6. J U N E 1992
Communications
Adaptive Cancellation Technique In Processing Myoelectric Activity of Respiratory Muscles Padma Akkiraju and D. C . Reddy
Abstract-Spectral analysis of myoelectric activity of the chest muscles is a promising diagnostic tool for respiratory diseases. Reduction of the corrupting cardiac activity from the recorded myoelectric activity of the chest muscles is essential because of the spectral overlap of both the signals. The adaptive noise cancellation technique was used to reduce the interfering cardiac activity from the recorded myoelectric activity. The adaptive noise canceller implemented in a transversal structure was found to successfully reduce the corrupting cardiac activity. This paper describes the cancellation of the corrupting cardiac activity from the recorded myoelectric activity using adaptive noise cancellation technique and the characteristic features of spectra of EMG signals.
nique for the present problem. The assumption of no correlation between the contaminating and contaminated signal is perfectly valid here because the sources of EMG and ECG are different. (Throughout this paper, ECG for cardiac activity and EMG for myoelectric activity have been used.) In the present work, Widrow’s adaptive noise canceller [ I ] was used to solve the problem of reducing the interfering ECG activity from the recorded EMG. The adaptive noise canceller in transversal structure, with ECG as reference input and EMG as primary input was found to successfully reduce the interfering cardiac activity. The filter coefficients were adjusted using the least mean squares (LMS) algorithm [ I ] , [2] to approach the set of weights for which the output has minimum noise in the least squared sense. Then the EMG signals were processed with these converged set of weights to obtain clean EMG.
11. ADAPTIVENOISECANCELLER A N D CONCEPTOF CORRELATION CANCELLATION
I. INTRODUCTION The importance of respiratory muscle function in the patho physiological features of the chronic obstructive pulmonary disease (COPD) syndrome is highlighted by the observations such as prominent use of accessory inspiratory muscles, impaired motion of the diaphragm, asynchronous motion of the abdomen, etc. Therefore, the electromyographic interference pattern detected by the surface electrodes placed on the external intercostal muscles could hopefully provide signals that could enable us to distinguish between a normal person and a patient with COPD [4],[ 7 ] , [ 8 ] . The acquisition of clean EMG (Electromyography) of the chest muscles is difficult because of the interfering caridac potentials developed in the human body because of the generator activity of the heart. Removal of the cardiac activity by direct subtraction is simply impossible because it is not feasible to obtain an identical copy of the cardiac activity that is corrupting the myoelectric activity because of the various nonlinearities associated with the signal path and recording. The analysis of spectra of both the cardiac activity and the myoelectric activity of the chest muscles clearly indicated the spectral overlap of both the signals between 1 and 50 Hz. Also, since the energy associated with both the signals is mainly contained in the above frequencies, any type of linear filtering cannot be used. Adaptive noise cancellation technique has been proved very effective for similar problems [ 1 ] , [3] where the signal and noise spectra overlap. Therefore, it was proposed to use the same techManuscript received March 14, 1989; revised January 1 I , 1990. P. Akkiraju was with the Department of Electrical and Computer Engineering, Osmania University, Hyderabad, India. She is now with Cirrus Logic, Inc., Fremont, CA 94538. D. C. Reddy is with the Department of Electrical and Computer Engineering, College of Engineering, Osmania University, Hyderabad, India. IEEE Log Number 9108132.
The block diagram of the adaptive noise canceller is shown in the Fig. 1 . The primary input to the noise canceller is the corrupted no where s is the original signal and no is the noise signal s corrupting it. The reference input n , is the separately recorded noise. The reference input n , is uncorrelated with the signal s but correlated with the noise no in some unknown way. The noise is filtered through an adaptive filter to produce the output y which is subtracted from the primary input s no to produce the output that is best fit in least squares sense to the signal s. This objective is accomplished by feeding the output of the filter back to adaptive filter and adjusting the coefficients (or weights) of the filter through an adaptive algorithm that minimizes the total output power. Hence, in adaptive noise cancellation system, the system output seems as an error signal for the adaptive process. The output is
+
+
To minimize the error in mean squared sense, taking expectation of e we get
’,
~ ( e ’ [ n l )= ~ ( s [ n +] no[n] - y[nl)’. Since s[n]and n , [ n ] and hence s [ n ] and y [ n ] are not correlated,
+
~ ( e * [ n l=) ~ ( s ’ [ n ] ) ~ ( n ~ [-n y[nl)*. ] The signal power remains constant during the adaptation. Therefore minimizing E ( e 2 [ n ] )minimizes E(no[n]- ~ [ n ] and ) ~ ,hence y [ n ] becomes approximately equal to no[n].Then, e [ n ] approximates s [ n ] ,thus minimizing the noise in the LMS sense.
0018-9294/92$03,00 0 1992 IEEE
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 39. NO. 6. JUNE 1992
653
EM;
111. IMPLEMENTATION DETAILS A . Equipment
The following equipment was used for signal acquisition and processing:
source
Purpose
...
Equipment
Signal acquisition
Ag-Agcl electrodes for EMG signals Suction electrodes for ECG signals Signal monitoring Nihon Kohden’s polygraph system RM and conditioning 6000 Signal recording Instrumentation FM tape recorder type 7005 (B&K) [dc to 60 kHz] Spectral analysis of the Dynamic spectrum analyzer HP-3561A signals [ 100 kHz] AID conversion Computer oscilloscope Model 4802 Sampling frequencies up to 100 kHz Adaptive noise canceller PC-AT, Fortran77 implementation
input
Fig. 1 . Adaptive noise canceller.
B. Signal Recording
1
I
m-no.of weights.
output y[n]
Fig. 2. Transversal structure of the adaptive filter. The adaptive filter is usually implemented as an finite impulse response (FIR) filter with finite number of weights m , because of its computational stability. The filter weights are adjusted using LMS algorithm. This algorithm adjusts the weights by estimating the gradient of e 2 [ n ] .By approximation, it takes e’ itself as an . this simple estimate, the LMS algorithm estimate of E ( e 2 [ n ] ) With adjusts the weights according to as follows:
w[n + I]
=
w[n] + 2pp[n]e[n]
+
I ] and w [ n ] are (m X 1) mawhere p [ n ] = s [ n ] + no[n].w[n trices of the filter weights. p [ n ] is also ( m X l ) matrix of the m latest signal samples including the current one. p is the convergence factor satisfying the condition 0 < p < 2/lmaXwhere I,, is the largest eigen value of the autocorrelation matrix of its input. The LMS algorithm adjusts the weights sample by sample basis. As processing of the signal takes place, filter gradually learns the required correlation and adjusts the weights accordingly. If there are any correlations between the primary and reference signals, the filter will adjust the coefficients to cancel these correlations from the output error signal. In the case where the noise is not time varying, the weights will converge after the filter leams about noise. These converged set of weights can be used further. But in the case of time-varying noise, the weights have to be dynamically updated and the leaming operation becomes simultaneous. The present problem is to obtain myoelectric activity of the chest muscles which is free of the corrupting cardiac activity. So the no of the Fig. 1 where no is recorded myoelectric activity is s the noise which is the ECG. It forms the primary input of the adaptive noise canceller. And, separately recorded ECG is n , of Fig. 1 and forms its reference input. Since the noise, which is the ECG, is nontime-varying, the weights converge after learning the correlation. The converged set of weights can be used for further processing. Fig. 2 shows the transversal structure of filter with EMG and ECG as primary and reference inputs.
+
EMG signal were recorded using Ag-AgC1 electrodes whose frequency response is flat up to 2 kHz, sufficient enough to record EMG signals. These Ag-AgC1 surface electrodes were chosen because the magnitude of the recorded EMG signals is high with surface electrodes [4]. The ECG signals were recorded using suction electrodes. Both the EMG and ECG signals were monitored using Nihon Kohden’s (NH) polygraph system and were recorded on FM tape recorder. Both ECG and EMG were recorded using a common channel of the NH system. The channel was equipped with biopotential amplifiers whose frequency response and time responses were designed suiting both EMG and ECG recording [SI. EMG signals from various positions on the chest were picked up using Ag-AgCI electrodes and were recorded using the FM tape recorder. Since the objective of obtaining a clear EMG is to use it for the diagnostic purpose, it was suggested that the EMG of the intercostal muscles be picked up when a subject executes deep and rapid (DR) breath maneuvers. (Usually the study of a respiratory disease is done by examining some features of the patient’s chest while he executes deep and rapid breath.) Moreover, the magnitude of the EMG signal recorded for normal breathing is low compared to the background noise. If such a noisy signal is used, it takes large number of iterations for the adaptive noise canceller to converge. The magnitude of the recorded signal has increased considerably for DR breath maneuver. Three electrodes were used for EMG recording. Two electrodes were placed on the intercostal space and the third electrode was placed on the arm connecting the body to ground. NH polygraph system is fully equipped different accessories which enable recording without disturbances (or artifacts) because of any physical movement during recording. ECG signals were also recorded from the same places as that of EMG signals. The ECG signals were recorded by placing suction electrode on the intercostal spaces. Both the EMG and the ECG recording was bipolar.
C . Spectral Analysis Frequency spectra of the recorded EMG and ECG signals were studied using the dynamic spectrum analyzer. Spectral analysis of EMG and ECG signals recorded from various interspaces on the chest yielded intersecting information about the energy distribution, variation of the EMG magnitude with the position of the elec-
654
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 39, NO. 6. JUNE 1992
trodes, and also about the interfering ECG activity. The ECG spectrum extends from 1 to 75 Hz with most of its energy in the frequencies 1-50 Hz. The fundamental spectral coefficient was at 1.25 Hz and the other harmonics were identified at 2.5, 3.75, 5 Hz, and so on. The period of the waveform was observed to be 0.8 sec. The spectrum of the EMG signal appeared noisy because of the random appearance of the signal. EMG spectrum extends between 0.5 and 200 Hz, with most of the energy lying between 0.5 and 75 Hz. The low frequencies of the EMG spectrum can be attributed to the fact that the frequency spectrum of the EMG signal shifts to lower frequencies during sustained contraction of chest muscles because of deep and rapid (DR) breath maneuvers [4], [8]. Spectral analysis of the EMG signals with different time record lengths revealed a few interesting features. The signals recorded from the upper part of the chest have less apparent EMG. These signals have dominant spectral peaks at 1.25, 2.5, 3.75 Hz, etc. These peaks are that of the corrupting ECG signal. The ECG signal has been already studied to spectral peaks at 1.25, 2.5, 3.75 Hz. In these signals, EMG spectral coefficients were very small compared to the ECG spectral coefficients. EMG activity is more pronounced as we move towards the lower part of the chest; i.e., at the 4th, Sth, and 6th interspaces. It was also noted that EMG is less pronounced in the waveforms recorded from the sternal margin than those from the midclavicular and midaxillary linear lines. Among all positions on the chest, EMG content of the signals from the 4th, 5th, 6th interspaces on the midaxillary line was large and hence more suitable for diagnostic purpose. A thorough study of the spectra of EMG signals from various interspaces indicated spectral coefficients at 0.8, 1.6, 2.4, 3.2 Hz, and so on. These signals are those recorded when the subject was executing DR breath maneuver. Interestingly the spectrum of the EMG signal recorded for one deep and rapid breath also indicated spectral coefficients at 0.8, 1.6, 3.2, 4.8 Hz, and so on. Since the spectra of the EMG activity corresponding to DR breath, with time record lengths taken as multiples of duration of one deep and rapid breath, indicated spectral coefficients at the same frequencies as those of EMG corresponding to one DR breath, we may say that the EMG of the continuous DR breath is periodic with a period equal to the duration of the activity for one DR breath. However this periodicity of breathing is not apparent in the recorded time domain signals since the respiratory muscles were not given enough time to depolarize during continuous DR breathing.
2 volts
o sec
volts
U'! 0 HZ
16 sec
-Time
I
25 HZ
-Frequency
Fig. 3. ECG signal from the 3rd intercostal space on the mid axillary line on the right side of the chest. +2 volts
y1
f
-2 VOllSl
I
I -Tme
I
I
I
12.5 sec
'
0 HZ
1
-Frequency
32 HZ
(a) +2
volts
volts I
t
I
I
0 sec
-Time
5 sec
D. Signal Processing
(b) Fig. 4. (a) EMG signal from the 5th intercostal space on the mid axillary line on the right side of the chest. (b) EMG signal recorded for one deep and rapid breath.
The recorded EMG and ECG signals were sampled at a frequency greater than twice their maximum frequency; i.e., at 500 Hz meeting the Nyquist criteria for signal reconstruction. The samples were stored on floppy disks. These samples were fed to the Adaptive noise canceller implemented in Fortran 77 on IBM PC-AT. The corrupted EMG was the primary input to the filter and ECG was the reference input. The adaptive filter was implemented as FIR filter of ten coefficients. The coefficients were adjusted using LMS algorithm. The step size p was computed from the maximum eigen value of the auto-correlation matrix of the ECG signal's samples (i.e., the reference input to the filter). On average a step size of 0.01 was decided meeting the convergence criteria [2]. This step size was found to work satisfactorily obtaining convergence of weights after two thousand iterations. The processing was done in two stages. First, the converged set of weights were obtained after the system adapts. Next, these converged weights were used to process the signals. The output of the filter after the weights converge is relatively free of noise.
With a step size of 0.01 and with ten filter coefficients, the adaptive noise canceller was found to work satisfactorily with a reduction in the ECG activity up to 85 % . The filter coefficients were observed to converge around 2000 iterations. Spectra of EMG signals indicated 85% reduction in the spectral coefficients of ECG activity. Fig. 3 shows the ECG signals recorded from the third intercostal space on right side of the chest on the midaxillary line. The frequency spectrum of this signal is also shown along with the signal in Fig. 3. The top frame of Fig. 3 is the frequency spectrum. One can clearly see the spectral peaks at 1.25, 2.5, 3.75 Hz. and so on. Frequency spectra of the ECG signals recorded from other parts of the chest have also showed the same spectral characteristics. Fig. 4(a) shows the EMG signals recorded for continuous DR breathing from the positions as indicated. Fig. 4(b) shows the EMG
IV. RESULTS
IEEE TRANSACTlONS ON BIOMEDICAL ENGINEERING, VOL. 39. NO. 6. J U N E 1992 +2
+2
I
VOllSl
volts
o
,
sec
,
,
,
,
,
,
.
I
-Time
I
,
I
,
,
.
,
.
1 . 6 sec
655
TABLE 1 SPECTRAL COEFFICIENTS FROM THE FFT MAGNITUDE SPECIRUM OF THF EMG SIGNAL FROM THE 4TH INTERCOSTAL SPACE ON THE MIDCLAVICULAR LINEON THE RIGHT SIDEO F THE CHEST Frequency in Hz
Normalized Magnitude before Processing
Normalized Magnitude after Processing
1.60 2.50 4.80 5.00 I I .25 16.25 18.75 23.75 29.60 31.20 35.20 36.80
0.26310 0.52030 1 .00000 0.96490 0.75430 0.91228 0.67.540 0.46491 0.55263 0.49122 0.28947 0.44736
0.7 I92 1 0.26315 1 .00000 0.6.5798 0.71928 0.64912 0.57010 0.29825 0.5 5203 0.49 122 0.29820 0.46491
peaks at 1.25,2.5, 3.75 Hz, and so on. These are the spectral peaks identified in spectra of ECG signals. Reduction of these spectral peaks confirms the effectiveness of the technique in reducing the corrupting ECG signals. Table I shows these results. The reduction in the magnitude of the spectral coefficients of the ECG signals in all the processed EMG signals was computed to be approximately
85%. EMG Signals recorded from various other parts of the chest were also processed and the adaptive noise cancellation technique worked well with all of them. Figures showing those signals and their processing are not included because of the limitation on the space.
V . CONCLUSION
o
sec
-Time
I 6 sec
(c) Fig. 5 . (a) ECG signal. Reference input to the noise canceller. (b) Corrupted EMG signal. Primary input to the noise canceller. (c) EMG signal after the noise cancellation using the converged set of weights. [Notice the reduction in QRS complexes of ECG signal when compared to (b).]
signal recorded for a single deep and rapid breath. A close observation of these signals along with the ECG signals clearly shows the Q R S complexes of the corrupting ECG signals and the randomly appearing EMG signals. Fig. 5(a)-(c) show the snap shots of the results obtained by processing the EMG signals from the 4th intercostal space on the mid clavicular line on the right side of the chest. Fig. 5(a) shows the reference input to the noise canceller, i.e. the separately recorded ECG signal. Fig. 5(b) shows the recorded EMG signal corrupted by ECG signal. The ECG signal’s negative peaks can be observed in this figure. Fig. 5(c) shows the filtered EMG signal relatively free of noise. The ECG signal peaks are considerably reduced enchancing the EMG signal. A close look at the EMG signals before and after processing [Fig. 5(b), (c)] show the reduction in Q R S complexes of the ECG signal. Apart from this, comparison of the spectra of the EMG signals after noise cancellation with those of before processing, showed reduction in magnitude of the spectral
The adaptive noise cancellation technique was found to be effective in reducing the corrupting ECG activity. The EMG signals recorded from 4th, 5th, and 6th intercostal spaces are rich in the myoelectric activity of the respiratory musculature. Obtaining relatively noiseless and clean EMG of the chest muscles for diagnostic purpose is now feasible by using adaptive noise canceller. A study of the clean EMG signals from these positions would give us valuable information for diagnosing COPD’s. Power spectral densities, cumulative power difference function, mean frequency of the frequency spectrum of the EMG signals could serve as good measures in differentiating abnormalities pertaining to various respiratory diseases [4], [ 7 ] . REFERENCES [I] B. Widrow and S . Steams, Adaptive Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [2] 0. Sophcles, Optimal Signal Processing. New York: Macmillan, 1985. [3] V. K. Iyer and P. A. Ramamoorthy, “Reduction of heart sounds from lung sounds by adaptive filtering,” IEEE Trans. Biomed. E n g . , vol. BME-33, Dec. 1986. [4]F. B. Stulen and C. J . De Luca, “Frequency parameters of the myoelectric signal as a measure of muscle conduction velocity,” IEEE Trans. Biomed. E n g . , vol. BME-28, July 1981. 151 Technical Reference Manual, Nihon Kohden’s Polygraph System. [6] J . G. Webster, Medical Instrumentation, Application and Design. 171 E. Kwatny, D. H. Thomas, and H. G. Kwatny, “An application of signal processing techniques, IEEE Trans. Biomed. E n g . , vol. BME-17, Oct. 1970. [8] P. L. Davies, “Diaphragmatic electromyographic studies, the application of frequency domain and time domain analysis,” Amer. Rev. Respiratory Diseases, vol. 119. Feb. 1979. ”
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 39, NO. 6. J U N E 1992
Communications
Adaptive Cancellation Technique In Processing Myoelectric Activity of Respiratory Muscles Padma Akkiraju and D. C . Reddy
Abstract-Spectral analysis of myoelectric activity of the chest muscles is a promising diagnostic tool for respiratory diseases. Reduction of the corrupting cardiac activity from the recorded myoelectric activity of the chest muscles is essential because of the spectral overlap of both the signals. The adaptive noise cancellation technique was used to reduce the interfering cardiac activity from the recorded myoelectric activity. The adaptive noise canceller implemented in a transversal structure was found to successfully reduce the corrupting cardiac activity. This paper describes the cancellation of the corrupting cardiac activity from the recorded myoelectric activity using adaptive noise cancellation technique and the characteristic features of spectra of EMG signals.
nique for the present problem. The assumption of no correlation between the contaminating and contaminated signal is perfectly valid here because the sources of EMG and ECG are different. (Throughout this paper, ECG for cardiac activity and EMG for myoelectric activity have been used.) In the present work, Widrow’s adaptive noise canceller [ I ] was used to solve the problem of reducing the interfering ECG activity from the recorded EMG. The adaptive noise canceller in transversal structure, with ECG as reference input and EMG as primary input was found to successfully reduce the interfering cardiac activity. The filter coefficients were adjusted using the least mean squares (LMS) algorithm [ I ] , [2] to approach the set of weights for which the output has minimum noise in the least squared sense. Then the EMG signals were processed with these converged set of weights to obtain clean EMG.
11. ADAPTIVENOISECANCELLER A N D CONCEPTOF CORRELATION CANCELLATION
I. INTRODUCTION The importance of respiratory muscle function in the patho physiological features of the chronic obstructive pulmonary disease (COPD) syndrome is highlighted by the observations such as prominent use of accessory inspiratory muscles, impaired motion of the diaphragm, asynchronous motion of the abdomen, etc. Therefore, the electromyographic interference pattern detected by the surface electrodes placed on the external intercostal muscles could hopefully provide signals that could enable us to distinguish between a normal person and a patient with COPD [4],[ 7 ] , [ 8 ] . The acquisition of clean EMG (Electromyography) of the chest muscles is difficult because of the interfering caridac potentials developed in the human body because of the generator activity of the heart. Removal of the cardiac activity by direct subtraction is simply impossible because it is not feasible to obtain an identical copy of the cardiac activity that is corrupting the myoelectric activity because of the various nonlinearities associated with the signal path and recording. The analysis of spectra of both the cardiac activity and the myoelectric activity of the chest muscles clearly indicated the spectral overlap of both the signals between 1 and 50 Hz. Also, since the energy associated with both the signals is mainly contained in the above frequencies, any type of linear filtering cannot be used. Adaptive noise cancellation technique has been proved very effective for similar problems [ 1 ] , [3] where the signal and noise spectra overlap. Therefore, it was proposed to use the same techManuscript received March 14, 1989; revised January 1 I , 1990. P. Akkiraju was with the Department of Electrical and Computer Engineering, Osmania University, Hyderabad, India. She is now with Cirrus Logic, Inc., Fremont, CA 94538. D. C. Reddy is with the Department of Electrical and Computer Engineering, College of Engineering, Osmania University, Hyderabad, India. IEEE Log Number 9108132.
The block diagram of the adaptive noise canceller is shown in the Fig. 1 . The primary input to the noise canceller is the corrupted no where s is the original signal and no is the noise signal s corrupting it. The reference input n , is the separately recorded noise. The reference input n , is uncorrelated with the signal s but correlated with the noise no in some unknown way. The noise is filtered through an adaptive filter to produce the output y which is subtracted from the primary input s no to produce the output that is best fit in least squares sense to the signal s. This objective is accomplished by feeding the output of the filter back to adaptive filter and adjusting the coefficients (or weights) of the filter through an adaptive algorithm that minimizes the total output power. Hence, in adaptive noise cancellation system, the system output seems as an error signal for the adaptive process. The output is
+
+
To minimize the error in mean squared sense, taking expectation of e we get
’,
~ ( e ’ [ n l )= ~ ( s [ n +] no[n] - y[nl)’. Since s[n]and n , [ n ] and hence s [ n ] and y [ n ] are not correlated,
+
~ ( e * [ n l=) ~ ( s ’ [ n ] ) ~ ( n ~ [-n y[nl)*. ] The signal power remains constant during the adaptation. Therefore minimizing E ( e 2 [ n ] )minimizes E(no[n]- ~ [ n ] and ) ~ ,hence y [ n ] becomes approximately equal to no[n].Then, e [ n ] approximates s [ n ] ,thus minimizing the noise in the LMS sense.
0018-9294/92$03,00 0 1992 IEEE
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 39. NO. 6. JUNE 1992
653
EM;
111. IMPLEMENTATION DETAILS A . Equipment
The following equipment was used for signal acquisition and processing:
source
Purpose
...
Equipment
Signal acquisition
Ag-Agcl electrodes for EMG signals Suction electrodes for ECG signals Signal monitoring Nihon Kohden’s polygraph system RM and conditioning 6000 Signal recording Instrumentation FM tape recorder type 7005 (B&K) [dc to 60 kHz] Spectral analysis of the Dynamic spectrum analyzer HP-3561A signals [ 100 kHz] AID conversion Computer oscilloscope Model 4802 Sampling frequencies up to 100 kHz Adaptive noise canceller PC-AT, Fortran77 implementation
input
Fig. 1 . Adaptive noise canceller.
B. Signal Recording
1
I
m-no.of weights.
output y[n]
Fig. 2. Transversal structure of the adaptive filter. The adaptive filter is usually implemented as an finite impulse response (FIR) filter with finite number of weights m , because of its computational stability. The filter weights are adjusted using LMS algorithm. This algorithm adjusts the weights by estimating the gradient of e 2 [ n ] .By approximation, it takes e’ itself as an . this simple estimate, the LMS algorithm estimate of E ( e 2 [ n ] ) With adjusts the weights according to as follows:
w[n + I]
=
w[n] + 2pp[n]e[n]
+
I ] and w [ n ] are (m X 1) mawhere p [ n ] = s [ n ] + no[n].w[n trices of the filter weights. p [ n ] is also ( m X l ) matrix of the m latest signal samples including the current one. p is the convergence factor satisfying the condition 0 < p < 2/lmaXwhere I,, is the largest eigen value of the autocorrelation matrix of its input. The LMS algorithm adjusts the weights sample by sample basis. As processing of the signal takes place, filter gradually learns the required correlation and adjusts the weights accordingly. If there are any correlations between the primary and reference signals, the filter will adjust the coefficients to cancel these correlations from the output error signal. In the case where the noise is not time varying, the weights will converge after the filter leams about noise. These converged set of weights can be used further. But in the case of time-varying noise, the weights have to be dynamically updated and the leaming operation becomes simultaneous. The present problem is to obtain myoelectric activity of the chest muscles which is free of the corrupting cardiac activity. So the no of the Fig. 1 where no is recorded myoelectric activity is s the noise which is the ECG. It forms the primary input of the adaptive noise canceller. And, separately recorded ECG is n , of Fig. 1 and forms its reference input. Since the noise, which is the ECG, is nontime-varying, the weights converge after learning the correlation. The converged set of weights can be used for further processing. Fig. 2 shows the transversal structure of filter with EMG and ECG as primary and reference inputs.
+
EMG signal were recorded using Ag-AgC1 electrodes whose frequency response is flat up to 2 kHz, sufficient enough to record EMG signals. These Ag-AgC1 surface electrodes were chosen because the magnitude of the recorded EMG signals is high with surface electrodes [4]. The ECG signals were recorded using suction electrodes. Both the EMG and ECG signals were monitored using Nihon Kohden’s (NH) polygraph system and were recorded on FM tape recorder. Both ECG and EMG were recorded using a common channel of the NH system. The channel was equipped with biopotential amplifiers whose frequency response and time responses were designed suiting both EMG and ECG recording [SI. EMG signals from various positions on the chest were picked up using Ag-AgCI electrodes and were recorded using the FM tape recorder. Since the objective of obtaining a clear EMG is to use it for the diagnostic purpose, it was suggested that the EMG of the intercostal muscles be picked up when a subject executes deep and rapid (DR) breath maneuvers. (Usually the study of a respiratory disease is done by examining some features of the patient’s chest while he executes deep and rapid breath.) Moreover, the magnitude of the EMG signal recorded for normal breathing is low compared to the background noise. If such a noisy signal is used, it takes large number of iterations for the adaptive noise canceller to converge. The magnitude of the recorded signal has increased considerably for DR breath maneuver. Three electrodes were used for EMG recording. Two electrodes were placed on the intercostal space and the third electrode was placed on the arm connecting the body to ground. NH polygraph system is fully equipped different accessories which enable recording without disturbances (or artifacts) because of any physical movement during recording. ECG signals were also recorded from the same places as that of EMG signals. The ECG signals were recorded by placing suction electrode on the intercostal spaces. Both the EMG and the ECG recording was bipolar.
C . Spectral Analysis Frequency spectra of the recorded EMG and ECG signals were studied using the dynamic spectrum analyzer. Spectral analysis of EMG and ECG signals recorded from various interspaces on the chest yielded intersecting information about the energy distribution, variation of the EMG magnitude with the position of the elec-
654
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 39, NO. 6. JUNE 1992
trodes, and also about the interfering ECG activity. The ECG spectrum extends from 1 to 75 Hz with most of its energy in the frequencies 1-50 Hz. The fundamental spectral coefficient was at 1.25 Hz and the other harmonics were identified at 2.5, 3.75, 5 Hz, and so on. The period of the waveform was observed to be 0.8 sec. The spectrum of the EMG signal appeared noisy because of the random appearance of the signal. EMG spectrum extends between 0.5 and 200 Hz, with most of the energy lying between 0.5 and 75 Hz. The low frequencies of the EMG spectrum can be attributed to the fact that the frequency spectrum of the EMG signal shifts to lower frequencies during sustained contraction of chest muscles because of deep and rapid (DR) breath maneuvers [4], [8]. Spectral analysis of the EMG signals with different time record lengths revealed a few interesting features. The signals recorded from the upper part of the chest have less apparent EMG. These signals have dominant spectral peaks at 1.25, 2.5, 3.75 Hz, etc. These peaks are that of the corrupting ECG signal. The ECG signal has been already studied to spectral peaks at 1.25, 2.5, 3.75 Hz. In these signals, EMG spectral coefficients were very small compared to the ECG spectral coefficients. EMG activity is more pronounced as we move towards the lower part of the chest; i.e., at the 4th, Sth, and 6th interspaces. It was also noted that EMG is less pronounced in the waveforms recorded from the sternal margin than those from the midclavicular and midaxillary linear lines. Among all positions on the chest, EMG content of the signals from the 4th, 5th, 6th interspaces on the midaxillary line was large and hence more suitable for diagnostic purpose. A thorough study of the spectra of EMG signals from various interspaces indicated spectral coefficients at 0.8, 1.6, 2.4, 3.2 Hz, and so on. These signals are those recorded when the subject was executing DR breath maneuver. Interestingly the spectrum of the EMG signal recorded for one deep and rapid breath also indicated spectral coefficients at 0.8, 1.6, 3.2, 4.8 Hz, and so on. Since the spectra of the EMG activity corresponding to DR breath, with time record lengths taken as multiples of duration of one deep and rapid breath, indicated spectral coefficients at the same frequencies as those of EMG corresponding to one DR breath, we may say that the EMG of the continuous DR breath is periodic with a period equal to the duration of the activity for one DR breath. However this periodicity of breathing is not apparent in the recorded time domain signals since the respiratory muscles were not given enough time to depolarize during continuous DR breathing.
2 volts
o sec
volts
U'! 0 HZ
16 sec
-Time
I
25 HZ
-Frequency
Fig. 3. ECG signal from the 3rd intercostal space on the mid axillary line on the right side of the chest. +2 volts
y1
f
-2 VOllSl
I
I -Tme
I
I
I
12.5 sec
'
0 HZ
1
-Frequency
32 HZ
(a) +2
volts
volts I
t
I
I
0 sec
-Time
5 sec
D. Signal Processing
(b) Fig. 4. (a) EMG signal from the 5th intercostal space on the mid axillary line on the right side of the chest. (b) EMG signal recorded for one deep and rapid breath.
The recorded EMG and ECG signals were sampled at a frequency greater than twice their maximum frequency; i.e., at 500 Hz meeting the Nyquist criteria for signal reconstruction. The samples were stored on floppy disks. These samples were fed to the Adaptive noise canceller implemented in Fortran 77 on IBM PC-AT. The corrupted EMG was the primary input to the filter and ECG was the reference input. The adaptive filter was implemented as FIR filter of ten coefficients. The coefficients were adjusted using LMS algorithm. The step size p was computed from the maximum eigen value of the auto-correlation matrix of the ECG signal's samples (i.e., the reference input to the filter). On average a step size of 0.01 was decided meeting the convergence criteria [2]. This step size was found to work satisfactorily obtaining convergence of weights after two thousand iterations. The processing was done in two stages. First, the converged set of weights were obtained after the system adapts. Next, these converged weights were used to process the signals. The output of the filter after the weights converge is relatively free of noise.
With a step size of 0.01 and with ten filter coefficients, the adaptive noise canceller was found to work satisfactorily with a reduction in the ECG activity up to 85 % . The filter coefficients were observed to converge around 2000 iterations. Spectra of EMG signals indicated 85% reduction in the spectral coefficients of ECG activity. Fig. 3 shows the ECG signals recorded from the third intercostal space on right side of the chest on the midaxillary line. The frequency spectrum of this signal is also shown along with the signal in Fig. 3. The top frame of Fig. 3 is the frequency spectrum. One can clearly see the spectral peaks at 1.25, 2.5, 3.75 Hz. and so on. Frequency spectra of the ECG signals recorded from other parts of the chest have also showed the same spectral characteristics. Fig. 4(a) shows the EMG signals recorded for continuous DR breathing from the positions as indicated. Fig. 4(b) shows the EMG
IV. RESULTS
IEEE TRANSACTlONS ON BIOMEDICAL ENGINEERING, VOL. 39. NO. 6. J U N E 1992 +2
+2
I
VOllSl
volts
o
,
sec
,
,
,
,
,
,
.
I
-Time
I
,
I
,
,
.
,
.
1 . 6 sec
655
TABLE 1 SPECTRAL COEFFICIENTS FROM THE FFT MAGNITUDE SPECIRUM OF THF EMG SIGNAL FROM THE 4TH INTERCOSTAL SPACE ON THE MIDCLAVICULAR LINEON THE RIGHT SIDEO F THE CHEST Frequency in Hz
Normalized Magnitude before Processing
Normalized Magnitude after Processing
1.60 2.50 4.80 5.00 I I .25 16.25 18.75 23.75 29.60 31.20 35.20 36.80
0.26310 0.52030 1 .00000 0.96490 0.75430 0.91228 0.67.540 0.46491 0.55263 0.49122 0.28947 0.44736
0.7 I92 1 0.26315 1 .00000 0.6.5798 0.71928 0.64912 0.57010 0.29825 0.5 5203 0.49 122 0.29820 0.46491
peaks at 1.25,2.5, 3.75 Hz, and so on. These are the spectral peaks identified in spectra of ECG signals. Reduction of these spectral peaks confirms the effectiveness of the technique in reducing the corrupting ECG signals. Table I shows these results. The reduction in the magnitude of the spectral coefficients of the ECG signals in all the processed EMG signals was computed to be approximately
85%. EMG Signals recorded from various other parts of the chest were also processed and the adaptive noise cancellation technique worked well with all of them. Figures showing those signals and their processing are not included because of the limitation on the space.
V . CONCLUSION
o
sec
-Time
I 6 sec
(c) Fig. 5 . (a) ECG signal. Reference input to the noise canceller. (b) Corrupted EMG signal. Primary input to the noise canceller. (c) EMG signal after the noise cancellation using the converged set of weights. [Notice the reduction in QRS complexes of ECG signal when compared to (b).]
signal recorded for a single deep and rapid breath. A close observation of these signals along with the ECG signals clearly shows the Q R S complexes of the corrupting ECG signals and the randomly appearing EMG signals. Fig. 5(a)-(c) show the snap shots of the results obtained by processing the EMG signals from the 4th intercostal space on the mid clavicular line on the right side of the chest. Fig. 5(a) shows the reference input to the noise canceller, i.e. the separately recorded ECG signal. Fig. 5(b) shows the recorded EMG signal corrupted by ECG signal. The ECG signal’s negative peaks can be observed in this figure. Fig. 5(c) shows the filtered EMG signal relatively free of noise. The ECG signal peaks are considerably reduced enchancing the EMG signal. A close look at the EMG signals before and after processing [Fig. 5(b), (c)] show the reduction in Q R S complexes of the ECG signal. Apart from this, comparison of the spectra of the EMG signals after noise cancellation with those of before processing, showed reduction in magnitude of the spectral
The adaptive noise cancellation technique was found to be effective in reducing the corrupting ECG activity. The EMG signals recorded from 4th, 5th, and 6th intercostal spaces are rich in the myoelectric activity of the respiratory musculature. Obtaining relatively noiseless and clean EMG of the chest muscles for diagnostic purpose is now feasible by using adaptive noise canceller. A study of the clean EMG signals from these positions would give us valuable information for diagnosing COPD’s. Power spectral densities, cumulative power difference function, mean frequency of the frequency spectrum of the EMG signals could serve as good measures in differentiating abnormalities pertaining to various respiratory diseases [4], [ 7 ] . REFERENCES [I] B. Widrow and S . Steams, Adaptive Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [2] 0. Sophcles, Optimal Signal Processing. New York: Macmillan, 1985. [3] V. K. Iyer and P. A. Ramamoorthy, “Reduction of heart sounds from lung sounds by adaptive filtering,” IEEE Trans. Biomed. E n g . , vol. BME-33, Dec. 1986. [4]F. B. Stulen and C. J . De Luca, “Frequency parameters of the myoelectric signal as a measure of muscle conduction velocity,” IEEE Trans. Biomed. E n g . , vol. BME-28, July 1981. 151 Technical Reference Manual, Nihon Kohden’s Polygraph System. [6] J . G. Webster, Medical Instrumentation, Application and Design. 171 E. Kwatny, D. H. Thomas, and H. G. Kwatny, “An application of signal processing techniques, IEEE Trans. Biomed. E n g . , vol. BME-17, Oct. 1970. [8] P. L. Davies, “Diaphragmatic electromyographic studies, the application of frequency domain and time domain analysis,” Amer. Rev. Respiratory Diseases, vol. 119. Feb. 1979. ”
