You are on page 1of 42

EMG SIGNAL DECOMPOSITION

USING PATTERN CLASSIFICATION


TECHNIQUES

Under the guidance of


Dr. VINOD 

KUMAR 

PROFESSOR &HEAD  By
 

 niranjana rao
 kakarla
 M.Tech.(M&I)
 08528010
 I.I.T. Roorkee
Topics to be discussed:-
EMG signal composition
Steps involved in decomposition
Single classifier approaches
Multiple classifier approaches
Comparative study


EMG SIGNAL COMPOSITION
Ø Electromyography is the detection of muscle activity
associated with muscle contraction

Ø MOTOR UNIT: a motor unit is an α-motoneuron and all the


muscle fibres innervated by its axon

Ø MUP waveform: the summation of the spatially and
temporally dispersed potentials, created by impulses
propagating along the individual muscle fibers of a motor
unit

Ø MUPT: is the collection of MUP waveforms generated by one


motor unit separated by their inter-discharge intervals


Fig. 1. MUPT with MUPs separated by their inter-discharge
intervals[1]

Ø The superposition of the MUPTs of all recruited motor units


and background noise comprises EMG signal

 Fig. 2. A 1-s epoch of raw EMG signal[2]


E M G S IG N A L D E C O M P O S IT IO N
 process of resolving a composite EMG signal in to its
constituent MUPTs

 Figure.3. Schematic representation of the detection and


decomposition of the intramuscular EMG signal[3].


Applications of decomposition
 The shapes and occurrence times of MUPs provide an
important source of information to assist in the diagnosis of
neuromuscular disorders.

 reflect the structural and physiological changes of a MU


 central nervous system recruitment and control of MUs can


be
 studied.
 Major steps involved in EMG signal decomposition are

 1. segmentation

 2. feature extraction

 3. clustering

 4. supervised classification


Segmentation of the composite signal

Segmentation refers to the detection of all MUAPs


generated by MUs active during signal acquisition.

 Fig. 4. (Top) Raw EMG segment recorded from biceps brachii using

concentric needle electrode-a typical "mixed interference pattern" in which


the individual MUAP's are difficult to identify and characterize because of

superimpositions. (bottom) Same segment after filtering using the second-


order differentiating filter[4].
Fe a tu re extra ctio n
 Feature

Ø The detected MUAPs must be represented using a vector for


pattern recognition.
Ø The characteristics of the MUAPs used for this representation
are called features.

 Feature spaces

Ø The multi-dimensional space composed of all possible feature


values is called the feature space.
Ø

 feature spaces used to represent MUAPs for signal
decomposition are:

1. peak to peak voltage, number of phases, duration, number of

turns, etc.

2. Fourier transformation coefficients

3. Wavelet coefficients

4. time samples of the band-pass filtered signal

5. time samples of the low-pass differentiated signal.



CLUSTERING OF DETECTED MUAPs

 partitioning detected MUAPs into a number of groups or


clusters.





 
 





 


 The goal of clustering algorithm is

 1. Determine prototypical shape of each motor


unit
 2. Determine the number of motor units
Ø The clustering algorithm does not have to analyse the entire
composite signal
Ø it does not have to assign every detected MUAP.
Ø Once prototypical MUAPs are available the complete set of
detected MUAPs can be classified using supervised
classification techniques.
SINGLE CLASSIFIER APPROACHES
 CERTAINITY CLASSIFIER

Ø It is a non-parametric template matching classifier

Ø Uses certainty based approach for assigning MUPs to MUPTs

Ø For a set of M MUPT class labels Ω = {ω1,ω2,……..,ωM} the decision


functions for the assignment of MUP mj are evaluated for only

the two MUPTs with the most similar templates to MUP mj

Ø
Ø It determines two types of decision functions
 1. based on shape information
 2. based on firing pattern information
Ø
 The shape information decision functions include:

 1 .Normalized shape certainty CND : represents the distance

from a candidate MUP mj to the template of a MUPT ωi


1.

 2.Relative shape certainty CRD :represents the distance from a

candidate MUP mjto the template of the closest MUPT relative to

the distance from the same MUP to the second closest MUPT.
1.

1.
Ø The firing pattern information is presented by the firing certainty
based function CFC with respect to the established firing pattern of
the MUPT

Ø The decision ofCassigning


j a MUP mj to MUPT ωi is based on the value
i

of overall certainty
Ci j = CNDi
j j
.CRDi j
.CFCi
 where i=1,2.

Ø MUPT mj is assigned to MUPT ωi if overall certainty is the


greatest and if it is greater than the minimal certainty threshold
(Cm) for which a classification is to be made.

Ø Other wise MUP mj is left unassigned.


FUZZY k-NN CLASSIFIER
Ø The fuzzy k-NN classifier uses a fuzzy non-parametric classification
procedure based on the nearest neighbor classification rule.
Ø By passes probability estimation and goes directly to decision
functions.

Ø It determines for each candidate MUP mj a MUPT ωi membership μωi


(mj) representing the shape based strength of
j
membership of mj
AFAi
in MUPT class ωi

Ø Also determines firing assertion decision function assesingthe time


of occurrence of MUPT mj with respect to the established firing

pattern of MUPT class ωi



Ø The overall assertion value
j for assigning MUP mj to MUPT class
Ai
ωi is defined as:

Ø

Ai j = µ ω i ( m ).
j
j
AFAi

Ø MUP mj is assigned to the MUPT class ωi with the highest assertion


value and if this value is above the minimum assertion value
threshold (Am) of the MUPT ωi to which a classification is to be
made.

Ø Otherwise MUP mj is left unassigned.



MATCHED TEMPLATE FILTER CLASSIFIER
Ø This method uses correlation measure as an estimate of the degree
of similarity between MUP and MUPT templates.
Ø Two matched template filters are used for supervised MUP
classification
 1. normalized cross correlation
 2. pseudo- correlation

Ø The MTF classifier also determines for MUP mj a firing time similarity
j
S FSi
decision function with respect to the established firing pattern
of the MUPT


Ø The decision of assigning a MUP to a MUPT is based on the overall
similarity
j
function
Si
 Wherej
Si = ρjωi x( ). j
SFSi
Ø MUPT mj is assigned to MUPT ωiif the value ofj
is greatest and it is
Si
greater than minimal similarity threshold (sm) for which a
classification is to be made.

Ø Otherwise MUP mj is left unassigned.



Multiple Classifier Approaches
Ø Classifier fusion system architecture belongs to the
parallel category of combining classifiers.

Fig.6.Classifier fusion system basic


architecture[6]
Ø The decision aggregation module in a classifier fusion system
combines the base classifier outputs to achieve a group
consensus.
 Majority Voting Aggregation

Ø A MUP x is classified to belong to MUPT class ωi if over half of


the classifiers say x ∈ ωi .

 Average rule aggregation

Ø Combines the set of Kdecision confidences


∑ Cf ik ( x)
 Qi ( x) = k =1

K

ω( x)= argmax iM=(1Q (ix))


One-Stage Classifier Fusion
Ø Does not contain the ensemble members selection module

Ø Uses a fixed set of base classifiers

Ø Choosing base classifiers can be performed directly through

exhaustive search with the performance of the fusion being

the objective function.

Ø As the number of base classifiers increases, this approach

becomes computationally too expensive.


Diversity-Based One-Stage Classifier Fusion
Ø Contains an ensemble members selection module

Ø The ensemble choice module selects the subsets of classifiers


that can be combined to achieve better accuracy

Ø The kappa statistic is used to select base classifiers having an


excellent level of agreement to form ensembles having
satisfactory classification performance.
Hybrid Classifier Fusion
Ø Does not contain the ensemble members selection module

Ø Uses a fixed set of base classifiers

Ø It uses a hybrid aggregation module which is a combination of

two stages of aggregation

Ø The first aggregator is based on the abstract level and the

second is based on the measurement level


 The hybrid aggregation scheme works as follows:
 First stage:
Ø The outputs of the ensemble of classifiers are presented to
the majority voting aggregator.
Ø If all classifiers state a decision that a MUP pattern is left
unassigned, then it stays unassigned.
Ø If over half of the classifiers assign a MUP pattern to the same
MUPT class, then that MUP pattern is allocated to that MUPT
class and no further assignment is processed.
 Second stage:

Ø This stage is activated for those MUP patterns for which only

half or less than half of the ensemble of classifiers in the

first stage specify a decision for a MUP pattern to be

assigned to the same MUPT class.

Ø The outputs of the ensemble of classifiers are presented to

the average rule aggregator, or the trainable aggregator

represented by the Sugeno fuzzy integral.


Ø For each MUP pattern, the overall combined confidence
values representing the degree of membership in each
MUPT class are determined
Ø
Ø MUP pattern is assigned to the MUPT class for which its
determined overall combined confidence is the largest and
if it is above the specified aggregator confidence threshold
set for that MUPT class
Ø
Ø otherwise the MUP pattern is left unassigned
Diversity-Based Hybrid Classifier Fusion
Ø The diversity-based hybrid classifier fusion scheme is a two-
stage process
Ø
Ø consists of two aggregators with a pre-stage classifier
selection module for each aggregator.

Ø The ensemble candidate classifiers selected for aggregation


are decided through assessing the degree of agreement
using the kappa statistic measure

 The diversity-based hybrid fusion scheme works as follows:

 First stage:

Ø classifiers selected for aggregation by the first aggregator are


those having the maximum degree of agreement
Ø The outputs of the classifiers are presented to the majority
voting aggregator.
Ø If all the classifiers state a decision that a MUP pattern is left
unassigned it stays unassigned.
 Second stage:

Ø This stage is used for those MUP patterns for which only half
or less than half of the ensemble classifiers in the first
stage specify a decision for a MUP pattern to be assigned to
the same MUPT class
Ø
Ø classifiers selected for aggregation at the second combiner
are those having a minimum degree of agreement
considering only the unassigned category

Ø The outputs of the classifiers are presented to the average
rule aggregator or the trainable aggregator represented by
Sugeno fuzzy integral aggregator.

Ø For each MUP pattern, the overall combined confidence


values representing the degree of membership in each
MUPT class are determined

Ø MUP pattern is assigned to the MUPT class for which its


determined overall combined confidence is the largest and
if its above the specified aggregator confidence threshold
set for that MUPT class.

Comparative Study
 The single classifier and multi-classifier approaches are
compared in terms of the difference between the correct
classification rate CCr% and error rate Er%.

 number of MUPs correctly classified
CCr = ×
100
 total number of MUPs detected


number of MUPs erroneously classified
 Er = × 100
number of MUPs assigned

 The base classifiers used for experimentation are

 1. four ACC classifiers e1,e2, e3, e4

 2. four AFNNC classifiers e5,e6,e7,e8

 3. four ANCCC classifiers e9,e10 ,e11 ,e12

 4. four ApCC classifiers e13 ,e14 ,e15 ,e16


Classifiers e1, e5, e9, e13 were fed with time-domain

first-order discrete derivative features

Classifiers e2, e6, e10 , e14 were fed with time-

domain first-order discrete derivative features

Classifiers e3, e7, e11 , e15 were fed with wavelet-

domain first order discrete derivative features

Classifiers e4, e8, e12 , e16 were fed with wavelet-

domain first-order discrete derivative features


Table .1 Mean and mean absolute deviation (MAD) of the difference between
correct classification rate CCr and error rate Er for the different single classifier
approaches across the three EMG signal data sets[9]
C la ssifie r In d e p e n d e n t R e la te d R e a l sig n a ls
sim u la te d sig n a ls sim u la te d sig n a ls
e1 8 1 .9 (4 .9 ) 7 5 .0 (2 .5 ) 7 8 .5 (0 .9 )

e2 83.9 (4.9) 76.5 (1.8) 72.0 (0.3)

e3 82.3 (4.1) 75.6 (1.6) 71.9 (1.4)

e4 84.7 (4.0) 76.4 (1.5) 67.1 (1.4)

e5 85.2 (2.5) 73.7 (1.0) 80.9 (0.9)

Best 90.4 (1.7) 80.7 (2.3) 77.5 (2.6)


single classifier e6
e7 83.1 (2.4) 73.3 (0.5) 73.5 (0.4)

e8 88.9 (1.5) 79.0 (1.8) 73.4 (2.4)

Average of 8 single 85.0 (3.3) 76.2 (1.6) 74.3 (1.2)


classifiers
Table .1 continues……..
e9 79.3 (3.2) 59.2 (0.4) 69.0 (1.3)

e10 80.6 (3.1) 54.6 (2.6) 63.0 (0.7)

e11 76.1 (2.7) 59.4 (1.0) 58.7 (0.4)

e12 77.7 (2.7) 56.0 (0.8) 49.5 (0.3)

e13 78.0 (3.4) 62.6 (0.7) 71.3 (0.4)

e14 77.8 (2.8) 59.4 (0.9) 66.1 (2.7)

e15 77.5 (2.9) 62.2 (0.1) 68.0 (1.5)

e16 77.6 (2.2) 58.9 (0.9) 56.6 (3.6)

Average of 16 single 81.5 (0.1) 67.7 (0.4) 68.5 (0.0)


classifiers
Table .2 Mean and mean absolute deviation (MAD) of the difference between correct
classification rate CCr and error rate Er for the different single classifier and multi-classifier
approaches across the three EMG signal data sets

cla ssifie r In d e p e n d e n t sim u la te d R e la te d R e a l sig n a ls


sig n a ls sim u la te d sig n a ls
S in g le cla ssifie rs

W e a ke st o f 8 S in g le 84.7 (4.0) 76.4 (1.5) 67.1 (1.4)


C la ssifie rs e 4
Weakest of 16 Single 77.7 (2.7) 56.0 (0.8) 49.5 (0.3)
Classifier e12
Best Single Classifier e6 90.4 (1.7) 80.7 (2.3) 77.5 (2.6)

Average of 8 Single 85.0 (3.3) 76.2 (1.6) 74.3 (1.2)


Classifiers
Average of 16 Single 81.5 (0.1) 67.7 (0.4) 68.5 (0.0)
Classifier
Table .2 continues……

O n e -S ta g e C la ssifie r Fu sio n [ 2 ]

Majority Voting (fixed of 86.0 (4.6) 79.2 (3.0) 77.3 (4.3)


8)
Average Fixed Rule (fixed 88.0 (2.5) 82.0 (0.7) 85.1 (1.2)
of 8)
Sugeno Fuzzy Integral 82.3 (2.7) 78.3 (1.9) 80.9 (2.9)
(fixed of 8)
Diversity-based One-Stage Classifier Fusion[2]

Majority Voting 6/8 87.6 (4.2) 80.1 (2.7) 78.8 (4.8)

Average Fixed Rule 6/8 88.5 (2.2) 82.1 (1.1) 84.9 (0.8)

Sugeno Fuzzy Integral 6/8 84.6 (2.4) 80.2 (1.1) 82.0 (0.8)
 Table .2 continues…………..
H yb rid C la ssifie r Fu sio n [ 7 ]

A M V A FR ( fixe d o f 6 ) 9 1 .8 (1 . 8 ) 8 4 . 6 (1 . 3 ) 8 2 .7 (2 .5 )

A M V S FI ( fixe d o f 6 ) 9 1 .8 (1 . 8 ) 8 4 . 6 (1 . 3 ) 8 2 .5 (1 .7 )

D ive rsity -b a se d H yb rid C la ssifie r Fu sio n [ 6 ]

A D M V A FR – 6 / 8 9 1 .6 (1 . 8 ) 8 4 . 4 (0 . 7 ) 8 5 .5 (0 .9 )

A D M V S FI – 6 / 8 9 1 .2 (1 . 8 ) 8 4 . 0 (0 . 8 ) 8 5 .2 (0 .9 )

A D M V A FR – 6 / 1 6 9 0 .0 (3 . 2 ) 8 3 . 2 (0 . 7 ) 8 3 .7 (0 .6 )

A D M V S FI – 6 / 1 6 8 9 .6 (3 . 3 ) 8 2 . 5 (0 . 8 ) 8 2 .8 (0 .4 )

AMVAFR, ADMVAFR – stands for Adaptive (or Diversity-based) Majority Voting


with Average Fixed Rule hybrid classifier fusion scheme, respectively.
AMVSFI, ADMVSFI – stands for Adaptive (or Diversity-based) Majority Voting with
Sugeno Fuzzy Integral one-stage classifier fusion scheme, respectively.
Conclusion
Ø one-stage aggregator classifier fusion and its diversity-based
variant schemes have performance better than the average
performance of the base classifiers
Ø Better than the performance of the best base classifier except
across the independent simulated signals.
Ø The hybrid classifier fusion and its diversity-based variant
approaches have performances that not only exceed the
performance of any of the base classifiers forming the
ensemble but also reduced classification errors for all data
sets studied
References
1. Rasheed S, Stashuk D and Kamel M (2008) Fusion of multiple
classifiers for motor unit potential sorting. Biomedical
Signal Processing and Control, 3(3):229–243
2. Rasheed S, Stashuk D and Kamel M (2008) Diversity-based
combination of non-parametric classifiers for EMG signal
decomposition. Pattern Analysis & Applications, 11:385–
408
3. Stashuk D W (2001) EMG signal decomposition: how can it be
accomplished and used? Journal of
Electromyography and Kinesiology, 11:151–173
4. McGill K C, Cummins K and Dorfman L J (1985) Automatic
decomposition of the clinical electromyogram. IEEE
Transactions on Biomedical Engineering, 32(7):470–477
5.
6.
5. Nikolic M, Sørensen J A, Dahl K et al. (1997) Detailed analysis of
motor unit activity. In Proceedings of the 19th Annual
International Conference of the IEEE Engineering in Medicine
and Biology Society Conference, 1257–1260
6. Sarbast Rasheed, Daniel W. Stashuk, Mohamed S.
Kamel(2002) Integrating Heterogeneous Classifier
Ensembles for EMG Signal Decomposition Based on
Classifier Agreement IEEE transactions on information
technology in biomedicine, 1(11):1-17
7. Rasheed S, Stashuk D and Kamel M(2007) A hybrid classifier
fusion approach for motor unit potential classification during
EMG signal decomposition. IEEE Transactions on Biomedical
Engineering, 54(9):1715–1721
8. Roberto Merletti, Philip Parker (2004) Electromyography:
Physiology, Engineering, and Non-Invasive Applications.
Wiley-IEEE Press, 1st edition
9. Amine Nait-ali (2009) Advanced Biosignal Processing . Springer,
1st edition
 

5.

You might also like