Content uploaded by Franz Marxreiter
Author content
All content in this area was uploaded by Franz Marxreiter on May 09, 2019
Content may be subject to copyright.
Abstract— In this study, we intended to differentiate patients
with essential tremor (ET) from tremor dominant Parkinson
disease (PD). Accelerometer and electromyographic signals of
hand movement from standardized upper extremity movement
tests (resting, holding, carrying weight) were extracted from 13
PD and 11 ET patients. The signals were filtered to remove
noise and non-tremor high frequency components. A set of
statistical features was then extracted from the discrete wavelet
transformation of the signals. Principal component analysis
was utilized to reduce dimensionality of the feature space.
Classification was performed using support vector machines.
We evaluated the proposed method using leave one out cross
validation and we report overall accuracy of the classification.
With this method, it was possible to discriminate 12/13 PD
patients from 8/11 patients with ET with an overall accuracy of
83%. In order to individualize this finding for clinical
application we generated a posterior probability for the test
result of each patient and compared the misclassified patients,
or low probability scores to available clinical follow up
information for individual cases. This non-standardized post
hoc analysis revealed that not only the technical accuracy but
also the clinical accuracy limited the overall classification rate.
We show that, in addition to the successful isolation of
diagnostic features, longitudinal and larger sized validation is
needed in order to prove clinical applicability.
I. INTRODUCTION
Tremor is defined as rhythmic, oscillating and involuntary
movements and often involves the upper extremities.
Differential diagnosis of tremor commonly comprises
essential tremor (ET) and Parkinson disease (PD) amongst
other rare forms. The prevalence of ET is high with up to 5%
with people over 65 years old [1]. PD has an overall
prevalence of approximately 0.5 increasing up to 2% with
age [2]. Tremor is one of the core symptoms in PD patients
and the initial clinical presentation of the tremor dominant
PD phenotype.
* These authors contributed equally to this work.
This study is supported by the FAU Emerging Fields Initiative
(EFIMoves, 2_Med_03).
N. Haji Ghassemi, P. Kugler, C. Pasluosta and B. M. Eskofier are with
the Digital Sports Group, Pattern Recognition Lab, Department of
Computer Science, Friedrich Alexander University Erlangen-Nürnberg
(FAU), Erlangen, Germany.
F. Marxreiter, J. Schlachetzki, C. Pasluosta, and J. Klucken are with the
Department of Molecular Neurology, University Hospital Erlangen,
Erlangen, Germany.
Axel Schramm is with the Department of Neurology, University
Hospital Erlangen, Erlangen, Germany.
Prototypically, tremor in ET features a symmetric postural
and kinetic tremor of the arms, whereas tremor in PD is
characterized by an often unilateral resting and postural
tremor. Differential diagnosis however, especially in early
disease stages remains problematic because tremor in PD
may not only occur at rest, but also at posture and/or during
action. Further, tremor at rest is not pathognomonic for PD
and has been also observed in ET. Due to these overlapping
symptoms, misdiagnosis of ET and PD tremor may occur in
20-30% of the cases [3]. In light of the difference in
prognosis as well as current and evolving treatment options
for PD and ET, an accurate and reliable diagnosis is urgently
needed. Despite these challenges, the diagnostic evaluation
is highly dependent on the experience of the clinician and
typically includes medical history and physical examination.
Neuroimaging [4] has been considered as one potential
diagnostic option to discriminate of PD and ET. However,
ability, costs, and invasiveness of this diagnostic option have
to been taken into consideration [5].
There is a growing interest in methods based on
accelerometer and surface electromyography (EMG)
electrodes, since they are readily available, non-invasive and
cost-efficient diagnostic tools. Combination of these signals
can be used to detect frequency and muscle activity [6].
Several studies differentiated ET from PD based on
accelerometer and EMG signals [6] [7]. These studies yield
reasonable differentiation between PD and ET and overall
classification accuracy was used as the main metric for
evaluation of performance. However, high classification
accuracy is not enough to support clinical decision-making.
A physician should be able to clinically validate
classification results. Thus, classification procedure should
be translated into an individual report for the patients. In this
work, we propose using posterior probability to assess the
classification result in an individual manner. This meta
information is essential for us to integrate our method in
clinical assessment. We utilized a SVM-based classifier, by
which we can compute the posterior probability of both
diseases. To the best of our knowledge, this is the first
attempt to develop such a method for differential diagnosis.
II. METHODOLOGY
Our classification system is composed of the following
steps: (1) signal preprocessing, (2) discrete wavelet
transform, (3) feature extraction, (4) dimensionality
reduction, and (5) classification. The schematic diagram of
the system is shown in Fig 1. We implemented the entire
procedure in Matlab R2015a (MathWorks).
Combined Accelerometer and EMG Analysis to Differentiate
Essential Tremor from Parkinson’s Disease
Nooshin Haji Ghassemi*, Franz Marxreiter*, Cristian F. Pasluosta, Patrick Kugler, Johannes
Schlachetzki, Axel Schramm, Bjoern M. Eskofier, Member, IEEE, EMBS, Jochen Klucken
Figure 1. Block diagram of the automated classification
A. Data Acquisition
Data acquisition was performed as part of a standard
diagnostic tremor test in the neurophysiology lab of the
Department of Neurology, University Hospital Erlangen,
Erlangen, Germany. Typically, tremor frequencies and
synchronization of agonist/antagonists are reported providing
only a limited amount of information from the raw data
recorded. Subjects were comfortably seated and placed their
arms in a resting position on their legs. Prior to electrode
placement, the skin was cleaned by cotton and ethanol. Two
calibrated accelerometers were placed on the dorsal side of
both hands of the subjects. Bipolar Ag/AgCl surface EMG
electrodes were placed on the extensor and flexor muscles of
the left and right forearm by an electrophysician. A total of
six channels were recorded, two accelerometer signals and
four surface EMG signals. All signals were recorded at 1000
Hz using a Schwarzer Topas EMG system, Natus, USA.
Three 30-seconds tests were performed for each patient, the
first with the arms fully relaxed and rested on the legs (rest),
the second with their arms and hands extended in a horizontal
position (hold) and the third in the same position with
weights of 1 kg attached to the forearm (weight). To discard
transitions from the recorded data, we considered only the
signal from 10 to 25 seconds.
13 PD patients (tremor dominant forms) and 11 ET
patients were included for this study. For our experimental
procedures, we followed the declaration of Helsinki 1975, as
revised in 2000. Initial diagnosis was made by a movement
disorder specialist prior to tremor analysis. If possible,
patients received clinical follow up examinations to confirm
or change the initial diagnosis. PD and ET were diagnosed
according to consensus criteria of the German Society of
Neurology. Consensus criteria for PD are similar to the UK
PDS Brain bank criteria for diagnosis of PD [3] Consensus
criteria for ET are based on the consensus criteria of the
movement disorders society [9] Patients’ demographics and
clinical characteristics are summarized in Tab I. PD severity
of patients was staged using the Hoehn & Yahr scale and
motor performance was clinically evaluated using the
UPDRS-III rating.
TABLE I. PATIENT CHARACTERISTICS.
B. Preprocessing and Filtering
The recording physician visually inspected both EMG and
accelerometer signals during data acquisition. Measurements
with bad signal quality were either repeated or not included
in the study. To remove non-tremor related high frequency
components, the accelerometer signals and EMG signals
were filtered with a Butterworth low pass filter with a cutoff
frequency of 70 Hz. The EMG signals were additionally
high pass filtered with a cutoff frequency of 20 Hz and DC-
rectified to amplify the visibility of the tremor bursts in the
signals.
C. Wavelet-based Feature Extraction
EMG signal is non-stationary, meaning that the
characteristic of the signal changes over time. Discrete
wavelet transform (DWT) [10] is widely used for analysis of
non-stationary signals. DWT decomposes a signal to
different level of coefficients corresponding to different
frequency bands in a way that the coefficients include all
information of the original signal. Each level of coefficients
has different time-frequency resolution. In this work, we
used Haar mother wavelet, since it has been widely used for
wavelet analysis of EMG signal. The signals were
decomposed in 10 levels in order to have an effective feature
extraction from each coefficient in the next step.
Four EMG and two accelerometer signals were
decomposed by DWT method. Then, a set of standard
statistical features was extracted from the coefficients: mean,
standard deviation, skewness, kurtosis, entropy, energy, root
mean square, and mean absolute value. In the majority of PD
and ET patients, one hand is more affected by the disease
than the other hand. Feature extraction is performed based
on the most affected side of the patients rather than left or
right hand side, in order to obtain more descriptive features.
D. Dimensionality Reduction
Dimensionality reduction was performed to reduce the
dimensionality of the feature vector and avoid overfitting in
the classification step. This is particularly important in our
study since the number of features (524 for each test) is
much higher than the number of subjects (24). Feature
reduction was achieved by using principal component
analysis (PCA) [11]. PCA projects the feature space into
principal components in the direction of maximum variance.
PD
ET
Number of Patients
13
11
Age at examination(Range)
67 ± 11 (45-88)
66 ± 13 (43-79)
Gender (Male/Female)
7/6
4/7
Disease duration
4 ± 4,6
13 ± 14,6
UPDRS III
17 ± 9,7
H+Y
2 ± 0,9
No
Init.
Diag.
Classific.
result
Probability
Clinical
confirm.
ET
PD
1
'ET'
'ET'
90,7%
9,3%
confirmed
2
'ET'
'ET'
61,1%
38,9%
confirmed
3
'ET'
'ET'
54,3%
45,7%
OTHER
4
'ET'
'ET'
75,5%
24,5%
confirmed
5
'ET'
'ET'
76,0%
24,0%
confirmed
6
'ET'
'ET'
63,0%
37,0%
no follow up
7
'ET'
'ET'
50,0%
50,0%
confirmed
8
'ET'
'ET'
95,6%
4,4%
no follow up
9
'ET'
'PD'
3,6%
96,4%
OTHER
10
'ET'
'PD'
43,2%
56,8%
no follow up
11
'ET'
'PD'
46,8%
53,2%
no follow up
12
'PD'
'ET'
83,0%
17,0%
confirmed
13
'PD'
'PD'
43,3%
56,7%
no follow up
14
'PD'
'PD'
31,4%
68,6%
confirmed
15
'PD'
'PD'
47,4%
52,7%
confirmed
16
'PD'
'PD'
46,4%
53,6%
confirmed
17
'PD'
'PD'
38,4%
61,6%
no follow up
18
'PD'
'PD'
40,4%
59,7%
confirmed
19
'PD'
'PD'
48,2%
51,8%
confirmed
20
'PD'
'PD'
34,5%
65,5%
confirmed
21
'PD'
'PD'
48,2%
51,9%
confirmed
22
'PD'
'PD'
42,1%
57,9%
confirmed
23
'PD'
'PD'
34,1%
65,9%
no follow up
24
'PD'
'PD'
39,7%
60,3%
confirmed
These new components compose a feature space with
reduced dimensionality. In our experiment, the number of
components was optimized empirically to achieve the
highest classification performance.
E. Classification
We trained a binary Support Vector Machines (SVM)
classifier [12] to distinguish between two classes: ET and
PD. Two kernels, linear and radial basis function (RBF)
were examined in this work. SVM has a cost parameter,
which controls number of misclassification of training
examples. The RBF kernel has an additional parameter,
gamma, which controls how far the influence of a single
training example reaches. A grid-search was employed as a
method of model selection to adjust the SVM parameters.
Since performing a complete grid-search is very time
consuming, it was applied in two stages via coarse grid, and
then, fine grid. In the coarse grid, the range of cost
parameter was {0.001, 0.01, 0.1, 1, 10, 15, 20, 50, 100,
1000} and the range of gamma was {0.003, 0.03, 0.3, 3, 9,
15, 20}. In the fine, the parameters were examined in a range
of ±5%, ±10%, ±15%, and ±20% of their selected values.
The input feature vector was normalized to zero mean and
unit standard deviation.
As previously described, we considered the three tests,
rest, hold and weight separately and trained a classifier for
each test. From clinical point of view, it is important to
investigate which of these tests yield a better differentiation
between ET and PD.
F. Evaluation
Accuracy was computed for evaluating the performance of
our method, meaning the rate of correctly classified patients.
The evaluation was implemented using leave one out cross-
validation (LOOCV). LOOCV was chosen due to the small
number of data. Finding principal components and SVM
parameters were performed for each training fold and the
result was applied for the test subject. In addition, we
computed the posterior probability of the classification. The
SVM algorithm not only predicts the class of a data but can
also report the posterior probability as an indicator of
certainty of the classification. The posterior probability is a
valuable piece of information in clinical applications since it
allows for an individual result. It provides a measure of
similarity for each patient assessed to the reference group.
This information can complement the diagnostic workup for
the clinician.
III. RESULTS
Tab. II shows the results for our classification method
optimized for the number of PCA components and the
parameters of SVM and the kernels. We analyzed each test
of rest, hold and weight separately. The best performance
was achieved for the weight test. Tab. III presents individual
classification result for each patient and the posterior
probabilities for each disease.
In order to better understand the misclassified patients and
to associate different probability levels, we obtained
additional clinical information from clinical records or
follow up visits if available in a non-standardized form to
confirm the initial diagnose that was used in the
classification experiments and to evaluate if the clinical
heterogeneity of the small patient population might affect
the classification outcome.
Table II. CLASSIFICATION ACCURACY RESULT
IV. DISCUSSION
The preliminary results of this study suggest that our method
is effective to discriminate patients with ET from PD
patients. The best discrimination between ET and PD was
achieved when there was a load in the subjects’ hands. This
finding underlines the ability of the diagnostic test, the
sensor-paradigm and the algorithms used to identify
characteristic differences between the two tremor forms.
However, the classification accuracy is not sufficient to
identify each patient with the reference group correctly,
thereby limiting the individual diagnostic application.
Therefore, we introduced a probability score for each patient
and compared the resulting similarity levels with the
available clinical information, in particular, for the
misclassified patients.
TABLE III. INDIVIDUAL CLASSIFICATION RESULT and
CLINICAL CONFIRMATION
Tests
No. PCA
Components
SVM Parameters
Accuracy
Rest
3
Linear kernel
cost = 0.001
79%
Hold
3
RBF kernel
cost = 14.25
gamma = 0.003
75%
Weight
3
RBF kernel
cost = 10
gamma = 0.3
83%
The accuracy of the clinical diagnosis “ET” is typically
lower than “PD”, since the presence of other PD specific
motor symptoms in addition to the tremor typically allow for
higher accuracy for PD. In this work, three patients with
initial diagnosis of ET (patient 9-11) and one PD patient
(patient 12) were misclassified. Patient 9 is an important
index patient for this clinical validation procedure since the
probability for ET was extremely low (3,6%). Importantly,
in the follow-up visits it became evident, that the initial
clinical diagnosis “ET” could not be confirmed. In fact, the
diagnosis was changed to cervical dystonia with irregular
tremor. For patients 10 and 11 the probability “ET” was
above 40%, but did not reach the 50% classification
boundary. Further, patient 7 barely reached 50% probability
for “ET”. Unfortunately, no follow up visits were
documented for these patients, thus the clinical confirmation
of the initial diagnosis could not be performed. Likewise, the
diagnosis of ET for patient 3 was changed in follow-up
examinations. Even though this patient was classified as ET,
the probability was only 54,3% suggestion that probabilities
around 50% generate another level of uncertainty for
individual test results. The diagnosis of the misclassified PD
patient 13 was clinically confirmed. Medical history did not
reveal any signs that might explain the misclassification
from the clinical point of view except that the patient was at
the very first stage of PD (disease duration was 0 years),
suggesting that at very early stage of tremor-dominant PD
the classification accuracy might be limited.
Any pattern recognition method is bound to limitations. A
major limitation of proof-of-concept studies is the small
number of patients. Larger data set brings better
generalization and as a result better classification result.
Besides, our study also revealed that the validity of the
ground truth is limited by the diagnostic accuracy of clinical
examinations of PD as well as ET at a given time point,
especially in early disease stages is not perfect. Therefore,
careful reevaluation and the assessment of the response to
treatment are often needed to finally decide upon the final
diagnosis.
Despite of all limitations, our method brought good
classification accuracy. Furthermore supporting that the
study design and analysis enables the identification of
different tremor forms. In this regard, the concept of
probability definition is a first attempt to translate successful
classification paradigms into individualized results that can
complement the diagnostic workup in clinical settings. It is
foreseeable that with increasing numbers of future patients
that undergo the instrumented tremor testing the reference
group can be increased and refined. This may also make lead
to a better classification accuracy for patients in early
disease stages. Moreover the acceptance of treating
physicians and patients of the individual probability results
can be evaluated in longitudinal studies.
V. CONCLUSION
In this work, accelerometer and EMG signals were analyzed
to differentiate ET from PD using pattern recognition
methods. With the proposed method we were able to
discriminate ET from PD patient with an overall accuracy of
83%. Additionally, we propose the posterior probability of
the classification outcome as a clinical indicator of patients
presenting disease-specific symptoms. The clinical
validation revealed that studies using clinically confirmed
and small sized patient cohorts share the risk that the ground
truth or gold standard for the classification experiment might
be also affected by the clinical accuracy of the diagnoses.
Thus, larger study cohorts, better clinical validation (e.g.
neuroimaging, etc.) and/or standardized follow-up
paradigms are required to clinically validate instrumented
tests for diagnostic workup in the future. Our findings also
underline that studies aiming at clinical translation of
instrumented movement analysis have to include technical
and clinical accuracy considerations in the study design to
ultimately prove clinical applicability.
ACKNOWLEDGMENT
We would like to thank all participants of this study as well
as the staff helping with data collection. First author
acknowledges financial support from the Bavarian Research
Foundation.
VI. BIBLIOGRAPHY
[1]
J. Benito-León and E. D. Louis, "Essential tremor: emerging views of
a common disorder ," Nature Clinical Practice Neurology, vol. 2, no.
12, pp. 666–678, 2006.
[2]
D. Hirtz, D. J. Thurman , K. Gwinn-Hardy, M. Mohamed, A. R.
Chaudhuri, and R. Zalutsky, "How common are the "common"
neurologic disorders?," Neurology, vol. 68, no. 5, pp. 326-337, 2007.
[3]
A. J. Hughes, S. E. Daniel, L. Kilford, and A. J. Lees, "Accuracy of
clinical diagnosis of idiopathic Parkinson's disease: a clinico-
pathological study of 100 cases," Journal of Neurology, Neurosurgery,
and Psychiatry, vol. 55, no. 3, pp. 181–184, 1992.
[4]
L. Passamonti, A. Cerasa, and Q. Aldo, "Neuroimaging of Essential
Tremor: What is the Evidence for Cerebellar Involvement?," Tremor
Other Hyperkinet Mov, vol. 2, 2012.
[5]
K. Badiavas, E. Molyvda, I. Iakovou, M. Tsolaki, K. Psarrakos, and
N. Karatzas, "SPECT imaging evaluation in movement disorders: far
beyond visual assessment," vol. 38, no. 4, pp. 764-773, 2011.
[6]
J. Timmer, M. Lauk, and G. Deuschl, "Quantitative analysis of tremor
time series," vol. 101, no. 5, pp. 461-468, 1996.
[7]
A. Hossen, M. Muthuraman, Z. Al-Hakim, J. Raethjen, G. Deuschl,
and U. Heute, "Discrimination of Parkinsonian tremor from essential
tremor using statistical signal characterization of the spectrum of
accelerometer signal," Bio-medical materials and engineering, vol. 23,
no. 6, pp. 513-531, 2013.
[8]
J. Raethjen, M. Lauk, B. Koster, U. Fietzek, L. Friege, J. Timmer, C.
H. Lucking, and G. Deuschl, "Tremor analysis in two normal cohorts,"
Clin Neurophysiol, vol. 115, no. 9, pp. 2151-2156, 2004.
[9]
G. Deuschl, P. Bain, and M. Brin, "Consensus statement of the
Movement Disorder Society on Tremor. Ad Hoc Scientific
Committee.," Movement Disorders: Official Journal of the Movement
Disorder Society, vol. 13, no. 3, pp. 2–23, 1998.
[10]
C. S. Burrus, R. A. Gopinath, and H. Guo, "Introduction to Wavelets
and Wavelet Transforms," A Primer, 1997.
[11]
S. Wold, K. Esbensen, and P. Geladi, "Principal component analysis,"
Chemometrics and Intelligent Laboratory Systems, vol. 2, no. 1-3, pp.
37-52, 1987.
[12]
M. A. Hearst, S. T. Dumais, E. Osman, J. Platt, and B. Scholkopf,
"Support vector machines," IEEE Intelligent Systems and their
Applications, vol. 13, no. 4, pp. 18 - 28, 1998.