Comparison of Parametric and Non-Parametric Spectral Estimation

Methods for Automatic Tremor Detection against Clinical Evaluation

O. Martinez-Manzanera, J. H Elting, J. W. van der Hoeven and N. M. Maurits

Department of Neurology, University Medical Center Groningen (UMCG), University of Groningen,

Groningen, The Netherlands

Keywords: Tremor, Accelerometry, Biomedical Signal Processing, Psychogenic Tremor.

Abstract: Psychogenic tremor (PT) is a condition where the person affected suffers from tremor with variable

characteristics that can make it difficult to diagnose. To help in the diagnosis an automatic tremor detection

method applied to long-term kinematic recordings is proposed. The recorded signal is divided in segments

which are analyzed and classified automatically as “tremor” or “no tremor”. The classification is done

according to the location of the dominant frequency of the power spectral density (PSD) of each segment.

Different PSD estimation methods are explored to determine the optimum method for segments of short

length. The performance of each method is compared against a clinical assessment of tremor.

1 INTRODUCTION

Psychogenic movement disorders (PMD) are

characterized by the presence of abnormal

movements that cannot be attributed to an organic

neurological disorder and are considered to be

psychologically mediated (Kranick et al., 2011). PT

is the most common form of PMD (Jankovic et al.,

2006). The diagnosis of a movement disorder is

mainly a clinical process where patients are

interviewed and undergo clinical observation. For a

diagnosis of PT the movement characteristics must

be incongruent with any organic tremor and the

tremor may not be fully explained by an organic

disease (Jankovic et al., 2006). PT often shows

variable amplitude and frequency, suggestibility and

entrainment and it changes character or is

suppressed when the patient is distracted (Kenney et

al., 2007). While these features are useful clues the

certainty of a final diagnosis largely depends on the

experience of the examiner (Jankovic et al., 2006).

These features can be quantified using

electromyographic (EMG) (O’Suilleabhain and

Matsumoto, 1998) or kinematic recordings (Salarian

and Russmann, 2007). A clinician can detect

episodes of tremor in these recordings by assessing

the signals qualitatively (by visual inspection) and

quantitatively (by using PSD estimation).

In patients with PT (where tremor symptoms are

variable) long term recordings could be beneficial

for accurate diagnosis.

Kinematic recordings have been used to assess

tremor duration (Pareés et al., 2012). The presence

of tremor was compared with a self-report from the

patient from the same period resulting in an

overestimation of tremor by the patient. Detailed

analysis of the signals would require a large time

investment of a clinician. In this study we therefore

compare several automatic tremor detection methods

based on PSD estimation applied to long-term

accelerometry recordings. The goal is to evaluate the

accuracy of the automatic detection methods in

identifying tremor compared to a clinician’s

assessment.

2 METHODS

Kinematic recordings obtained from the diagnostic

work-up of 15 patients with different disorders (12

males, 3 females, mean age=68.2, standard

deviation=9.7 years, 5 parkinsonism, 4 essential

tremor, 2 enhanced physiological tremor, 2 PT, 1

dystonia, 1 ataxia) at UMCG were used in this study.

The signal obtained from a uniaxial accelerometer

placed on the dorsal side of the hand of the most

affected limb was used for analysis. Methods to

estimate the PSD of a signal can be divided in

Martinez-Manzanera O., H. Elting J., W. van der Hoeven J. and Maurits N..

Comparison of Parametric and Non-Parametric Spectral Estimation Methods for Automatic Tremor Detection against Clinical Evaluation.

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

parametric (the signal is represented by a model plus

noise) and non-parametric (no model assumption).

For this study both methods were used to determine

the dominant frequency in the accelerometer signal.

2.1 Non-Parametric Methods

Two methods to estimate the PSD based on the Fast

Fourier Transform (FFT) were used. The modified

periodogram (Hann window) was used since it is

one of the less computationally intensive methods.

Also the Welch method (a very popular method to

reduce the variance of the PSD) (Welch, 1967) was

used (2, 3 or 8 windows with 50% overlap). For both

methods, the dominant frequency was selected as the

frequency of the peak with the highest amplitude

within the 0-20 Hz frequency band.

2.2 Parametric Methods

We selected an autoregressive (AR) model because

it is suited for estimating spectra that are

characterized by their peaks, making it the most

appropriate method for the analysis of tremor data

(Spyers-Ashby et al., 1998). AR modelling of a time

series is based on the assumption that each value of

the series can be predicted as a weighted sum of the

previous values (and posterior values for the Burg

method) of the same series plus an error term

(Takalo et al., 2005). The number of values used in

the prediction is the model order. In this study we

used the Akaike criterion (Akaike, 1974) and four

criteria based on the evaluation of the frequency

responses from the 2nd order filter to the 50th order

(Fig. 1). For each response we determined the

frequency of the peak with the highest amplitude

(HighAmp) and the frequency of the peak with the

highest frequency (HighFreq; up to 20 Hz) of all the

responses. ModeHighAmp is the most occurring

frequency in all the responses when HighAmp is

Figure 1: Frequency response of models (2

to 50

order)

obtained from one segment and their dominant frequencies

according to five criteria (see text).

searched. ModeHighFreq is the most occurring

frequency in all the responses when HighFreq is

searched (Fig. 1).

2.3 Tremor Classification

Two experienced clinicians evaluated the

accelerometer signal per segments of 4 seconds.

They visually assessed the signal and used a PSD

estimation tool (based on FFT) to determine the

dominant frequency of a specific segment, when

needed. The clinicians classified the segment as

tremor when the signal was dominantly sinusoidal

and the frequency was consistent with a tremor. In

the automatic methods, the criterion for a tremor

segment was a dominant frequency between 2.5-10

Hz. To compare the methods the F1-score was used.

It is an evaluation metric that combines the positive

predictive value and the sensitivity in a single

number.

3 RESULTS

Compared to the postulated gold standard (1

clinician), the evaluation by the 2

clinician resulted

in a better F1 score than the automatic methods. The

performance of the automatic methods is similar,

only HighFreq has very poor specificity due to its

tendency to localize high frequency peaks (Fig. 2).

The results vary depending on the patient suggesting

that some conditions are more difficult to assess

using solely the accelerometer signal.

Figure 2: F1-score for each method (applied to segments

of length 1, 2 and 4 s). Results are plotted with a grey

circle, the mean and 95% confidence intervals are shown

in red.

4 DISCUSSION

The intervals in which tremor is found are similar

for almost all automatic methods (Fig. 3) These

similarities may indicate that there are signal

segments that contain characteristics that made them

be classified as tremor by the automatic methods but

not by the clinician (e.g. the clinician may classify a

segment as tremor using a more constrained

frequency bandwidth than 2.5-10 Hz,

unconsciously). Constraining this bandwidth in the

automatic methods would probably improve the

results as Fig. 3 (top) suggests. This would result in

a higher probability that the segment classified as

tremor by the automatic methods actually

corresponds to a true episode of tremor. We propose

that in long term recordings having a trade-off

between a high positive predictive value at the

expense of an acceptable level of sensitivity is

beneficial, since the detection of false positive

tremor segments would be avoided and enough

tremor segments could still be used. Since peak

detection might not be the most accurate method to

identify tremor, other methods (e.g. ratio of tremor

area under the curve in frequency response) will be

studied in the future.

The method selected as gold standard may not be

optimal either, because it concerns a subjective

interpretation. Moreover, in our current approach,

neither the clinicians, nor the automatic methods

incorporated other information as used in daily

routine to diagnose tremor (EMG, video recordings).

REFERENCES

Akaike, H., 1974. A new look at the statistical model

identification. Automatic Control, IEEE Transactions

Jankovic, J., Vuong, K. D. & Thomas, M., 2006.

Psychogenic tremor: long-term outcome. CNS

spectrums, 11(7), pp.501–8.

Kenney, C. et al., 2007. Distinguishing psychogenic and

essential tremor. Journal of the neurological sciences,

263(1-2), pp.94–9.

Kranick, S., Gorrindo, T. & Hallett, M., 2011.

Psychogenic movement disorders and motor

conversion: a roadmap for collaboration between

neurology and psychiatry. Psychosomatics, 52(2)

O’Suilleabhain, P. E. & Matsumoto, J. Y., 1998. Time-

frequency analysis of tremors. Brain : a journal of

neurology, 121 ( Pt 1, pp.2127–34).

Salarian, A. & Russmann, H., 2007. Quantification of

tremor and bradykinesia in Parkinson’s disease using a

novel ambulatory monitoring system. Biomedical …,

54(2), pp.313–22.

Spyers-Ashby, J. M., Bain, P. G. & Roberts, S. J., 1998. A

comparison of fast Fourier transform (FFT) and

autoregressive (AR) spectral estimation techniques for

the analysis of tremor data. Journal of neuroscience

methods, 83(1), pp.35–43.

Pareés et al., 2012. Believing is perceiving: Mismatch

between self-report and actigraphy in psychogenic

Figure 3: Results for patient 1 (Psychogenic tremor). Top: Dominant frequencies for each 1 s segment (HighAmp method).

Yellow band: range for tremor detection. Bottom: Tremor detection by each method for each segment.

tremor. Brain 135;117-123

Takalo, R., Hytti, H. & Ihalainen, H., 2005. Tutorial on

univariate autoregressive spectral analysis. Journal of

clinical monitoring and computing, 19(6), pp.401–10.

Welch, P., 1967. The use of fast Fourier transform for the

estimation of power spectra: a method based on time

averaging over short, modified periodograms. Audio

and Electroacoustics, IEEE Transactions on.