Sensitivity and specificity of a portable system measuring postural tremor

Post on 16-Sep-2016

213 views 0 download

Transcript of Sensitivity and specificity of a portable system measuring postural tremor

Neurotoxicology and Teratology, Vol. 19, No. 2, pp. 95-104,1997 Copyright 0 1997 Elsevier Science Inc.

Printkdk the USA. All rights reserved 0892.0362/97 $17.00 + .OO

PII SO892-0362(96)00179-l

Sensitivity and Specificity of a Portable System Measuring Postural Tremor

RODERICK EDWARDS AND ANNE BEUTER

Dkpartement de Kinanthropologie, Universit6 du Qutbec Li Mont&al, C.P. 8888, succ. Centre-ville, Montrdal, PQ, Canada H3C 3P8

Received 2 May 1996; Accepted 31 July 1996

EDWARDS, R. AND A. BEUTER. Sensitivity and specificity of a portable system measuring postural tremor. NEURO- TOXICOL TERATOL 19(2) 95-104, 1997.-A portable, accelerometric system measuring tremor was evaluated. That is, the validity and consistency of measurements as well as its ability to discriminate pathologic from physiological tremor were investigated. Control subjects and patients with Parkinson’s disease were tested with this portable system and with an inde- pendent system that gave precise displacement data using lasers. It was found that amplitude of postural tremor as measured by the two systems differed significantly, but further investigation revealed that this difference was due 1) to the difference between amplitude of acceleration and amplitude of displacement, and 2) to changes in tremor over the time between tests, rather than to any inaccuracy or unreliability in the portable system. The other characteristics of tremor reported by the por- table system were also valid and reasonably reliable in test-retest experiments, with the exception of the “harmonic index,” which proved less stable. Most of the reported characteristics were distributed differently for the control group and for the patients with Parkinson’s disease, but the large overlaps between distributions would make diagnosis difficult when tremor is not very pronounced. These results suggest that until better discriminating measures of tremor are available, tremor tests should be repeated and combined with other tests of motor function. 0 1997 Elsevier Science Inc.

Tremor quantification Neurological deficits Portable system

BEHAVIORAL testing of occupational populations based on neuropsychology and experimental psychology started in the 1960s and has continuously evolved since. For example, the number of tests has considerably expanded and tests have become computerized (1). Recently researchers have become aware that subtle changes in neuromotor control might repre- sent a pathognomonic sign of neurotoxicity. As a result, a few valid and reliable neuromotor test batteries have been devel- oped. One of them, composed of the Tremor Analysis Test System and the Coordination Ability Test System developed in Denmark in 1992 by Danish Product Development (DPD) is now available commercially to quantify exposure to neuro- toxic agents such as metals, solvents, or pesticides. A sample of 61 normal subjects was tested with the Tremor Analysis Test System for standardization by Dr. Sigurd Mikkelsen at the Clinic of Occupational Health, Copenhagen County Hos- pital [(4) p. 311. Typical items included in neuromotor evalua- tion include measures of reaction time, finger tapping, rapid alternating movements, tremor, hand dexterity, and postural

-

sway. Traditionally, neuromotor test batteries have included subsets of these items. Among these tests, tremor is probably one of the most difficult movements to record and analyse ad- equately (3). This is an especially important issue because it appears that tremor is relatively vulnerable to neurotoxic in- fluences but the recorded changes are usually subtle and the variability in the normal population is high.

Tremor is defined as involuntary and continuous oscilla- tory movements of body parts such as extremities, jaws, eyes, or head that are normally present in healthy subjects (i.e., physiological tremor) but may change as a function of fear, anxiety, cold, anger, or hypoglycaemia (i.e., enhanced physio- logical tremor) (16). Sometimes enhanced physiological tremor is perfectly normal (it is basically physiological tremor with a larger amplitude); however, at other times it may re- flect metabolic disturbances or pathologies of the nervous sys- tem (8).

Like all movement abnormalities, physiological tremor tends to fluctuate over time in a significant way that makes it

Kequests tor reprints should be addressed to Anne Beuter, DCpartement de Kinanthropologie, UniversitC du QuCbec B MontrCal, C.P. 8888, succ. Centre-ville, Montrtal, PQ, Canada H3C 3P8. Fax: (514) 987-6616.

9.5

96

harder to classify in normal and pathological types (8,9,11,13). Authors have attempted to classify tremor by aetiology, by behavioral characteristics, by frequency, by electromyogra- phy, by pharmacological response, or by origin (i.e., central vs. peripheral). Other tremors defined as psychogenic (14) psychotic (lo), and orthostatic (i.e., rapid, irregular, and asyn- chronous) (18) are usually excluded from these classifications (8). These last authors indicate that today no single classifica- tion scheme is available to be used diagnostically. One addi- tional difficulty is that other movement abnormalities such as clonus. chorea, or tics may also have, like tremor. an oscilla- tory nature.

Tools and methodologies used to quantify and analyse tremor generally lack standardisation and it is not unusual to find completely opposite conclusions in studies dealing with a similar problem (3). We decided to select a pathology charac- terized by relatively well-known symptoms and to examine critically the ability of a typical test to detect differences in the subjects’ performance.

Therefore, the purpose of this study was to determine the sensitivity (detecting small differences in tremor), specificity (detecting differences specific to a particular condition), and consistency of a simple, portable accelerometer system (the DPD system). First. tremor amplitude in patients with Parkin- son’s disease and in healthy normal control subjects was quan- tified using precise displacement data from a laser-based sys- tem (LB). Second, all subjects were classified into two groups (i.e.. low-tremor amplitude and high-tremor amplitude groups) and tested using the DPD system. Third. the five characteristics calculated by the DPD system were compared between high- and low-tremor amplitude groups and between groups defined by a neurological assessment of presence or absence of Parkinson’s disease symptoms. Finally. validity was tested by a simultaneous recording of tremor by the two systems and consistency was checked by testing another group of subjects twice using the DPD system.

Fifty-four subjects were included in the primary study, 21 of them in the early stages (I and II) (12) of Parkinson’s dis- ease. Control subjects were selected in the same age range. All the subjects were assessed by a neurologist on the same day as the tremor recordings were taken. The clinical exami- nation revealed that all of the control subjects were free of neurological disorders except one who had a moderate rssen- tial tremor. and two others who were assessed as having mild essential tremor. These last two. however, had tremor indis- tinguishable from normal tremor in our recordings. All sub- ,jects performed the two types of test (DPD and LB) with each hand. The LB test was done twice in succession for each hand. with a 30-s tracking task between the two trials.

Limitatiot2s

Tremor was measured using the DPD and LB systems within a period of about 1 h so some differences may be present due to changes in conditions over this time (e.g.. fa- tigue, anxiety). Changes in amplitude were sometimes ob- served visibly over periods of a few minutes. Also, patients were under medication and this may have contributed to hide some of the symptoms. but it is well known that tnedications have a limited effect on tremor (7). To what extent medica- tions alter tremor characteristics aside from amplitude is un- known.

EDWARDS AND BEUTER

The Danish Product Development Tremor Test (DPD)

This tremor test has been developed by Danish Product Development (Fig. 1). It measures postural tremor succes- sively in each hand during about 8.2 s while subjects look at a light stylus that they hold horizontally 10 cm away from their navel. The stylus is sensitive in a plane perpendicular to the tube axis and is individually calibrated with a calibration file. Two axis microaccelerometers embedded in the tip of the sty- lus measure accelerations in orthogonal directions. Samples arc taken at 500 Hz but are subsampled at 31.25 Hz after be- ing put through a low-pass filter (5). The raw accelerations, af- ter calibration, are reported to be accurate to within l-10% (standard deviation) across the frequency range used [(4). p. 201. The stylus is connected to a personal computer via a datalogger. Fourier transforms are calculated from the two time ser-ies and combined by a Pythagorean sum to give a sin- gle power spectrum.

Recorded data (time series and power spectrum) and group statistics can be visualized immediately and power spectrum data can be exported in ASCII format for further analysis. Measures derived from acceleration data are based on the Fourier power spectrum. which gives a power distribu- tion of the data in the frequency domain. It is composed of 1 I6 discrete values in the 0.9-15 Hz range. each approxi- mately 0.12 Hz apart. This spectrum is supposed to react to “deviant” tremor patterns.

FIG. I. The Danish Product Development Tremor Test

POSTURAL TREMOR MEASURING SYSTEM

1.

2.

3.

4.

5.

The specific parameters examined (4) are as follows:

Tremor intensity (I) is the RMS of acceleration recorded in the 0.9-15 Hz band during the 8.192-s test period and is expressed in m/s*. This is usually called “amplitude” in the literature and we use the latter term throughout. Centre Frequency (F50) is the median frequency of the ac- celeration in the 0.9-15 Hz band during the 8.2-s test pe- riod: 50% of the area under the spectrum is at frequencies above the centre frequency and 50% is below. Units are Hz or s I. Based on a test performed by DPD on 61 control subjects, the normal human means for left and right hand are 7.5 and 7.2 Hz, respectively. Standard Deviation of Centre Frequency (SFSO) indicates the degree of irregularity of the tremor. Sixty-eight percent of the area under the spectrum lies within 1 SD of the cen- tre frequency. A simple rhythmic tremor has a small SF50 indicating that most of the area is within a narrow fre- quency band. Based on a test performed by DPD on con- trol subjects, the normal human means for left and right hand are 3.8 and 3.5 Hz, respectively. Harmonic Index (HI) compares the tremor spectrum with that of a single harmonic oscillation, which has a HI = 1 .OO. A tremor composed of a few dominating frequencies has a high HI. It is defined as the area above the normal- ized spectrum (i.e., the highest peak is set to a power of I) between 0.Y and 14.9 Hz and below a power of 1. Based on a test performed by DPD on control subjects, normal tremor has a HI centered around 0.88, reflecting its irregu- lar character, but values up to 0.95 are not unusual. Tremor Index (TI) is a single measure incorporating the four previous measures. Any value deviating significantly from the norm will contribute a smaller than usual amount to the TI. The method of calculation is as follows: Let the recorded value for each of the four characteristics be de- noted K,. Let the normal human mean and SD for each characteristic be denoted hl, and S,. respectively. Then the TI is given by

TI = F c(ai exp (- I (K, - MI)/S, I).

where I;is a scaling factor to make the mean Tl of the test sample of normal human subjects equal 100 and the ai are l/6 for all except amplitude, for which a, is 113. (There is also a fifth component. SHI, a measure of dispersion of the Harmonic Index. which was used in previous versions of the DPD system but is no longer reported. though it is still apparently used in the calculation of the TI).

Past studies have shown that a LB system (2) measuring displacement can be used to record human tremor with high precision. In the present study. the LB system was used to quantify the amplitude of postural tremor and to classify the subjects into two groups: one group of subjects whose tremor was larger (detrended RMS > 0.17 mm) and the other group whose tremor was smaller (detrended RMS < 0. I7 mm).

The LB system is placed at a fixed distance from the finger tip. This system uses analog output sensors based on optical triangulation range measurement. The laser beam emitted from the light-emitting element (semiconductor laser) passes through the projector lens to a target. A part of the diffuse- reflected laser light passes through the receiver lens to a spot on the position-sensitive device. The position of the light spot varies actor-ding to the detected distance. The change in out-

97

put currents caused by a deflection of the light spot OII the po- sition-sensitive device is used to determine distance. Because this deflection is independent of the volume of the incident light. a stable distance measurement can be made.

Proctdures Used with the Laser-Bused System (LB)

The subject is comfortably seated in a chair with the elbow joints flexed at 90” and resting on a foam-padded support. The subject’s forearm is lying pronated on the padded sup- port. The index finger is extended while the remaining fingers of the tested hand rest in a semiflexed position on a specially molded soft support. Interphalangeal joints of the index are blocked by a light splint (-7 g) and vertical displacement of the index at the metacarpo-phalangeal joint is measured at the center of the finger nail (-10 cm from the joint). Visual feedback is presented on an oscilloscope screen placed 80 cm in front of the subject. A line corresponding to the finger posi- tion is displayed on the screen and subjects arc asked to try to keep this line on a fixed reference line. The analog signals re- corded from the laser are sampled at 200 H7 for 30 s using an acquisition system (Experimenter’s Workbench. DataWave Technologies, Longmont, CO).

The displacement data from the lasers were detrended by subtracting cubic polynomials fitted to 200 points (which cor- responds to 1 s) every 0.5 s, over five 1024-point segments. A gradual transition was made (by linear combination) between the overlapping parts of successive cubits. This was intended to capture the shape of the curve smoothly. The magnitudes of the five segments were then averaged in the Fourier do- main and tremor amplitude was calculated as the RMS of dis- placement of the KY-15 Hz component of the detrended and averaged signal. This method was used to remove the low-fre- quency component of the signal (due to drift. breathing, and hear-tbeat). and was preferred to a low-pass filter because it also removed the contribution of these factors to higher fre- quency parts of the spectrum (due to the effect known as “leakage”). Though curve fitting is usually avoided in spectral analysis. as it can create small artifacts in the spectrum. here we are only calculating amplitude of tremor (essentially a time domain characteristic).

RtSC!LIS

The distribution of tremor amplitudes recorded with the LB system for the 216 trials (54 subjects X 2 hands X 2 trials) had a bimodal form (Fig. 2). with a large group below about 0.17 mm and a smaller group with a mode above this value. This was used as a criterion for dividing the trials into low- and high-amplitude groups. A particular hand of a particular subject was deemed to belong to the high-amplitude group if either of the trials for that hand did. This gave 12 cases (sub- ject. hand) in the high-amplitude group and 96 cases in the low-amplitude group for subsequent analysis using the DPD system. Note that the high-amplitude group includes four cases (out of 66) from the control group and the low-ampli- tudc group includes 34 cases (out of 42) from the patient group (including the nonaffected hand in particular).

Kesults for the five characteristics reported by the DPD system for each hand of each subject in the two groups (as de- fined by the LB system) arc summarized in Table 1 (means

98 EDWARDS AND BEUTER

I I I I I I

-5 -4 -3 -2 -1 0

Log intensity (mm)

FIG. 2. Histogram of LB tremor amplitude from detrended displacement data (54 subjects x 2 hands x 2 trials). The arrow divides the high- and low-amplitude groups. (Note that the high-amplitude group includes trials from control subjects and the low-amplitude group includes trials from patients).

and SDS). Figure 3 gives a histogram of DPD amplitudes in the two groups.

The high- and iow-tremor amplitudes as calculated by the DPD system are not neatly separated, although there is a sig- nificant difference in mean amplitude for the two groups. The other characteristics vary by differing amounts between the two groups. Centre frequency is distributed similarly for both groups. Harmonic index and tremor index show some separa- tion, although the means are still within 1 SD of each other. Frequency dispersion separates the two groups slightly more. Therefore, the other three characteristics used in calculating the tremor index (centre frequency, frequency dispersion, and harmonic index), although somewhat correlated with ampli- tude, extract different information from the tremor.

The differences in amplitude results from the DPD and LB systems, indicated by the large overlap in DPD amplitude be-

TABLE 1

DPD SYSTEM

High-Amplitude Group Low-Amplitude Group Characteristic (n = 12) (n = 96)

Amplitude 0.314 2 0.372 0.095 + 0.036

Log(amplitude)* - 1.614 ? 0.944 -2.402 -t 0.291

Centre freq. 6.267 -c 1.059 6.408 2 1.083

Freq. dispersion 1.592 5 1.335 2.848 ? 1.048

Harmonic index 0.937 k 0.045 0.903 z 0.050

Tremor index 59.83 i 39.20 86.74 5 27.80

Values are means t SDS of spectral characteristics, 54 subjects, both hands (high- vs. low-amplitude groups).

*Amplitude is given because it is reported by the DPD system, but the distribution of log(amplitude) for our sample is closer to being Gaussian.

tween the high- and low-amplitude groups, suggested that fur- ther tests be done to check validity and consistency of the measuring systems.

Validity

Validity of the DPD system measurements was assessed by testing a small group of subjects simultaneously with the DPD and LB systems. Six subjects were tested four times each (two left, two right). The DPD stylus was held in the usual position and a light (<l g) card with a 90” fold was fixed to the end of it. Two lasers were directed at the two faces of the card to record x and y displacements in the plane of measurement of the accelerometers. The LB system recordings were then pro- cessed to match the method of acquisition used by the DPD system. The signals were matched in time as closely as possi- ble, and acceleration spectra were calculated from the dis- placement time series (see Appendix).

The resulting spectra were directly compared with those obtained from the DPD system. First, they were smoothed by averaging the power in each of 28 frequency ranges of width 0.5 Hz between 0.9 and 14.9 Hz. Then correlation coefficients were calculated between pairs of spectra obtained simulta- neously by the two systems. The mean of these coefficients was 0.950 and they ranged (with one exception) from 0.895 to 0.993. The exceptional coefficient was 0.665, and this was due to a large peak at the 1 Hz edge of the spectrum reported by the DPD system. This may have been caused by a slight roll- ing of the pen. (With this one removed, the mean of the coef- ficients is 0.962.) Although we attempted to mimic the DPD acquisition procedure exactly, there were slight discrepancies in processing method, due primarily to different sampling rates. The above correlations are thus very reasonable (see Discussion section below).

Also, the four characteristics that are used to form the

POSTURAL TREMOR MEASURING SYSTEM 99

‘High’ tremor group

-3 -2 -1 0 1

Log intensity (m&Z)

‘Low’ tremor group

-3 -2 -1 0 1

Log intensity (m&Z)

FIG. 3. Histograms of DPD tremor amplitude, for the high-tremor amplitude group (above) and the low-tremor amplitude group (below).

tremor index were calculated from the acceleration spectra derived from the LB system and compared with the corre- sponding values from the DPD system (see Table 2). It should be noted that the correlation between the DPD and LB am- plitudes is not as high if detrended displacement amplitude in- stead of acceleration is used with the LB system (for ampli- tude: 0.841; for log amplitude: 0.844). In fact, detrended displacement amplitude and acceleration amplitude, both cal- culated from the same LB data, had similar correlations (for amplitude: 0.851: for log amplitude: 0.859). If the displace- ment is not detrended, the correlation disappears entirely. The correlation coefficient for DPD acceleration amplitude vs. LB displacement amplitude (not detrended) was -0.096: for log amplitude. it was -0.154.

We also calculated correlation coefficients between the DPD system amplitude (acceleration) and the LB system am- plitude (detrended displacement) from the original sample of 54 subjects (see Table 3).

Clearly, the simultaneous recordings by the two systems. when processed in the same way, produce essentially the same results, whereas our original recordings with the two systems, which were not simultaneous and differed also in the variable recorded (displacement vs. acceleration). produce very differ- ent results, even for something as basic as amplitude.

Consistency

To get an idea of how much variability there is over short time periods for the same subject (same hand). we conducted

another test with two trials for each hand, trials separated by 1-2 min (and exceptionally, up to 11 min). A different group of 20 subjects, referred by a local physiatry department, was tested with the DPD system. Correlation coefficients were calculated for each of the five characteristics. As a compari- son, we calculated correlation coefficients (for detrended dis- placement tremor amplitude only) between two trials of the LB system (about 34 s apart) on the original sample of 54 sub- jects. Correlations for these have been calculated separately for the control group and the group of patients with Parkin- son’s disease (see Table 4).

Amplitude as measured by the DPD system is consistent between successive trials with a short delay between trials.

TABLE 2

CORRELATIONS BETWEEN SYSTEMS (DPD AND LB) FOR TREMOR AMPLITUDE AND SPECTRAL CHARACTERISTICS

IN SIMULTANEOUS RECORDlNGS

Trials Characteristic rho

DPD vs. LB ccntre freq. DPD vs. LB freq. dispersion

DPD vs. LB harmonic index DPD vs. LB amplitude DPD vs. LB log(amplitude)

I? = 24 = 6 subjects X 2 hands X 2 trials.

0.055 0.8 19

0.896 0.989

0.989

100 EDWARDS AND BEIJTER

TABLE 3 TABLE 5 (‘ORREI,ATfONS BETWEEN TREMOR AMPLI’I [IDES AS

CALCULATED BY DPD (ACCELERATION) AND LB (DETRENDED DISPLACEMENT) SYSTEMS

SPECTRAL C‘HARAC‘TERISTICS OF THE C’ONTROL GROUf AND THft PATIENTS WITH PARKINSON‘S DISEASE

(DPD ABOVE AND LB BELOW)

‘Trial5 rho

LB 1 vs. DPD amplitude

LB I vs. DPD log(amplitude)

LB 2 vs. DPD amplitude

LB 2 vs. DPD log(amplitudc)

II = IOX = 54 suhiects X 2 hands.

0.2YS

0.444

0.270

0.475

Frequency dispersion is quite consistent (i.e., a concentration of power in a small frequency range tends to remain) but other characteristics are less robust and tend to vary to a greater degree. even when the amplitude does not. The ampli- tude is similarly consistent across trials with the LB system for both patient and control groups (despite the presence of a tracking task between the two trials: see the Method section).

Of the 54 subjects in the primary study, the one who had a large essential tremor was removed from the sample for this part of the study. to avoid skewing the results for the group without Parkinson’s disease. The values of the DPD system’s five characteristics were compared between the group with and without Parkinson’s disease, using the most affected hand of each subject (defined as highest DPD amplitude, or if am- plitude was equal between the two hands, lowest tremor in- dex) (set Table 5 and Fig. 4). The displacement amplitude as measured by the LB system is included in the table for com- parison (worst hand defined as highest detrended displace- ment amplitude).

There are differences in the distributions of the character- istics for the Parkinson’s and non-Parkinson’s groups but there is considerable overlap in every case. A one-sided t-test to determine whether the tremor index for the group with Parkinson’s disease is lower than for the group without Par- kinson’s disease (null hypothesis: TI(P) 2 TI(NP); alternative hypothesis: TI(P) < TI(NP)) rejects the null hypothesis at the 95% level (p = 0.029).

TABLE 4

CORRELATIONS BETWEEN FIRST AND SECOND TRIALS FOR TREMOR AMPLITUDE AND SPECTRAL

CHARACTERISTICS (BOTH HANDS)

TI-ialc I, Characlrli\l~c rho

DPD vs. DPD 40 centre freq. 0.736

DPD vs. DPD 40 freq. dispersion 0.812

DPD vs. DPD 40 harmonic index 0.263 DPD vs. DPD 40 tremor index 0.575

DPD vs. DPD 40 amplitude 0.945

DPD vs. DPD 40 log(amplitudc) 0.946

LB vs. LB (control) 66 amplitude 0.828

LB vs. LB (control) 66 log(amplitude) 0.801 LB vs. LB (patients) 42 amplitude 0.971

LB vs. LB (patients) 42 log(amplitude) O.YO7

Chal-actcristics

Amplitude

Lo&amplitude)

C‘entre freq.

Fry. dispersion

Harmonic index

Tremor index

0.2 12 2 0.2Y2

- I.933 c 0.73s

6.310 z 0.82h

I.805 2 1.174

O.Y40 -c 0.033

72.00 z 39.00

(‘ontrol Group (!I = i’)

0.093 2 0.022

2.405 2 0.247

6.475 i 0.074

3.1 IY L 0.912

O.XYS ‘, 0.054

‘J2.03 t 24.49

,I :. 42 II - 6-l

Amplitude (LB) 0.1 13 i 0.136

Log(amplitude)(LB) -2.55Y z 0.798

*Both trials were included for the LB bystem.

O.OhO i 0.034

-2.946 + 0.S 17

DISCUSSION

High- vs. Low+Amplitude Groups

It is expected that centre frequency, frequency dispersion, and harmonic index do not neatly separate the high- and low- tremor amplitude groups. In fact, if other measures were to be used that correlate highly with amplitude they would not carry any extra information. The fact that there was some separation with frequency dispersion and with harmonic index is probably due to the fact that higher amplitude tremors generally tend to be more concentrated in narrow fr-equency ranges.

More surprising was that there was also considerable ovcr- lap between the high- and low-tr-emor amplitude groups for the DPD tremor amplitude measure. although there was a sig- nificant difference in means for the two groups. A careful look at individual amplitude results revealed that a few sub- jects showed a high amplitude from the DPD system and vir- tually no tremor for the LB system and vice versa. Potential causes for these discrepancies are:

I. imprecision in the DPD system (validity); 2. an inherent tendency of tremor amplitude to vary over-

time with changing conditions (consistency); and 3. a difference between amplitude measures (acceleration vs.

detrended displacement).

Validity

The first of these potential causes of discrepancy is a yues- tion of validity. The simultaneous recordings with the DPD and LB systems showed that this is not a primary factor. The spectra obtained in parallel fashion by the two systems were not identical, but the small discrepancies could be due to mi- nor differences in processing (see Appendix). The correlation coefficients between parallel smoothed spectra are high enough to show that they are recording essentially the same data, and when the power in the spectrum is lumped into a single ampli- tude figure, the correlation is even more convincing.

The centre frequency, as calculated from the two parallel recordings, was also very consistent, and the frequency disper- sion and harmonic index only slightly less so. The lower corre- lation for frequency dispersion (0.819) seems largely to be due to the exceptional case with the spurious peak in the DPD spectrum: when it is removed. the coefficient becomes 0.937.

POSTURAL TREMOR MEASURING SYSTEM

a

101

b

-3 -2 -1 0 1 -3 -2 -1 0 1

Log intensity (mM2) Log intensity (mls”2)

c d

4 5 6 7 8 9 4 5 6 7 8 9

Centre frequency (Hz) Centre frequency (Hz)

e

0 1 2 3 4 5

Frequency dispersion (Hz)

0 1 2 3 4 5

Frequency dispersion (Hz)

h

0.75 0.80 0.85 0.90

Harmonic index

1 .oo 0.70 0.75 0.80 0.95 0.90 0.95 1 .oo

Harmonic index

0 50 too 150 0 50 100 150

Tremor index Tremor index

FIG. 4. Histograms of each of the five DPD characteristics using the worst hand only. for the group with Parkinson’s disease (left) and the control group (right) for log amplitude (a. b). centre frequency (c, d), frequency dispersion (e. f), harmonic index (g. h). and tremor index (i. j).

The harmonic index, however, has a coefficient of about 0.89 (even with the exceptional case removed) and so appears to be slightly less robust with respect to minor differences in ac- quisition or processing methods.

The problem of spurious low-frequency peaks in the DPD spectrum can best be dealt with by ensuring that the subject doesn’t rotate the pen at all during acquisition, by explana- tion, and, if necessary, redoing the test, because the frequency information is available instantaneously. This is an inherent concern with accelerometric systems.

The second potential cause of the discrepancies is a ques- tion of consistency. Our test-retest reliability study shows that amplitude is consistent between trials conducted within a few

minutes of each other (even more consistent than with the LB displacement amplitude, though there we had a tracking task between trials, which may have affected the tremor). Fre- quency dispersion was also quite consistent across trials, as was centre frequency, though with a slightly lower correlation coefficient. Harmonic index. however. was quite inconsistent across trials.

Although this reassures us that the test is reliable as a sys- tem for measuring tremor amplitude, it does not preclude the likelihood that over longer periods of time tremor characteris- tics, amphtude in particular, can vary a great deal. In fact, there are considerable discrepancies between severity of pos- tural tremor as determined by the DPD system and the LB system and as assessed by the neurologist on the same day, which can only be due to tremor coming and going between the tests. For example, the neurologist found no postural

102 EDWARDS AND BEUTER

tremor for three of the subjects with Parkinson’s disease who had relatively high tremor amplitude measures from both the DPD system (worst hand amplitudes: 1.38, 0.15, 0.17 m/s*) and the LB system (worst hand amplitudes: 0.23, 0.30, 0.92 mm) in- cluding the highest for each system. Recall that for detrended displacement, high amplitude was defined as more than 0.17 mm and the normal range for the DPD system is 0.11 ? 0.03. Another three with low-tremor amplitudes according to both the DPD (worst hand amplitudes: 0.11,0.09.0.08 m/s*) and LB (worst hand amplitudes: 0.07, 0.07, 0.04 mm) results were as- sessed by the neurologist as having significant postural tremor. In another case, the DPD test found a large tremor in the right hand (0.23 m/s*), agreeing with the neurologist, but the LB sys- tem did not (0.05 mm). In yet another case, the LB system found a large tremor in the right hand (0.20 mm) reported by the neurologist, but the DPD system did not (0.10 m/s*).

These differences cannot be due to inherent unreliability in the DPD measurements (or in the LB measurements) as the test-retest results as summarized in Table 4 are too consis- tent. Hence, they must be due to genuine changes in the am- plitude of tremor of some subjects over the testing period (the neurological assessment, the DPD test, and the LB test in some cases spanned a period of 2 h or more). In fact, we have some LB recordings of tremor that visibly changed in ampli- tude over times of 30 s to several minutes (6).

Displacement vs. Acceleratiort

The third potential cause of discrepancies in amplitude measurements is the choice of whether to measure displace- ment or acceleration (or even velocity) amplitude. There is, of course, a clearly defined relationship between acceleration and displacement, and it is theoretically a simple matter to calculate one from the other in either the time or frequency domains (3,7,19). The basic effect is that acceleration (in rela- tion to displacement) emphasises higher frequencies and sup- presses lower frequencies. A pronounced tremor concen- trated in a narrow frequency band will yield a high-amplitude measure in either case (unless the frequency is very low), but with a broad, noisy spectrum the effect will depend on how the power is distributed. It is certainly true that two oscilla- tions of the same displacement amplitude, located at different frequencies. will yield different acceleration amplitudes and vice versa.

The detrending removes the large-scale and very low-fre- quency component of the displacement data, which is not a part of tremor and is naturally supressed in acceleration mea- surement. However, the above effect is still present in the re- maining data.

In recordings of human tremor, this effect may be quite significant, in light of the comparisons made above, especially between our original LB and DPD results in Table 3, but even acceleration and displacement amplitude calculated from the same time series are not identical. Presumably, the neurolo- gist, who assesses severity of tremor visually, is measuring dis- placement rather than acceleration amplitude. There has been considerable debate over which is preferable, and we do not resolve this debate here, but our results point out that there can be a significant difference between the two methods and that it is important to address this issue.

Control Group vs. Group with Parkinson’s Disease

Finally, the DPD system is designed to detect anomalies in tremor (compared to control subjects), so we asked if it could separate the groups with and without Parkinson’s disease. The

distributions of most of the characteristics for the two groups were clearly different, but there were nevertheless consider- able overlaps for all the characteristics. There were a few sub- jects with Parkinson’s disease with high tremor amplitude having a sharp narrow peak in the spectrum, which made the means for amplitude, harmonic index, and frequency disper- sion different from those of the control group. However. many of the subjects with Parkinson’s disease had tremor that was well within the normal ranges in amplitude and in the other characteristics. Interestingly, the centre frequency was virtually the same in the two groups. The tremor index, which combines the four characteristics in a single number, showed the same overlap between the groups. Diagnosis on the basis of tremor index by itself, or any of the individual characteris- tics, would therefore not be very reliable.

There are several possible factors contributing to these overlaps between results for Parkinson’s and non-Parkinson’s groups. First. the neurological assessment is itself uncertain: thcrc is a considerable error rate in diagnosis (15.17). Second, most of the subjects were in early stages of the disease (stages I and II), which increases the difficulty of detection (though this was the intention of the study). Some subjects may have been in a state of transition from normal to pathologic tremor. Third. most of the patients with Parkinson’s disease were on medication and had taken their medication on the day they were tested (the length of time before being tested varied). Medication generally reduces tremor to some degree, though not completely. Fourth, there is variation over time in ampli- tude of tremor. as mentioned above. so it is possible that some of the subjects with Parkinson’s disease did not have signifi- cant tremor at the moment of testing. In fact, during the ex- periment we noted that some patients with Parkinson’s dis- ease did not have visible postural tremor whereas some control subjects had mild essential tremor.

How well did the centre frequency, frequency dispersion, and harmonic index reported by the DPD system perform. in general? In our study, ccntre frequency did not vary much in distribution between the group with and without Parkinson‘s disease, or between the “high-amplitude” and “low-ampli- tude” groups. Most of the literature, however, suggests that subjects with Parkinson’s disease should have lower fre- quency tremor. This is perhaps a matter of how one defines “frequency of tremor.” In the case of a pronounced, single peak in the power spectrum, there is no ambiguity, but in a broad, noisy spectrum there are many ways to select a fre- quency to try to reflect the tremor by a single figure. Taking the largest peak is one way, but is subject to random fluctua- tions in the peaks and may even depend significantly on the frequency resolution in the spectrum, though smoothing or other means of spectral estimation reduce the variance of the power estimates.

Centre (median) frequency is an attempt to give a stable measure that is always well defined. For pronounced tremors, it will correspond with the dominant peak exactly, but in other cases there may be problems. If the spectrum has two peaks of similar size. for example, the centre frequency may fall between them in a region with little power. In this case, a small change in power in the peaks could cause a large change in the centre frequency. However, in both the simultaneous recordings by the DPD and LB systems and in the test-retest recordings with the DPD system, centre frequency was rea- sonably consistent.

Frequency dispersion showed some separation between high-amplitude and low-amplitude groups as we11 as between the group with and without Parkinson’s disease. but in both

POSTURAL TREMOR MEASURING SYSTEM 103

cases, with considerable overlap. It was fairly consistent in the test-retest results with the DPD system and in the simulta- neous recordings, though less consistent than centre fre- quency or amplitude. Again, if the boundaries of the region of the spectrum around the centre frequency having 68% of the power fall in an area of low power, small changes in distribu- tion could cause large movements in these boundaries.

Harmonic index also separated the high-amplitude and low-amplitude groups and the group with and without Parkin- son’s disease to some extent. but was very unreliable in the test-retest experiment and slightly less even in the simulta- neous recording experiment. This measure is perhaps unsta- ble because it depends on the normalization of the power spectrum to the height of the highest peak. Random differ- ences from trial to trial. or differences in processing methods. can cause significant changes in the height of a peak. which will have a correspondingly large effect on the harmonic in- dex. If a peak becomes half the height but twice as wide, con- taining the same amount of power, for example, the area un- der the spectrum will more than double, and so the harmonic index will be twice as far from 1.0. There is clearly some value in a measure like this, but it would be better if defined so as to be more stable across trials.

If there is something anomalous about the tremor of the subjects with Parkinson’s disease aside from amplitude, it is not clearly evident in the spectral characteristics reported by DPD. The DPD system does give a graphical display of the power spectrum (and the spectrum can be exported to an ASCII file) so there is the possibility of looking more closely at details of the spectrum or of calculating other characterir- tics from it.

Concluding Remarks

We have verified that the DPD system accurately mea- sures tremor spectra and certain characteristics of tremor. Some aspects of the data acquisition process used are not en- tirely clear from the documentation provided. although most of these points we were able to clear up later by contacting the company. The DPD system is easy to use, robust. and porta- ble. though care must be taken to make sure the test is done correctly (e.g., no rotation of pen), and this should also be considered in comparing it with other systems.

What is evident from our study, however, is that accuracy is not the only concern. Other issues that must be addressed include: what are the best characteristics (in terms of specific- ity and reliabilitv) of tremor to measure? and how are changes in tremor over time to be dealt with?

Even tremor amplitude proved not to be a clear-cut issue. Tremor amplitude calculated from the displacement data of the LB system is not highly correlated with the amplitude re- corded by the DPD system (Table 3). This seems to be due partly to changes in the tremor over time and partly to the dif- ference between acceleration and displacement data. A high amplitude in displacement (even when detrended according to the scheme described above) does not always imply a high amplitude in acceleration or vice versa, so these should be considered different characteristics. Acceleration amplitude did not clearly separate all patients with Parkinson’s from the control subjects either, although the ones with the largest tremors stood out. Centre frequency in acceleration spectra did not seem to be very specific. at least to patients with Par- kinson’s disease. Frequency dispersion appeared to be some- what useful in separating the patients with Parkinson’s disease from the control subjects, and was fairly reliable across trials.

Harmonic index as defined by DPD was also useful in sepa- rating the groups but was not a very stable measure.

Thus, whereas it is clear that the characteristics reported by DPD have a general bearing on distinguishing pathologic tremor, they are crude tools. The DPD system did clearly identify subjects with very pronounced tremors, though many of the patients with Parkinson’s disease were not thus identi- fied. The one subject with a moderate essential tremor also showed clearly abnormal results. It does not appear to be sen- sitive enough to identify subclinical cases or to differentiate between different types of tremor (such as essential and par- kinsonian tremor).

It should be noted that we used the testing procedures sug- gested by the DPD company. In the light of our results and comments below. it might be possible to improve the perfor- mance of the system by having the subject perform a mental task such as counting backwards during testing. In addition, most patients tested were on medication. Although this means that we cannot draw conclusions about similar patients without medication, the system is really most useful if it can detect the pathology even when it is not visibly evident.

If we are to do better in understanding and diagnosing dif- ferent types of tremor, we will need measures that are, if pos- sible, more specific to pathology or to particular pathologies. but are still stable and reliable. If amplitude is to be used, the question of whether to use displacement or acceleration (or velocity) or a combination must at least be addressed. Aside from spectral characteristics, it may be that significant infor- mation is held in the morphology of tremor. This has appar- ently not been studied. The question of fluctuations in aspects of tremor over time seems particularly important. In princi- ple. there seem to be three possibilities for dealing with this is- sue. One way is to find characteristics of tremor, if they exist, which are insensitive to transients but sensitive to real under- lying neurological differences. so that, for example, changes in anxiety or fatigue levels of a subject would not significantly af- fect results (6). Another way (as mentioned above) is to find experimental methods to ensure that an intermittent tremor is active during recording (without seriously altering the tremor due to fatigue). Or, finally, the fluctuations in tremor could themselves be studied. After all, transitions in tremor may hold significant information themselves (19).

ACKNOWLEDGEMENTS

The authors wish to acknowledge financial support from the Cree Board of Health and Social Services of James Bay, the Natural Sci- ences and Engineering Research Council of Canada, and the Fonds pour la Formation de Chercheurs et I-Aide a la Recherche (Quebec). We would also like to thank Dr. John Heeboll of Danish Product De- velopment, Ltd. for his patient and enthusiastic assistance with the many questions we had regarding the DPD system.

APPENDIX: C’ALCULATING ACCELERATION AMPLITUDE FROM DISPLACEMENT DATA

An acceleration spectrum corresponding to the one pro- duced by the DPD system can be calculated from displace- ment time data by differentiating the displacement time series twice, applying the same filter as is used by the DPD system and then extracting the spectrum. It can also be done by ap- plying the filter to the displacement data, extracting the dis- placement spectrum from this and then multiplying the spec- trum by the appropriate function to get the acceleration spectrum (3,19). The latter method may be better in that it avoids the problem of noise amplification in the differentia-

104 EDWARDS AND BEUTER

tion process. Both methods were tried and the second method gave slightly higher correlations to the DPD results.

Here are the details of the procedure. The DPD system starts its data recording 1.5 s after the keystroke that begins the test. In its analysis, it uses the following 5.192 s of data. Our LB recordings were started by pressing a key at the same time as the DPD starting keystroke, as nearly as possible. Then the appropriate 8.192 s of data were extracted for analy- sis. First, the I and y displacement signals were filtered using the same type of low-pass filter as used by DPD (5). except

I.

2.

3 .

4.

5. 6.

7.

8.

9. 10.

that because our data were collected at 200 Hz rather than 500 Hz, a 12-point filter rather than a 32-point filter was used, and then subsampling was done at 33.33 Hz (i.e.. every sixth point) rather than 31.25 Hz (i.e., every 16th point). Then the spectra for each series (X and y) were calculated and were converted to acceleration spectra by multiplying by (2rrf)4, where fis fre- quency in Hz. Finally, the two spectra were added (compo- nentwise), which is equivalent to taking a Pythagorean sum of .r and y acceleration values because the power spectrum val- ues have units of acceleration squared.

REFERENCES

Anger. W. K.; Cassito, M. G.: Liang, Y. X.: Amador. R.; Hoo- isma, J.; Chrislip. D. W.; Mcrgler, D.; Keifer. M.: Hortnagl, J.; Fournier, L.; Dudek. B.; Zsogon, E.: Comparison of performance from three continents on the WHO-recommended Neurohehav- ioral Core Test Battery. Environ. Res. 62:125-147; 1993. Beuter, A.; Cordo, P.; de Geoffroy, A.: The measurement of tremor using simple laser systems. J. Neurosci. Methods 53:47-54; 1994. Beuter. A.; de Geoffroy, A.: Can tremor be used to measure the effect of chronic mercury exposure in human subjects? Neurotox- icology 17(1):213-228; 1996.. Danish Product Development. Ltd.: TREMOR 3.0 user’s man- ual. Snekkersten. Denmark; 1994. Danish Product Development, Ltd.: Personal communication; 1995. Edwards, R.; Beuter. A.: How to handle intermittent phenomena in time series analysis: a case study of neuromotor deficit. Brain Cogn. 321262-266; 1997. Elble, R. .I.; Keller. W. C.: Tremor. Baltimore: The Johns Hop- kins University Press; 1990. Findley, L. J.; Cleeves. L.: Classification of tremor. In: Quinn, N. P.; Jenner, P. G., eds. Disorders of movement: Clinical phar- macological and physiological aspects, New York: Academic Press: 1989:505-519. Gillespie, M. M.: Tremor. J. Neurosci. Nurs. 243(3):17(&174; 1991. Graham, J. D. P.: Static tremor in anxiety states. J. Neural. Neu- rosurg. Psychiatry 8:5760; 1945.

Il.

12.

13.

14.

IS.

16.

17.

18.

IO.

Hallett. M.: Classification and treatment of tremor. JAMA 266(8): 1115-1117: 1901. Hoehn, M. M.: Yahr, M. D.: Parkinsonism: Onset, progression and mortality. Neurology 17:427-442; 1967. Keller, W. C.; Huber, S. J.: Tremor disorders of aging: Diagnosis and management. Geriatrics 44(5)3241; 1989. Keller, W.: Lang. A.; Vetere-Overfield, B.; Findley. L.; Cleeves, L.: Factor, S.: Singer. C.: Weiner, W.: Psychogenic tremors. Neu- rology 3Y:lOY4-1099: 1989. Larsen. J. P.: DuPont, E.; Tandberg, E.: Clinical diagnosis of Par- kinson’s discasc. Proposal of diagnostic subgroups classified at different levels of confidence. Acta Neurol. Stand. 89:242-251: 1993. Marsden, C. D.: Origins of normal and pathologic tremor. In: Findley, L. J.; Capildeo, R., eds. Movement disorders: Tremor. New York: Macmillan; 1984:37-84. Rajput, A. H.: Rozdilsky, B.; Rajput, A.: Accuracy of clinical diagnosis in parkinsonism: A prospective study. Can. J. Neural. Sci. 18:275-278; 1991. Rothwell, J. C.: Orthostatic tremor. In: Quinn, N. P.: Jenner, P. G., cds. Disorders of movement: Clinical, pharmacological and phys- iological aspects. New York: Academic Press; 1989:521-527. Wyatt, R. H.: A study of power spectra analysis of normal finger tremors. IEEE Trans. Biomed. Eng. 15:334.5; 1968.