Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation

Using the Fundamental Frequency

Beatriz Lopo Ferreira

1 a

, Virginie Felizardo

2,3 b

, Nuno Cruz Garcia

1 c

, Mehran Pourvahab

Henriques Zacarias

2,3

, Leonice Pereira

and Nuno M. Garcia

LASIGE, Faculdade de Ci

encias, Universidade de Lisboa, Lisboa, Portugal

Instituto de Telecomunicac¸

oes, Lisboa, Portugal

Universidade da Beira Interior, Covilh

a, Portugal

Instituto de Biof

ısica e Engenharia Biom

edica, Faculdade de Ci

encias, Universidade de Lisboa, Lisboa, Portugal

Keywords:

Parkinson’s Disease, Rest Tremor, Accelerometer, Fundamental Frequency, Monitoring.

Abstract:

Parkinson’s disease (PD) is a chronic, progressive and neurodegenerative disease that affects more than 10

million people worldwide. One of the cardinal symptoms of this disease is tremor, which is characterized

as an involuntary, oscillatory movement of a body part. The tremor associated with PD can be divided into

rest tremor, postural tremor, and kinetic tremor and is characterized as a regular and asymmetric tremor.

Emerging methods involve the use of data from inertial sensors to measure, analyze and quantify tremor and

other symptoms of PD. In this publication, a method for the monitoring of rest tremor scores is explored.

This method is based on the number of windows in a signal with a fundamental frequency within the rest

tremor frequency band and has potential to be applied as a support for monitoring this symptom. This method

had a 87.88% success rate for predicting rest tremor scores on the X axis of a 4 hour accelerometer signal,

establishing promising results that will be further explored in future work.

1 INTRODUCTION

Parkinson’s disease (PD) is a chronic, progressive,

and neurodegenerative disorder (de Oliveira Andrade

et al., 2020). In addition to affecting more than

10 million people worldwide, making it the second

most common neurodegenerative disorder, PD is the

fastest growing neurological disease, being estimated

that 12 million people will be diagnosed by 2010

(Shawen et al., 2020; Burq et al., 2022). This disor-

der can cause patients to feel several motor and non-

motor symptoms, with the cardinal symptoms being

bradykinesia, rigidity, and tremor (de Oliveira An-

drade et al., 2020; Huo et al., 2020).

Tremor is considered the most common move-

ment disorder and is characterized as an involun-

tary, oscillatory movement of a body part. The

tremor associated with PD is characterized as a reg-

ular asymmetrical tremor and can be rest tremor, pos-

tural tremor, or kinetic tremor. This symptom usually

manifests itself at rest and at the onset of the disease

tends to affect the hands (Zajki-Zechmeister et al.,

https://orcid.org/0000-0002-7437-9493

https://orcid.org/0000-0001-6874-3263

https://orcid.org/0000-0002-6371-3310

2020).

Due to the heterogeneity of PD, accurate moni-

toring and assessment of symptoms is extremely im-

portant for the continuous selection of the most ad-

equate treatment plan, as the disease progresses and

the severity of symptoms increases (Huo et al., 2020;

Smid et al., 2022). However, the golden standard

method to evaluate PD and its symptoms, the Uniﬁed

Parkinson’s Disease Rating Scale (UPDRS) (Goetz

et al., 2008), requires a professional to perform the

evaluation, leading to a low degree of objectivity, im-

partiality, and sensitivity (de Oliveira Andrade et al.,

2020). This method has a high within-subject vari-

ability and low test-retest reliability. Moreover, the

physical exams performed in the clinical evaluation

provide only a small sample of PD symptoms, which

may not accurately represent those symptoms outside

the clinic (Burq et al., 2022).

Researchers have shown the feasibility of using

data from inertial sensors, such as accelerometers and

gyroscopes, to measure, analyze, and quantify tremor

and other motor signs of PD, with accuracies from ex-

isting trials exceeding 85% for the detection of tremor

and bradykinesia (Burq et al., 2022; Shawen et al.,

2020). From these sensors, accelerometers are con-

312

Ferreira, B. L., Felizardo, V., Garcia, N. C., Pourvahab, M., Zacarias, H., Pereira, L. and Garcia, N. M.

Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation Using the Fundamental Frequency.

DOI: 10.5220/0013364500003938

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2025), pages 312-319

ISBN: 978-989-758-743-6; ISSN: 2184-4984

sidered the minimum necessary sensor for the char-

acterization of human activity, despite it not being

clear if this sensor alone is enough to characterize PD

symptoms and improve disease detection (Huo et al.,

2020).

These sensors are cost-efﬁcient, widely available,

and don’t rely on the interpretation of an experienced

professional, making them less subjective. Nonethe-

less, due to their potential time-consuming measure-

ments and complexity, most of these devices are not

currently used for clinical or home monitoring (Zajki-

Zechmeister et al., 2020; Smid et al., 2022). More-

over, their implementation in continuous in-the-wild

monitoring is limited by the battery life and memory

capacity of the sensor devices, which may lead to in-

termittent usage and inconsistent positioning (Shawen

et al., 2020). Another challenge lies in the distinction

between symptoms, like tremor, and normal daily ac-

tivities in signals collected in uncontrolled living con-

ditions due to noise (San-Segundo et al., 2020).

There is currently a lack of established meth-

ods/models for the correlation of PD symptoms’ sen-

sor data with UPDRS scores and no commercial sys-

tems to complement clinical assessment of symptoms

(Huo et al., 2020; San-Segundo et al., 2020). Meth-

ods to evaluate PD can generally be divided into two

categories. One being machine learning models that

can use sensor data and features (typically from the

time and/or frequency domains) from that data to de-

tect the presence and severity of symptoms (Shawen

et al., 2020; San-Segundo et al., 2020). The other

methods involve the correlation of sensor data with

PD symptoms through the creation of methods based

on the physical and physiological characteristics of

symptoms (Huo et al., 2020).

Regarding the use of the fundamental frequency

for the assessment of tremor in current literature, in

(Kuosmanen et al., 2020), the authors use the fun-

damental frequency, extracted from the periodograms

obtained through the Welch’s method, to categorize

game sessions, during which data was collected, as

types of tremor. In (Bazgir et al., 2018), a short-

time Fourier transform (STFT) was implemented and

the fundamental frequency, along with other ﬁve fre-

quency domain features, were extracted. Four classi-

ﬁers were tested to estimate hand tremor and Naive

Bayesian achieved the highest accuracies of 89% and

almost 90% before and after feature selection, re-

spectively. The fundamental frequency was one of

the most discriminative features. In (Pierleoni et al.,

2019), a system, including a data collection device

and two detection and classiﬁcation for tremor and

freezing of gait, was developed. In this system, the

authors extract several frequency domain feature, in-

cluding the fundamental frequency, and deﬁne their

own parameter to estimate tremor severity using these

features. This system achieved an accuracy of 97.7%

for tremor. Moreover, the fundamental frequency is

one of the most commonly extracted features for the

analysis of voice changes in PD(Amato et al., 2023).

In this paper, a simple method with a low com-

putational cost for the estimation of UPRDS rest

tremor scores based on the number of windows with

a fundamental frequency between 3 and 6 Hz is ex-

plored. Firstly, the dataset used and the steps to ap-

ply this method, including the preprocessing, feature

extraction, and analysis of the data, are explored in

the methods in section 2. Then the results are pre-

sented and discussed in sections 3 and 4, respectively.

This last section also includes a brief discussion of

the dataset and the periodograms obtained with the

Welch’s method.

2 METHODS

In this section, the dataset and the method proposed

and implemented to estimate the rest tremor score at-

tributed to each individual using the UPDRS are pre-

sented.

2.1 Dataset

Due to a lack of time and resources needed to col-

lect a dataset of tremor data, an open-source dataset

collected by the authors of (Adams et al., 2021) was

used for the implementation of this framework. This

dataset contains accelerometer data from 34 individ-

uals, 17 people with Parkinson’s disease (PwPD) and

17 HC, collected at a sampling rate of 31.25 Hz using

5 lightweight MC 10 BioStamp RC sensors, placed

on the trunk, left and right anterior thighs, and left

and right anterior forearms. The data collection began

during a UPDRS clinical evaluation, performed dur-

ing the ON and OFF states of medication, and lasted

around two days.

Only the data collected from the anterior fore-

arms was used, since the study was focused on tremor

from the upper extremities. Additionally, for the in-

dividuals with ID 7 and ID 60, the dataset only con-

tains ﬁles from the sensor on the right forearm. This

dataset also contained demographic data, annotations

from the UPDRS evaluations, and clinical assessment

data, which includes the scores attributed to the rest

tremor of each individual during the evaluation. De-

mographic data and the rest tremor score given to the

anterior forearms of each individual during the ON

medication state are presented in table 1.

Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation Using the Fundamental Frequency

313

Table 1: Demographic and clinical data during the ON med-

ication state of individuals in the dataset.

ID Group Gender Age

RT score

left right

5 Control F 74 0 0

6 PD M 73 2 0

7 Control F 52 0 0

8 Control F 77 0 0

10 PD F 72 0 2

12 PD F 64 1 1

13 PD F 60 0 0

14 Control F 56 0 0

15 PD M 65 0 1

16 Control F 62 0 0

17 PD M 74 0 3

18 Control F 66 0 0

20 Control F 68 0 0

22 Control M 68 0 0

23 PD F 68 0 2

24 PD M 62 0 0

25 PD F 72 0 1

27 Control F 54 0 0

30 Control F 68 0 0

33 PD M 46 0 0

35 PD M 67 2 0

36 PD M 69 3 0

38 PD M 78 1 0

39 Control F 74 0 0

40 PD F 75 1 0

41 Control M 75 0 0

42 PD M 84 0 0

43 Control F 69 0 0

44 PD F 63 0 0

45 Control M 64 0 0

58 Control M 39 0 0

60 Control F 65 0 0

62 Control F 56 0 0

63 PD M 37 0 1

The study focused on the evaluation of rest tremor

in-the-wild with free-living conditions. Other types

of tremor, like postural tremor or kinetic tremor, were

not evaluated since the dataset only contains the UP-

DRS scores attributed during the clinical evaluation

for this type of tremor.

From each forearm of each subject, a time period

of four hours was selected from the respective data

ﬁle. This interval starts ﬁve minutes after the end of

the clinical evaluation, known through the timestamps

in the annotations. Thus giving a ﬁve minute time

margin to guarantee no data from the clinical evalua-

tion is included in the selected data. Since the eval-

uation ended on the ON medication state, the patient

is considered to be in that state for the selected four

hours. The rest tremor scores attributed to the subjects

for that medication state are shown in table 1.

2.2 Preprocessing

Firstly, the magnitude of the accelerometer signal for

the selected interval was calculated with the square

root of the sum of squares of each axis. Given that

the dataset has a sampling frequency of 31.25 Hz, no

downsampling step was implemented.

A 4

order high-pass Butterworth ﬁlter with a cut-

off frequency of 0.5 Hz was applied to the axes and

the magnitude vector to remove the effects of gravita-

tional acceleration and low-frequency noise.

Lastly, the axes and magnitude vector of the se-

lected data were segmented into 10 second windows,

with no overlap. This 10 second interval allows the

averaging of periodograms of segments from these

windows performed for the Welch’s method. Due to

the averaging of periodograms, the lack of overlap be-

tween the windows is also important for the correct

implementation of the Welch’s method. This method

is better described in subsection 2.3.

2.3 Feature Extraction

The features extracted in this study belonged to the

frequency domain or time domain. These features

were extracted from each of the axes and the mag-

nitude vector and are shown in table 2.

Table 2: Features extracted from the frequency and time

domains.

Domain Features

Frequency AUC, pv, F0, F50, SF50,

and |F50 − F0|

Time RMS, range, mean, variance,

skewness, and kurtosis

The frequency domain features were extracted

from the PSD. The PSD was estimated using the

Welch’s method, which is used to determine the

power contained in a signal’s frequency components

(Kuosmanen et al., 2020). To apply this method, the

data from each 10 second window was divided into

four 2.5 second segments with an overlap of 50%. Af-

terwards the periodogram of each segment was calcu-

lated and all the periodograms in each 10 second win-

dow were averaged (Welch, 1967). The periodograms

are brieﬂy explored in section 4.2.

The frequency domain features extracted from the

averaged periodograms were the area under the curve

(AUC) between the frequencies of 3 and 6 Hz, peak

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

314

value (pv), fundamental frequency (F0), central fre-

quency (F50), frequency dispersion (SF50), and the

absolute value of the difference between F50 and F0

(|F50 − F0|). These features were the same as the

features extracted by (Kuosmanen et al., 2020). The

AUC was computed between the frequencies of 3 Hz

and 6 Hz using the trapezoidal rule. This feature was

calculated between these frequencies to determine the

power of the signal in the interval corresponding to

the rest tremor frequencies. The features extracted

from this domain are described in table 3. The time

domain features extracted were the root mean square

(RMS), range, mean, variance, skewness, and kur-

tosis. Despite extracting features from both the fre-

quency and time domains, only features from the fre-

quency domain were explored in this study.

2.4 Data Analysis

Since rest tremor is normally observed in the between

the frequencies of 3 and 6 Hz, the windows where

the fundamental frequency (F0) is in that frequency

band were selected. It can be assumed that the se-

lected windows contain the features when, if present,

rest tremor is the dominant type of tremor.

We noticed an increase in the amount of windows

with F0 between 3 and 6 Hz as the scores increased.

For this reason, the data was analyzed regarding the

number of windows that were selected in the 4 hour

time interval, previously selected in the preprocess-

ing. That number of windows was then correlated to

the UPDRS rest tremor score. After that number was

counted for X, Y, and Z axes and the magnitude of

both forearms of each subject, the mean of the three

axes was calculated. The 4 hour intervals were di-

vided into two intervals of 2 hours and four intervals

of 1 hours and the same analysis was also performed

for those intervals.

The values of the number of windows that corre-

spond to each tremor score were deﬁned and are pre-

sented in table 4. The values for the intervals of 2

hours and 1 hour were adapted from the values for the

4 hour interval. Since no subject in this dataset had

a rest tremor score of 4 from the clinical evaluation,

this score was excluded from the analysis.

The number of windows can be divided by 200,

100, and 50 for the 4, 2, and 1 hour intervals, respec-

tively, to get the ﬁnal UPDRS score prediction, since

these values represent the number of windows in each

interval corresponding to a score for their respective

time interval, like shown in table 4. The values ob-

tained from this division were rounded down to get

the ﬁnal score prediction.

3 RESULTS

In this section the results obtained for the data analy-

sis method proposed to estimate the rest tremor score

attributed to subjects during a UPDRS clinical evalu-

ation are presented.

The increase of the number of windows with the

increase in rest tremor score can be veriﬁed in table

5. This table displays, for the X axis of the 4 hour

interval, the number of instances when the data from

a forearm of any subject was attributed each score (Nº

samples), the mean number of windows, and the mean

predicted score, before rounding the prediction val-

ues, for each real score. For the samples that were

evaluated, the mean number of windows increases

with the increase of the real rest tremor score, for all

scores except score 3. Furthermore, the same increase

can be seen in the mean prediction value, except for

score 3. Excluding this last score, all mean values for

both the number of windows and the prediction are

within the established values for a correct prediction.

Table 6 contains the results for the UPDRS rest

tremor scores predictions with the number of win-

dows with F0 between 3 Hz and 6 Hz for patient 6.

This table does not include the ﬁnal rounded predic-

tion values that were obtained through the division of

the number of windows by 200.

The tremor scores rounded down predictions and

the rest tremor from the clinical evaluation for patient

6 in the 4 hour time period are shown in table 7. In

this case, all predictions failed for the right forearm

and the only failed prediction for the left forearm was

in the Y axis.

This method’s success rate was determined based

on the percentage of times the prediction was correct.

The success rates for the X, Y, and Z axes, the magni-

tude, and the mean in the 4, 2, and 1 hour time inter-

vals are shown in table 8. The X axis had the highest

success rate for the 4 and 2 hour time intervals and the

Z axis had the highest success rate for the 1 hour inter-

val. Despite this, the difference between the success

rate of these axes for the later time interval was only

0.38%. Furthermore, the X axis had the overall high-

est success rate for the 4 hour interval, with 87.88%.

Predictions using the magnitude vector failed most of

the time, with the magnitude having the worst suc-

cess rate for every interval. The second worst success

rate in every interval was obtained by the Y axis. The

mean number of windows in all axes had a higher suc-

cess rate than the Z axis in the 4 hour interval, how-

ever the X and Z axes had a higher success rate for

both the 2 and 1 hour intervals.

Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation Using the Fundamental Frequency

315

Table 3: Features extracted from the periodograms and theirs description (Kuosmanen et al., 2020).

Feature Description

AUC Total power of the signal.

pv Maximum value of the PSD.

F0 Frequency of maximum power. Can be used to determine the dominant tremor type in

a window, according to the type of tremor frequency band it belongs to.

F50 Central point where the periodograms are divided into two equal parts.

SF50 Width of the frequency band centered in F50 that contains 68% of the total power.

|F50 − F0| Absolute value of the difference between F50 and F0.

tip Calculated by dividing pv by SF50.

Table 4: UPDRS scores and the corresponding number of

windows for the 4,2, and 1 hour intervals.

UPDRS Number of windows

score 4 hours 2 hours 1 hour

0 < 200 < 100 < 50

1 200 - 400 100 - 200 50 - 100

2 400 - 600 200 - 300 100 - 150

3 > 600 300 - 400 150 - 200

Table 5: Number of samples, mean number of windows and

mean predictions for each real score in the X axis with the

4 hour time interval.

Real

Nº samples

Mean

score Nº windows Prediction

0 53 114.472 0.572

1 7 225.429 1.127

2 4 463.25 2.316

3 2 442.5 2.212

4 DISCUSSION

4.1 Dataset

Firstly, we discuss how the characteristics of this

dataset might have inﬂuenced the results.

The majority of the dataset is composed of sub-

jects who were given a rest tremor score of 0 the

UPDRS clinical evaluation. Furthermore, the dataset

contains only seven instances where rest tremors was

attributed a score of 1, four instances where it was

attributed a rest tremor score of 2 and two instances

where it was attributed a rest tremor score of 3. As

such, there are few samples of data where rest tremor

has a higher score of 3 or 4.

The authors of (Channa et al., 2021) consider a

sampling frequency of 100 Hz sufﬁcient to measure

motor features related to PD due to the frequency of

tremor in the upper extremities being lower than 13

Hz. Nonetheless, despite some authors using higher

sampling frequencies, many studies consider a sam-

pling frequency of 50 Hz adequate for the detec-

tion of human activity with an accelerometer (San-

Segundo et al., 2020; Kuosmanen et al., 2020; Sigcha

et al., 2021). However, the sampling frequency of this

dataset is 31.25 Hz, being lower than what is typically

used in other studies.

4.2 Welch’s Periodograms

A brief visual analysis, presented in this subsection,

was performed on the PSD plots obtained from the

Welch’s method. The plots show the periodogram of

all windows in the respective axes or magnitude.

The periodograms from the left forearm of patient

5 (HC group; UPDRS rest tremor score 0 for left and

0 for right) are shown in ﬁgure 1. The majority of

peaks in these periodograms are between 0 and 3 Hz,

which is typically not associated with PD’s tremor.

The authors of (Kuosmanen et al., 2020) considered

that dyskinesia is observed in this frequency band,

in (San-Segundo et al., 2020) this frequency band

is associated with normal human movement, and in

(Shcherbak et al., 2023) the frequency band between

1 and 3 Hz is associated with bradykinesia. In addi-

tion, when subjects from the PD group have a UPDRS

rest tremor score 0, their periodograms have similar-

ities with the periodograms of subjects from the HC

group.

Figure 2 shows the periodograms from the right

forearm of patient 12 (PD group; UPDRS rest tremor

score 1 for left and 1 for right), where more plots with

peaks in the rest tremor frequency band between 3 and

6 Hz can be seen, especially on the Y axis.

In ﬁgures 3 and 4, the periodograms of the right

forearm of patient 17 (PD group; UPDRS rest tremor

score 0 for left and 3 for right) and the left forearm

of patient 36 (PD group; UPDRS rest tremor score 3

for left and 0 for right), respectively, are shown. It

was noted that, for the two instances of UPDRS rest

tremor score 3, the periodogram displayed distinctive

peaks in the rest tremor, postural tremor, and kinetic

tremor frequency bands. In the existing literature, dif-

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

316

Table 6: Analysis and prediction performed on the data on both forearms of patient 6 for the 4 hour time period.

Number of windows Prediction

X Y Z Magnitude Mean X Y Z Magnitude Mean

P6 left 537 297 488 436 440.67 2.68 1.48 2.44 2.18 2.20

P6 right 263 213 239 317 238.33 1.32 1.065 1.195 1.585 1.19

Table 7: Predicted and real UPDRS rest tremor scores for

patient 6 in a 4 hour time period.

UPDRS Prediction

score X Y Z Mag Mean

P6 left 2 2 1 2 2 2

P6 right 0 1 1 1 1 1

Patient 5 - Left

X Y

Z Mag

Figure 1: Welch’s periodogram for the X, Y, and Z axes and

magnitude (Mag) from the left forearm of patient 5.

ferent authors tend to use different frequency bands

for the types of tremor, despite using similar values.

The frequency bands considered for this study were

between 3 and 6 Hz for rest tremor, between 6 and 9

Hz for postural tremor, and between 9 and 12 Hz for

kinetic tremor.

Some of the plots from data associated with a

tremor score 2 had similar peaks in the tremor fre-

quency bands, like in ﬁgure 5. Since no periodograms

from subjects of the HC group, who only had scores

of 0, had similar peaks, these types of plots could be

associated with PwPD and/or a higher UPDRS rest

tremor score.

All graphs in this subsection have the same PSD

axis scale, except for the periodograms of patient 36,

shown in ﬁgure 4. This was because the bigger scale

used for patient 36, whose periodograms, especially

on the Y axis, had peaks with higher PSD values,

didn’t allow the peaks in the periodograms of other

patients to be visible.

Patient 12 - Right

X Y

Z Mag

Figure 2: Welch’s periodogram for the X, Y, and Z axes and

magnitude (Mag) from the right forearm of patient 12.

X Y

Z Mag

Patient 17 - Right

Figure 3: Welch’s periodogram for the X, Y, and Z axes and

magnitude (Mag) from the right forearm of patient 17.

4.3 Data Analysis

For this dataset and the data that was evaluated, the

X axis achieved the best results in the prediction of

UPDRS rest tremor scores with the proposed method.

The discrepancies in the mean number of windows

and mean prediction for the real score 3 in table 5 can

be due to the low amount of samples for this score.

Furthermore, for the two instances where the data was

associated with a UPDRS rest tremor score 3, this

method classiﬁed one instance correctly and failed in

the other, predicting a rest tremor score 1. This may

explain the lower mean number of windows and mean

prediction than expected.

In table 7, all predictions failed for the right fore-

Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation Using the Fundamental Frequency

317

Table 8: Success rate for all axes, the magnitude and the mean of the number of windows in the 4, 2, and 1 hour intervals.

Interval

Success rate

X Y Z Magnitude Mean

4 87.88% 71.21% 81.82% 34.85% 83.33%

2 82.58% 59.09% 81.82% 37.88% 78.79%

1 77.27% 59.09% 77.65% 39.39% 74.24%

X Y

Z Mag

Patient 36 - Left

Figure 4: Welch’s periodogram for the X, Y, and Z axes and

magnitude (Mag) from the left forearm of patient 36.

X Y

Z Mag

Patient 6 - Left

Figure 5: Welch’s periodogram for the X, Y, and Z axes and

magnitude (Mag) from the left forearm of patient 6.

arm. It is possible that, since the amplitude of tremor

and consequently its tremor score can vary during the

day, no tremor was present during the evaluation, but

it was present during the 4 hour time period that was

evaluated. Moreover, another possibility is that noise

in the signal might have affected these predictions.

The prediction values were rounded down to ob-

tain the ﬁnal score predictions, however the analysis

of the values with decimals allows us to know how

close the prediction was to the limit between scores

and by how much it failed. If all windows for the X

axis in the 4 hour time period (which had the high-

est success rate) that failed the prediction by less than

0.1 were discarded, the success rate would increase to

92.06%. Nonetheless, discarding only the windows

with a failed prediction is not feasible in the imple-

mentation of this method in a real setting where it is

not known which predictions are correct or incorrect.

If all windows within 0.1 of the limit between scores

were discarded and their prediction was marked as in-

conclusive between the two scores they are closest to,

a success rate of 91.23% would be achieved. Further-

more, the indication of which scores the inconclusive

prediction is in between could be helpful in a clini-

cal setting, helping the specialist know what were the

closest scores to the prediction. For example, win-

dows with values between 0.9 and 1, including these

values would be discarded and marked as inconclu-

sive between the scores 0 and 1.

It is important to note that tremor may not be con-

stantly present and its amplitude can vary during the

day. As such, the connection between the number of

windows with F0 within 3 and 6 Hz and the UPDRS

rest tremor score could lie in a potential connection

between the rest tremor amplitude (referred to as rest

tremor score in this study) and the constancy of rest

tremor. This possible connection should be further

explored in future work. The results in this study may

have also been affected by noise, due to evaluation

in free-living and in-the-wild conditions, and the low

number of samples for the UPDRS rest tremor scores

2 and 3. The number of window intervals correspond-

ing to which score were loosely established to explore

this method and could be improved in future work.

However, the method shows promising results, which

could be further improved by reﬁning the number of

windows intervals corresponding to each rest tremor

score, as previously mentioned, and other parts of the

method.

5 CONCLUSIONS

This study is focused on the rest tremor associated

with Parkinson’s disease and the estimation of its rest

tremor score according to the UPDRS using data from

accelerometers. In this paper we explore a method

based on the number of windows in a signal where the

fundamental frequency (F0) is in the rest tremor fre-

quency band, indicating that, when present, the dom-

ICT4AWE 2025 - 11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

318

inant type of tremor is the rest tremor. This method

to estimate the rest tremor scores had promising re-

sults and has a low computational cost. After being

further explored and reﬁned, this method has the po-

tential of being implemented for the clinical support

of rest tremor evaluation, helping monitor rest tremor

and the progression of the disease by providing data

for follow-up consultations. This method is limited by

the use of tremor scores given during the clinical eval-

uation in a controlled environment as labels for data

collected in free-living, which can lead to some data

being incorrectly labeled. Thus, the method should be

tested with other datasets to verify its results and with

sensor data collected during the clinical evaluations,

helping validate the method to be applied in a con-

trolled environment. Moreover, the results shown the

capability of this feature in the assessment of tremor

and its use with classical machine learning or deep

learning models for this task should be further ex-

plored.

REFERENCES

Adams, J. L., Dinesh, K., Snyder, C. W., Xiong, M.,

Tarolli, C. G., Sharma, S., Dorsey, E. R., and Sharma,

G. (2021). A real-world study of wearable sen-

sors in parkinson’s disease. npj Parkinson’s Disease,

7(1):106.

Amato, F., Saggio, G., Cesarini, V., Olmo, G., and Costan-

tini, G. (2023). Machine learning-and statistical-based

voice analysis of parkinson’s disease patients: A sur-

vey. Expert Systems with Applications, 219:119651.

Bazgir, O., Habibi, S. A. H., Palma, L., Pierleoni, P., and

Nafees, S. (2018). A classiﬁcation system for assess-

ment and home monitoring of tremor in patients with

parkinson’s disease. Journal of Medical Signals &

Sensors, 8(2):65–72.

Burq, M., Rainaldi, E., Ho, K. C., Chen, C., Bloem,

B. R., Evers, L. J. W., Helmich, R. C., Myers, L.,

Marks, W. J., and Kapur, R. (2022). Virtual exam

for parkinson’s disease enables frequent and reliable

remote measurements of motor function. npj Digital

Medicine, 5(1). Cited by: 6; All Open Access, Gold

Open Access, Green Open Access.

Channa, A., Ifrim, R.-C., Popescu, D., and Popescu, N.

(2021). A-wear bracelet for detection of hand tremor

and bradykinesia in parkinson’s patients. Sensors,

21(3):981.

de Oliveira Andrade, A., Paixao, A. P. S., Cabral, A. M.,

Rabelo, A. G., Luiz, L. M. D., Dion

ısio, V. C., Vieira,

M. F., Pereira, J. M., Rueda, A., Krishnan, S., et al.

(2020). Task-speciﬁc tremor quantiﬁcation in a clini-

cal setting for parkinson’s disease. Journal of Medical

and Biological Engineering, 40(6):821–850.

Goetz, C. G., Tilley, B. C., Shaftman, S. R., Stebbins, G. T.,

Fahn, S., Martinez-Martin, P., Poewe, W., Sampaio,

C., Stern, M. B., Dodel, R., et al. (2008). Movement

disorder society-sponsored revision of the uniﬁed

parkinson’s disease rating scale (mds-updrs): scale

presentation and clinimetric testing results. Movement

disorders: ofﬁcial journal of the Movement Disorder

Society, 23(15):2129–2170.

Huo, W., Angeles, P., Tai, Y. F., Pavese, N., Wilson, S.,

Hu, M. T., and Vaidyanathan, R. (2020). A heteroge-

neous sensing suite for multisymptom quantiﬁcation

of parkinson’s disease. IEEE Transactions on Neural

Systems and Rehabilitation Engineering, 28(6):1397–

1406.

Kuosmanen, E., Wolling, F., Vega, J., Kan, V., Nishiyama,

Y., Harper, S., Van Laerhoven, K., Hosio, S., Fer-

reira, D., et al. (2020). Smartphone-based monitor-

ing of parkinson disease: quasi-experimental study to

quantify hand tremor severity and medication effec-

tiveness. JMIR mHealth and uHealth, 8(11):e21543.

Pierleoni, P., Belli, A., Bazgir, O., Maurizi, L., Paniccia, M.,

and Palma, L. (2019). A smart inertial system for 24h

monitoring and classiﬁcation of tremor and freezing

of gait in parkinson’s disease. IEEE Sensors Journal,

19(23):11612–11623.

San-Segundo, R., Zhang, A., Cebulla, A., Panev, S., Tabor,

G., Stebbins, K., Massa, R. E., Whitford, A., De la

Torre, F., and Hodgins, J. (2020). Parkinson’s dis-

ease tremor detection in the wild using wearable ac-

celerometers. Sensors, 20(20):5817.

Shawen, N., O’Brien, M. K., Venkatesan, S., Lonini, L.,

Simuni, T., Hamilton, J. L., Ghaffari, R., Rogers, J. A.,

and Jayaraman, A. (2020). Role of data measurement

characteristics in the accurate detection of parkinson’s

disease symptoms using wearable sensors. Journal of

neuroengineering and rehabilitation, 17(1):1–14.

Shcherbak, A., Kovalenko, E., and Somov, A. (2023). De-

tection and classiﬁcation of early stages of parkinson’s

disease through wearable sensors and machine learn-

ing. IEEE Transactions on Instrumentation and Mea-

surement, 72:1–9.

Sigcha, L., Pav

on, I., Costa, N., Costa, S., Gago, M.,

Arezes, P., L

opez, J. M., and De Arcas, G. (2021). Au-

tomatic resting tremor assessment in parkinson’s dis-

ease using smartwatches and multitask convolutional

neural networks. Sensors, 21(1):291.

Smid, A., Elting, J. W. J., van Dijk, J. M. C., Otten, B.,

Oterdoom, D. M., Tamasi, K., Heida, T., van Laar, T.,

and Drost, G. (2022). Intraoperative quantiﬁcation of

mds-updrs tremor measurements using 3d accelerom-

etry: A pilot study. Journal of Clinical Medicine,

11(9):2275.

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.

IEEE Transactions on audio and electroacoustics,

15(2):70–73.

Zajki-Zechmeister, T., K

ogl, M., Kalsberger, K., Fran-

thal, S., Homayoon, N., Katschnig-Winter, P., Wen-

zel, K., Zajki-Zechmeister, L., and Schwingenschuh,

P. (2020). Quantiﬁcation of tremor severity with a mo-

bile tremor pen. Heliyon, 6(8):e04702.

Uniﬁed Parkinson’s Disease Rating Scale Rest Tremor Score Estimation Using the Fundamental Frequency

319