Evaluation of the Performance of Wearables’ Inertial Sensors for the

Diagnosis of Resting Tremor in Parkinson’s Disease

Carlos Polvorinos-Fernández

, Luis Sigcha

, Laura Pereira de Pablo

, Luigi Borzí

4,5 c

Paulo Cardoso

, Nelson Costa

, Susana Costa

, Juan Manuel López

Guillermo de Arcas

and Ignacio Pavón

Instrumentation and Applied Acoustics Research Group (I2A2), Mechanical Engineering Department, ETSI Industriales,

Universidad Politécnica de Madrid, Spain

Department of Physical Education and Sports Science, Health Research Institute, & Data-Driven Computer Engineering

(D2 iCE) Group, University of Limerick, Limerick, V94 T9PX, Ireland

ALGORITMI Research Center, School of Engineering, University of Minho, Guimarães, Portugal

PolitoBIOMed Lab–Biomedical Engineering Lab, Politecnico di Torino, 10129 Turin, Italy

ANTHEA Lab–Data Analytics and Technologies for Health Lab, Department of Control and Computer Engineering,

Politecnico di Torino, 10129 Turin, Italy

Instrumentation and Applied Acoustics Research Group (I2A2), Physical Electronics,

Electrical Engineering and Applied Physics Department, ETS de Ingeniería y Sistemas de Telecomunicación,

Universidad Politécnica de Madrid, Spain

Keywords: Motor Symptoms, Wearables, Accelerometer, Gyroscope, Machine Learning.

Abstract: Currently, objective monitoring of resting tremor in Parkinson’s disease (PD) involves wearable devices and

machine learning. Smartwatches may present an affordable method for remote and unintrusive tremor

monitoring. However, the development of optimized systems is necessary to perform accurate monitoring in

free-living settings. In this study, the potential of inertial sensors to detect resting tremors is evaluated. A

smartwatch was placed on the wrist of six subjects with PD during the execution of MDS-UPDRS motor tasks.

Data were collected over eight weeks from triaxial accelerometer and gyroscope simultaneously and used to

implement machine learning algorithms to detect resting tremor. The best performance (accuracy 97.0% in

tremor detection) was achieved using accelerometer data analysed with a Random Forest classifier, while the

gyroscope showed lower performance (93.0%). The results indicates that the use of the accelerometer in

commercial smartwatches can offer effective results for detecting resting tremors, while reducing

computational workload. These results show opportunities for the development of robust free-living tremor

monitoring systems using commodity devices and algorithms using a single sensor.

1 INTRODUCTION

Parkinson's disease is a neurodegenerative disease

affecting the central nervous system, leading to motor

and non-motor manifestations. PD occurs when

https://orcid.org/0000-0002-4594-9477

https://orcid.org/0000-0002-9968-5024

https://orcid.org/0000-0003-0875-6913

https://orcid.org/0000-0002-7924-0060

https://orcid.org/0000-0002-9348-8038

https://orcid.org/0000-0001-7440-8787

https://orcid.org/0000-0001-7847-8707

https://orcid.org/0000-0003-1699-7389

https://orcid.org/0000-0003-0970-0452

neurons do not produce enough of the chemical

"dopamine" (Wirdefeldt, Adami, Cole, Trichopoulos,

& Mandel, 2011).

Globally, 7–10 million individuals are currently

affected by PD, with an upward trend in recent years.

PD is rare before the age of 50 and exhibits a greater

820

Polvorinos-Fernández, C., Sigcha, L., Pereira de Pablo, L., Borzí, L., Cardoso, P., Costa, N., Costa, S., López, J., de Arcas, G. and Pavón, I.

Evaluation of the Performance of Wearables’ Inertial Sensors for the Diagnosis of Resting Tremor in Parkinson’s Disease.

DOI: 10.5220/0012571600003657

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

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 820-827

ISBN: 978-989-758-688-0; ISSN: 2184-4305

prevalence among men compared to women. The

incidence of PD increases with age, affecting around

1% of the population aged 60 or older (Rocca, 2018).

The development of the disease over time is

dependent on the person who suffers from it. During

the disease, patients progress through different stages,

associated with the severity of the symptoms and

physical disability caused. Currently, the diagnosis

and monitoring of the disease is conducted by a

medical specialist who assesses a series of exercises

performed by the patients using standardized

guidelines (Bhidayasiri & Martinez-Martin, 2017).

Among the motor symptoms in PD, the most

common and diagnostically distinct motor symptoms

is tremor (Halli-Tierney, Luker, & Carroll, 2020).

Tremor can be defined as an involuntary oscillatory

movement of parts of the human body, such as in the

hands or feet. There are different types of tremors

associated with PD, classified as resting tremor, that

presents when patients are relaxed; and action tremor,

which occurs when holding a position against gravity

or during any voluntary movement (Gironell, 2018).

The frequency range in which this type of tremor

manifest is in the range of 3.5-7 Hz (Salarian, et al.,

2003), while common human movements are usually

found in the 0-20 Hz band (Mannini, Intille,

Rosenberger, Sabatini, & Haskell, 2013).

Currently, levodopa is the principal drug used to

treat PD actively (LeWitt, 2008). It acts by converting

to dopamine in the brain and works vigorously on

controlling tremors in patients.

The subjective nature of motor assessment based

in observation techniques and the sporadic follow-up

commonly performed in clinical settings hinders the

implementation of precise therapies. For this reason,

the need of tools to improve the diagnosis and

continuous monitoring is still required.

The use of smart technologies for diseases such as

PD is currently on the rise. Wearable technologies,

stand out for their low cost, battery life, non-

invasiveness, can bring an excellent technological

support to implement monitoring systems for PD.

The use of inertial sensors such as accelerometers

and gyroscopes included in wearable devices with an

appropriate processing of these data and a subsequent

implementation of artificial intelligence algorithms

can be a promising alternative for the monitoring of

motor symptoms in PD in free-living conditions.

In this work, it has been evaluated which of the

inertial sensors integrated in a smartwatch,

accelerometer or gyroscope, could provide better

performance in terms of accuracy for the

classification of tremor in PD patients using machine

learning models. The dataset (Sigcha, et al., 2023)

used contains weekly records from several

Parkinson's disease patients during various planned

activities, while they were wearing a smartwatch.

2 BACKGROUND

Currently, the most used method for the assessment

of PD is the Movement Disorders Society's review of

the Unified Parkinson's Disease Rating Scale (MDS-

UPDRS) (Goetz, et al., 2008). Motor symptoms are

evaluated in the part III of this guide on a scale of 0

to 4, with 0 assigned to the non-existence of

symptoms and 4 the label for the most severe value.

Despite this scale is widely used, the evaluation

can be subjective by the physician and depends on his

or her perception at the time, which may vary from

one neurologist to another. This, together with the

fact that patients make very occasional visits to the

clinic, has led many authors to study the possibility of

remote and objective symptom monitoring.

Therefore, in recent years, numerous studies have

evaluated the possibility of using wearable devices

for healthcare applications. Some studies have

focused on the development of specific devices, while

others have used commercial devices for the

evaluation of PD pathologies (Sigcha, et al., 2023).

Regardless of how the monitoring has been

approached, MEMS (Micro Electronic Mechanical

Systems) type sensors have been used due to their

small size and low cost. The most common sensors

used in motor symptom monitoring are the

accelerometer and the gyroscope. A study conducted

by (San-Segundo, et al., 2020) used accelerometers to

compare tremor detection in free-living conditions

and in the laboratory environment, achieving a 10%

and 5% error. (López-Blanco, et al., 2019) conducted

one year of monitoring using a smartwatch that

collected data from a gyroscope yielded a Spearman

coefficient between the mean of the resting tremor

scores and smartwatch measurements was 0.81.

The combination of accelerometer and gyroscope

data for tremor detection was evaluated in (Sun, et al.,

2021), where a watch integrating both sensors was

developed, achieving an accuracy of over 94%.

Despite the progress in tremor monitoring,

previous studies have not focused on evaluating

which inertial sensor (accelerometer or gyroscope)

can provide more information to assess this symptom.

This paper will to study the potential of inertial

sensors in a commercial smartwatch to detect restring

tremor and evaluate which dataset, the one collected

from the accelerometer or the one obtained from the

gyroscope, could provide more useful information.

Evaluation of the Performance of Wearables’ Inertial Sensors for the Diagnosis of Resting Tremor in Parkinson’s Disease

821

3 MATERIALS AND METHODS

3.1 Data Collection

The data used in this study were collected during the

TECAPARK project (TECAPARK, n.d.), using a

proprietary m-health application named Monipar

(Sigcha, et al., 2023). A consumer-grade smartwatch

and a smartphone were used to monitor motor

symptoms in PD patients. The Monipar dataset

contains weekly records from Parkinson's disease

patients during planned activities, including

standardized exercises and resting periods for their

upper limbs, while they were wearing a smartwatch.

Data was collected from 6 PD patients (3 males/3

females, 64.2 ± 8.2 years). These subjects were in

early stages of the disease according to the Hoehn and

Yahr scale (Hoehn & Yahr, 1998) (H&Y = 1).

Three participants did not present tremors while

the other three presented tremors. A trained specialist

evaluated tremor according to MDS-UPDRS 3.17,

assigning score from 0 (no tremor) to 2 (mild tremor).

The data collection was conducted over 8 weeks,

and, during the study, all patients maintained their

usual medication regimen.

To perform tasks such as signal labeling,

preprocessing and feature extraction, MATLAB

software (R2017a) was employed. For the evaluation

and the training of the models, Python (3.6), and the

libraries Pandas, and Scikit learn were chosen.

3.2 Acquisition Device (Smartwatch)

A consumer-grade smartwatch was used as the data

acquisition device during the measurement sessions

and was placed on the wrist of the most affected side.

The wearable device was used to collect vibration

signals in the time domain using the bult-in inertial

sensors (accelerometer and gyroscope) in three axes.

In the case of the accelerometer, in m/s

, and, for the

gyroscope, in rad/s.

In this study, a smartwatch with dimensions of

46.6×51.8×12.9 mm and a weight of 32.5g was used,

with WearOs® as the operating system. This

smartwatch is equipped with an LSM6DS3 type

package, which includes a 3-axis digital gyroscope

and a 3-axis digital accelerometer. The triaxial

accelerometer has a maximum measurement

amplitude of ±2 g, while the triaxial gyroscope has a

measurement range of ±2000 dps.

The smartwatch was set to record data at a

sampling rate of 50 Hz. This frequency has been

established as it is appropriate for the analysis of

human movement, as common human movements are

usually found in the 0-20 Hz band, while it also

allows recording the typical PD tremors in the range

of 3.5-7 Hz (Salarian, et al., 2003).

3.3 Experimental Protocol

In each measurement session, each patient performed

8 exercises designed to assess the motor status,

including resting periods between exercises. These

exercises were conducted using Monipar application,

which guides the user through exercises by displaying

the tasks to be performed on the mobile screen.

In specific, each exercise belongs to the MDS-

UPDRS part III. The exercises proposed are related to

the amplitude of resting and postural tremor of the

hands, movement of the hands towards the chest,

finger tapping, hand movements, pronation-

supination of the hands, getting up and gait.

Each exercise has a different duration

(explanation plus execution); some take 15 seconds,

while others may require 50 seconds. Furthermore,

there is a 30-second break between exercises, making

7 minutes the approximate duration of each single

measurement session. Furthermore, each patient's

sessions were video recorded for subsequent labeling.

For the completion of this work, only the data

related to resting tremor amplitude, assessed through

section 3.17 of the MDS-UPDRS were used. In this

task, the patient should sit quietly in a chair with

hands resting on the armrest (not on the lap) and feet

resting comfortably on the floor, for 10 seconds,

without any other indication. Figure 1 shows the

interface of the resting tremor exercise.

Figure 1: Exercise explanation in mobile application.

3.4 Data Labelling

For data labelling, Monipar automatic generated

labels were used for each of the exercises. The

HEALTHINF 2024 - 17th International Conference on Health Informatics

822

sequence of exercises is numbered from 1 to 8

according to the order in which they are performed.

The resting tremor exercise corresponds to label 1.

For tremor labeling, data were automatically

labeled using thresholds according to magnitude

analysis in the tremor band (3.5-7.5 Hz). Then, these

labels were verified and corrected using video

recordings for each test. Data was labeled according

to the MDS-UPDRS section 3.17 guidelines. In

specific, the following were assigned to the data: 0

(Normal) if no tremor is observed, 1 (Slight) if the

maximum amplitude of the movement is less than 1

cm, 2 (Mild) if the maximum amplitude is between 1

and 3 cm, 3 (Moderate) if it is between 3, and 10 cm

and as 4 (Severe) if it is greater than 10 cm.

Figure 2 shows the distribution of tremor labels.

Only labels 0,1,2 are available, with 0 being the most

common label, present in 78% of the data.

Figure 2: Observations distributed by tremor label.

In this paper, the tremor label was used in two

ways. On the one hand, to classify among MDS-

UPDRS scores. And on the other hand, to

differentiate between the presence or not of tremor,

so labels 1 and 2 will be grouped into a single label.

3.5 Algorithmic Approach

This paper presents machine learning models that

predict the level of tremor amplitude using

accelerometer and gyroscope data to identify which

of them is more useful for prediction. This work was

developed following the schema shown in Figure 3.

Figure 3: Algorithm development diagram.

To train and evaluate the proposed models, the

signal obtained from the smartwatch has been

processed. First, the signal obtained from each of the

three axes of each sensor was combined into one by

means of Euclidean Norm according to equations (1)

and (2). This is since the inertial sensors embedded in

the wearable device can have a random orientation,

so this combination has been performed to avoid

errors. In addition, the computational load during

training and prediction can be reduced.

𝐴

𝑐𝑐𝑒𝑙 



𝑎𝑐𝑐𝑒𝑙





𝑎𝑐𝑐𝑒𝑙





𝑎𝑐𝑐𝑒𝑙





(1)

𝐺𝑦𝑟𝑜 



𝑔𝑦𝑟𝑜





𝑔𝑦𝑟𝑜





𝑔𝑦𝑟𝑜





(2)

After calculating the Eucliden norm the signal

was filtered using a Butterworth band-pass filter of

order 3 to select the signal between 0.5 and 10 Hz.

This frequency range is suitable for human activity

recognition and relevant for the tremor present in PD

(Khan, Hammerla, Mellor, & Plötz, 2016).

Following, signal segmentation was performed

using 128-sample windows (2.56 seconds) using 50%

overlap. A total of 5158 windows have been defined.

This combination of segmentation and overlapping is

recommended for PD tremor analysis (Patel, et al.,

2009) using inertial sensors.

To establish the tremor amplitude label in each of

the windows, if the label is repeated for more than

half of the observations in each window, the assigned

value will be that label. If this does not occur, such

windows will be excluded.

Finally, the signal was transformed to the

frequency domain using the Fast Fourier Transform

(FFT), since it has been shown (Ahlrichs & Samà,

2014) that the signal in the frequency domain could

be representative to evaluate the tremor.

Both time and frequency domain features will be

obtained. The time variables are the easiest to obtain

since do not entail a very high computational cost;

however, they do not lead to very robust conclusions

due to their difficult interpretation in this domain in

view of the tremors associated with human

movement. Nevertheless, the frequency variables

allow an improvement in the detection of the tremors

despite being more computationally complex to

acquire due to the need to calculate the FFT.

For each domain, the same type of features was

obtained. The Table 1 shows the extracted features

with a brief explanation of each of them. The database

used will be composed of 18 characteristics, 9 from

the time domain and 9 from the frequency domain.

Evaluation of the Performance of Wearables’ Inertial Sensors for the Diagnosis of Resting Tremor in Parkinson’s Disease

823

Table 1: Features extracted from the filtered accelerometer

and gyroscope signals in each domain.

Feature Description

Standard

Deviation

Returns the standard deviation of

the si

nals in each domain.

Mean

Calculates the median value of all

the measurements.

Median

Finds the median value of the

filtered si

nals in each domain.

Percentile 25

The 25th percentile of the input

data for each domain si

nal slice.

Percentile 75

The 75th percentile of the input

data for each domain si

nal slice.

Skewness

Asymmetry of the filtered signals

in each domain.

Max

Finds the maximum of the values

for each window in each domain.

Min

Finds the minimum of the values

for each window in each domain.

Entropy

Returns the entropy of the filtered

nals in each domain.

To these 18 features, those obtained from the FFT

of the signal must be added. In this case, 65 additional

features were obtained. So, a total of 83 features were

calculated for each of the 5158 defined windows.

For the development of machine learning models,

the database was divided using Hold Out Validation.

In this case, 80% (4126 windows) of the data for

algorithm training and 20% (1032 windows) for

algorithm validation. Although all measurements

were collected during the same task, since human

movement, especially PD, is different from one to

another, the train-test distribution has been

randomized among the entire dataset.

As the target variable is categorical, the models

proposed are classification models. In this study, the

following models are proposed: Gradient Boosting

(XGB), AdaBoost (ADAB), KNeighbours (KNN),

Random Forest (RF), Logistic Regression (LR) and

Decision Tree (TREE). Evaluation of the models was

performed using accuracy, sensitivity, specificity,

precision and F1-score metrics.

4 EXPERIMENTS AND RESULTS

This section presents the results obtained in the

present study. Multiple experiments were conducted

to evaluate and determine which sensors provide the

best performance. Section 4.1 presents the results

related to the evaluation using a binary classification

between tremor and non-tremor, obtaining the best

sensor with the results of the best model obtained.

Section 4.2 presents the results using the

classification of the tremor level thresholds according

to the MDS-UPDRS scale and establishing the best

sensor and model obtained.

4.1 Results of the Training of Binary

Models

The proposed classification models shown in Section

3.5 were implemented and trained using the set of

features extracted from the time and frequency

domains for each triaxial signal. In specific, two

different sets of features were extracted to each

inertial sensor (accelerometer and gyroscope).

In this case, as an unbalanced database is used,

metrics such as accuracy or precision, which measure

the proportion of correct predictions out of the total

number of predictions, can be misleading as they

provide a very high percentage of correct predictions,

but they could be only correct predictions of no

tremor. Thus, the study has focused on analyzing the

F1-score, because this metric combines precision and

recall using their harmonic mean, so a maximum F1-

score implies maximizing both precision and recall

simultaneously. Figure 4 show the training results

obtained, as reflected in the F1-score for each of the

models for the accelerometer and the gyroscope.

Figure 4: F1-score comparison for each algorithm using

accelerometer and gyroscope data.

From the observation of figures 4, it is evident that

results obtained from the data provided by the

accelerometer suppose a better performance of the

prediction model over those achieved through the

gyroscope. The average F1-score value of the models

trained with the accelerometer data was 0.86 while

that obtained with the gyroscope dataset is 0.70.

However, to determine which is the best

classification model obtained, all calculated metrics

were considered. Table 2 and Table 3 show the

training results for each of the models.

0.00

0.20

0.40

0.60

0.80

1.00

XGB ADAB KNN RF LR TREE

Accelerometer Gyroscope

HEALTHINF 2024 - 17th International Conference on Health Informatics

824

Among all the machine learning algorithms

implemented, it is observed that the Random Forest

algorithm offers the best metrics, and in particular the

balance of sensitivity and specificity is highlighted.

Table 2: Metrics obtained for each machine learning

algorithm from accelerometer data.

Accuracy

Sensitivity

Specificity

Precision

F1-score

XGB

0,94 0,86 0,96 0,87 0,87

ADAB

0,94 0,87 0,96 0,86 0,86

KNN

0,94 0,86 0,97 0,88 0,87

0,95 0,87 0,97 0,90 0,88

0,93 0,81 0,97 0,87 0,84

TREE

0,94 0,86 0,96 0,86 0,86

Table 3: Metrics obtained for each machine learning

algorithm from gyroscope data.

Accuracy

Sensitivity

Specificity

Precision

F1-score

XGB

0,88 0,51 0,99 0,92 0,65

ADAB

0,89 0,61 0,97 0,84 0,71

KNN

0,90 0,61 0,98 0,89 0,72

0,90 0,68 0,97 0,86 0,76

0,88 0,57 0,97 0,85 0,68

TREE

0,86 0,68 0,90 0,67 0,68

The test data, 20% of the total dataset, has been

evaluated with the best model, Random Forest, to

identify which of the two dataset yields better results.

The metrics are shown in Table 4 while the related

normalized confusion matrices are shown in Figure 5.

Table 4: Metrics associated with Random Forest algorithm

for accelerometer and gyroscope for test set.

[%]

Accelerometer Gyroscope

Accuracy 96.98 92.99

Sensitivity 91.06 75.07

Specificity 98.56 97.70

Precision 94.38 89.54

F1-score 92.69 81.67

It can be appreciated that both inertial sensors

have above 75% in all the metrics that were

considered. However, there are notable differences

based on the selected inertial sensor. For all these

metrics, the accelerometer database shows better

performance than the gyroscope. While both sensors

obtain quite similar specificity values, significant

differences are shown in the other metrics, with the

most significant difference in sensitivity, where the

accelerometer provides a value of 91% while the

gyroscope obtains a value of 75%.

It is noteworthy that the gyroscope data produces

an error of 25% predicting no tremor when it is

tremor. 91.06% sensitivity and 98.56% specificity for

the data provided by the accelerometer indicate a very

high accuracy rate in the prediction of resting tremors.

Acceleromete

roscope

Figure 5: Confusion matrices for accelerometer and

gyroscope for test set.

4.2 Results of the Training of

Multiclass Models

This experiment will approach the lines of resting

tremor prediction in a more qualitative way, leaving

behind binary classification. It has sought to evaluate

the performance of the algorithms by faithfully

predicting the tremor label assigned by the MDS-

UPDRS scale, using multiple classification.

In this case, the model that has been proposed is

the one that has achieved the best results in the binary

classification model. The Table 5 shows the metrics

obtained from the model trained using the data from

the different data sources.

Table 5: Comparison of the accuracy of each class for

Random Forest (multiclass classification).

Accelerometer Gyroscope

[%]

0 1 2 0 1 2

Accuracy

96.8 89.3

Sensitivity

98.6 88.5 93.6 97.9 54.7 60.2

Specificity

98.6 88.5 93.6 97.9 54.7 60.2

Precision

97.9 88.9 100 93.7 64.6 75.5

F1-score

98.2 88.7 96.7 95.8 59.2 67.0

Evaluation of the Performance of Wearables’ Inertial Sensors for the Diagnosis of Resting Tremor in Parkinson’s Disease

825

It can be noticed that the prediction of true resting

tremors of amplitude according to MDS-UPDRS

(scores 1 and 2) has a much higher hit rate with the

accelerometer data than by the gyroscope. For the

detection of no tremor (label 0), the values obtained

by the accelerometer and the gyroscope are quite

similar, however, it is again the dataset obtained from

the accelerometer that provides the best results.

From the observation of Figure 6, for the

gyroscope data, it is a difficult task to discern between

tremor and no tremor. It makes an error of 36%

predicting tremor 1 when it is really rest, and 38%

predicting tremor 2 when it is tremor 1. This is not the

case for the accelerometer, with very low percentages

of error between predictions for no tremor and

different level of tremor (11% and 6%, respectively).

Acceleromete

roscope

Figure 6: Confusion matrix for Random Forest model.

5 CONCLUSIONS

Monitoring PD individuals is crucial for precisely

tracking their progression and treatments. This study

seeks to optimize this process and aims to implement

efficient algorithms to facilitate the monitoring.

With the rise of technology in recent decades, this

study indicates that a commercial smartwatch can

provide useful data to monitor resting tremor in PD in

subjects. Several models for the prediction of resting

tremor were implemented using data provided by

inertial sensors embedded in a smartwatch during the

performance of eight standardized exercises.

The results suggest that the use of the

accelerometer as only inertial sensor can provide

optimal results for prediction of resting tremor. The

use of a single inertial sensor in wearables could help

improve the battery performance and power

consumption of the device, as well as reduce the

computational load needed for data.

However, it should be noted that this study has

certain limitations that need to be considered in future

projects. It has been worked with 6 PD patients for 8

weeks, which may be a small sample. In addition, the

database is unbalanced, with more than 75% of the

samples from the same label. So, a larger number of

measurement sessions would be necessary to increase

the reliability of the study.

The binary prediction of PD resting tremor

achieves its best hit rate using a Random Forest model

with the accelerometer data, obtaining 91.06 % in the

sensitivity metric and 98.56 % in the specificity

metric. These results are remarkably satisfactory for

the automatic detection of resting tremor, and are

similar to those proposed in (Sun, et al., 2021) and

(San-Segundo, et al., 2020), but with the advantage

that it has been achieved using a single sensor.

To find a prediction that fits better the true level

of resting tremor, an experiment has been conducted

in which the Random Forest algorithm was trained for

a multiple prediction that differentiates the resting

tremor collected according to the MDS-UPDRS

labelling guide. Based on the results, it is observed

that, using the data provided by the accelerometer, it

is possible to predict very reliably, with a sensitivity

and specificity rate of 98.6% in no presence of resting

tremor, 88.5 % for tremor with MDS-UPDRS score

1, and 93.6 % for resting tremor score 2, with the

biggest challenge for the algorithm being the

differentiation between no tremor and tremor score 1.

A multiclass classification gives a more specific

idea of the severity of the disease, and in a future real-

world application, would contribute to the clinician's

understanding and follow-up of the patient's data.

ACKNOWLEDGEMENTS

This research has been possible thanks to the

financing of the project BIOCLITE PID2021-

123708OB-I00, funded by MCIN/AEI/10.13039/

501100011033/ FEDER, EU; and by “Ayudas para

contratos predoctorales para la realización del

doctorado con mención internacional en sus escuelas,

facultad, centros e institutos de I+D+i”, funded by

Programa Propio I+D+i 2022 from Universidad

Politécnica de Madrid. The authors acknowledge to

the Physical Education and Sports Science (PESS)

department, the Health Research Institute (HRI), and

the Data-Driven Computer Engineering (D2iCE)

Group at University of Limerick, and the

Instrumentation and Applied Acoustics Research

Group (I2A2) at Universidad Politécnica de Madrid.

REFERENCES

Ahlrichs, C., & Samà, A. (2014). Is “frequency

distribution” enough to detect tremor in PD patients

HEALTHINF 2024 - 17th International Conference on Health Informatics

826

using a wrist worn accelerometer? Proceedings of the

8th International Conference on Pervasive Computing

Technologies for Healthcare, (pp. 65-71). Oldenburg,

Germany.

Bhidayasiri, R., & Martinez-Martin, P. (2017). Clinical

Assessments in Parkinson's Disease: Scales and

Monitoring. International Review of Neurobiology,

132: 129-182.

Gironell, A. P.-S.-L. (2018). Tremor types in Parkinson

disease: a descriptive study using a new classification.

Parkinson's Disease, 2018: 4327597.

Goetz, C., Tilley, B., Shaftman, S., Stebbins, G., Fahn, S.,

Martinez-Martin, P., Dodel, R. (2008). Movement

Disorder Society-sponsored revision of the Unified

Parkinson's Disease Rating Scale (MDS-UPDRS):

Scale presentation and clinimetric testing results.

Movement Disorders, 23(15): 2129-2170.

Halli-Tierney, A. D., Luker, J., & Carroll, D. G. (2020).

Parkinson disease. American family physician,

102(11):679-691.

Hoehn, M., & Yahr, M. (1998). Parkinsonism: Onset,

progression, and mortality. Neurology, 17(5):427-42.

Khan, A., Hammerla, N., Mellor, S., & Plötz, T. (2016).

Optimising sampling rates for accelerometer-based

human activity recognition. Pattern Recognit, 73: 33-

40.

LeWitt, P. A. (2008). Levodopa for the treatment of

Parkinson's disease. New Englad Journal of Medicine,

359(23):2468-76.

López-Blanco, R., Velasco, M., Méndez-Guerrero, A.,

Romero, J., Del Castillo, M. D., Serrano, I. R., &

Benito-León, J. (2019). Smartwatch for the analysis of

rest tremor in patients with Parkinson's disease. Journal

of the Neurological Sciences, 401:37-42.

Mannini, A., Intille, S., Rosenberger, M., Sabatini, A., &

Haskell, W. (2013). Activity Recognition Using a

Single Accelerometer Placed at the Wrist or Ankle.

Medicine & Science in Sports & Exercise, 45(11):2193-

203.

Patel, S., Lorincz, K., Hughes, R., Huggins, N., Growdon,

J., Standaert, D., Bonato, P. (2009). Monitoring Motor

Fluctuations in Patients With Parkinson's Disease

Using Wearable Sensors. IEEE Transactions on

Information Technology in Biomedicine, 13(6):864-73.

Rocca, W. (2018). The burden of Parkinson's disease: a

worldwide perspective. The Lancet Neurology, 17(11);

928-929.

Salarian, A., Russmann, H., Vingerhoets, F., Burkhard, P.,

Blanc, Y., Dehollain, C., & Aminian, K. (2003). An

ambulatory system to quantify bradykinesia and tremor

in Parkinson’s disease. Proceedings of the 4th

International IEEE EMBS Special Topic Conference on

Information Technology Applications in Biomedicine,

(pp. 35–38). Birmingham.

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

G., Stebbins, K., Hodgins, J. (2020). Parkinson’s

Disease Tremor Detection in the Wild Using Wearable

Accelerometers. Sensors, 20(20):5817.

Sigcha, L., Borzì, L., Amato, F., Rechichi, I., Ramos

Romero, C., Cárdenas, A., Olmo, G. (2023). Deep

learning and wearable sensors for the diagnosis and

monitoring of Parkinson’s disease: A systematic

review. Expert Systems with Applications, 229A,

1205412.

Sigcha, L., Pavón, I., De Arcas, G., Costa, N., Costa, S.,

Arezes, P., & Polvorinos-Fernández, C. (2023).

Monipar Database: smartwatch movement data to

monitor motor competency in subjects with Parkinson's

disease (1.0) [Data set]. Retrieved from Zenodo:

https://doi.org/10.5281/zenodo.8104853

Sigcha, L., Polvorinos-Fernández, C., Costa, N., Costa, S.,

Arezes, P., Gago, M., Pavón, I. (2023). Monipar:

Movement data collection tool to monitor motor

symptoms in Parkinson's disease using smartwatches

and smartphones. Frontiers in Neurology, 14.

Sun, M., Watson, A., Blackwell, G., Jung, W., Wang, S.,

Koltermann, K., Pretzer-Aboff, I. (2021).

TremorSense: Tremor Detection for Parkinson's

Disease Using Convolutional Neural Network.

IEEE/ACM Conference on Connected Health:

Applications, Systems and Engineering Technologies

(CHASE), (pp. 1-10). Washington, DC, USA.

TECAPARK. (n.d.). Retrieved from https://www.i2a2.upm.

es/tecapark/

Wirdefeldt, K., Adami, H., Cole, P., Trichopoulos, D., &

Mandel, J. (2011). Epidemiology and etiology of

Parkinson’s disease: a review of the evidence.

European Journal of Epidemiology, 26 Suppl 1:S1-58.

Evaluation of the Performance of Wearables’ Inertial Sensors for the Diagnosis of Resting Tremor in Parkinson’s Disease

827