Parkinson’s Disease Detection Through Inertial Signals and

Posture Insights

Manuel Gil-Martín, Sergio Esteban-Romero, Fernando Fernández-Martínez and Rubén San-Segundo

Grupo de Tecnología del Habla y Aprendizaje Automático (T.H.A.U. Group), Information Processing and

Telecommunications Center, E.T.S.I. de Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain

Keywords: Parkinson’s Disease Detection, Inertial Signals, Fast Fourier Transform, Posture Insights, Lying, Sitting,

Convolutional Neural Networks.

Abstract: In the development of deep learning systems aimed at detecting Parkinson's Disease (PD) using inertial

sensors, some aspects could be essential to refine tremor detection methodologies in realistic scenarios. This

work analyses the effect of the subjects’ posture during tremor recordings and the required amount of data

to assess a proper PD detection in a Leave-One-Subject-Out Cross-Validation (LOSO CV) scenario. We

propose a deep learning architecture that learns a PD biomarker from accelerometer signals to classify

subjects between healthy and PD patients. This study uses the PD-BioStampRC21 dataset, containing

accelerometer recordings from healthy and PD participants equipped with five inertial sensors. An

increment of performance was obtained when using sitting windows compared to using lying windows for

Fast Fourier Transform (FFT) input signal domain. Moreover, using 5 minutes per subject could be

sufficient to properly evaluate the PD status of a patient without losing performance, reaching a window-

level accuracy of 77.71 ± 1.07 % and a user-level accuracy of 87.10 ± 11.80 %. Furthermore, a knowledge

transfer could be performed when training the system with sitting instances and testing with lying examples,

indicating that the sitting activity contains valuable information that allows an effective generalization to

lying instances.

1 INTRODUCTION

Biometrics research has experienced substantial

expansion in recent years, particularly finding

increased applications in the healthcare sector. The

scope of healthcare biometrics extends beyond

controlling access to electronic medical records and

patient identification; it encompasses medical

decision support tools designed for patient care.

These tools extract biomarkers that define patient

health, contributing to illness detection, analysis of

medication response, and the management of

chronic conditions such as Parkinson's Disease (PD).

PD is a neurodegenerative disorder characterized

by motor impairments like tremor, bradykinesia,

rigidity, and postural instability (Jankovic, 2008).

These impairments impact various motor functions,

including planning, programming, sequencing,

movement initiation, and execution (José, 1995).

Deep learning algorithms have being employed

for human motion recognition to model physical

activities using wearables or cameras (Manuel Gil-

Martin, San-Segundo, Fernandez-Martinez, &

Ferreiros-Lopez, 2020, 2021; Gil-Martín, San-

Segundo, Fernández-Martínez, & de Córdoba, 2020;

Zhang et al., 2017). Consequently, these

technologies can also be utilized to model tremor

movements associated with PD.

This work proposes a PD detection system based

on a deep learning architecture that allows analyzing

the effect of the subject’s posture performed while

recording the motion from inertial signals.

Additionally, this analyzes the recording time

required from each subject to evaluate the tremor

and distinguish between healthy people and PD

patients. The primary contributions of this research

are as follows:

 Analysis of the inertial signal domain and

sensors for PD detection.

 Assessment of different postures to detect PD

based on tremor symptom.

 Study of the required recording time to test a

patient and obtain an accurate detection.

1144

Gil-Martín, M., Esteban-Romero, S., Fernández-Martínez, F. and San-Segundo, R.

Parkinson’s Disease Detection Through Inertial Signals and Posture Insights.

DOI: 10.5220/0012451100003636

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

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 1144-1151

ISBN: 978-989-758-680-4; ISSN: 2184-433X

 Analysis of knowledge transfer for training

and testing the PD detection system using

different postures.

This paper is organized as follows. Section 2

reviews the literature of PD detection using inertial

sensors. Section 3 reviews the material and methods

used in this study, including a description of the

dataset, the signal processing, the deep neural

network, and the evaluation methodology. Section 4

describes the experiments and the obtained results.

Finally, section 5 summarizes the main conclusions

of the paper.

2 RELATED WORKS

Several researchers have explored the application of

machine learning for detecting motor symptoms

associated with PD through the use of wearable

sensors (Channa, Ifrim, Popescu, & Popescu, 2021;

Lang et al., 2019). However, there remain several

factors that could enhance PD detection systems in

real-world scenarios that could benefit the patients.

Concerning the extraction of features from

inertial signals, different features have used

proposed in previous works for PD detection based

on tremor. Most of these features are based on

measurements in the time domain (such as mean,

range, or cross-correlation) (Cole, Roy, De Luca,

Nawab, & Ieee, 2010; Garcia-Magarino, Medrano,

Plaza, & Olivan, 2016), in the frequency domain

(such as dominant frequency, energy content in a

particular band, or signal entropy) (Rigas et al.,

2012), or a combination of both domains (Dai,

Zhang, & Lueth, 2015). Moreover, other previous

works have concluded that features traditionally

used for speech processing (e.g., frequency analysis

using the Mel scale, cepstral coefficients) are also

effective in classifying human motion from

accelerometer data (San-Segundo, Manuel Montero,

Barra-Chicote, Fernandez, & Manuel Pardo, 2016;

San-Segundo, Navarro-Hellin, Torres-Sanchez,

Hodgins, & De la Torre, 2019; Vanrell, Milone, &

Rufiner, 2018).

As for tremor detection algorithms, previous

works have used a wide variety of machine learning

algorithms, such as decision trees (Garcia-Magarino

et al., 2016), random forests, hidden Markov models

(Rigas et al., 2012), and neural networks (Hathaliya

et al., 2022). For example a previous work

(Hathaliya et al., 2022) used a deep learning

architecture to model tremor obtaining a 92.4% of

accuracy using 6.4-second windows of raw samples

using a single sensor on the left anterior forearm.

However, the data distribution used in this work

seems to simulate a too optimistic scenario since

data from the same subjects were included in both

training and testing subsets and no distinction

between physical activities was performed. In

addition, there exists a lack of a study of the amount

of data required to properly assess unseen patients’

PD status.

Literature which mixes physical activity and PD

assessment is predominantly focused on

investigating whether an individual's likelihood of

developing PD is influenced by the extent of their

physical activity. Notably, prior studies have yielded

insights suggesting a correlation between higher

levels of physical activity and a lower incidence of

PD, particularly among women, with findings

underscoring the importance of these results in

strategic planning for interventions aimed at PD

prevention (Portugal et al., 2023). While the

literature has extensively explored the link between

overall physical activity and PD risk, a noticeable

gap exists in research focused on determining the

specific types of physical activities during which PD

detection is most discernible. Unlike general

physical activity assessments, postures offer a

unique perspective, as they involve more fixed

positions where tremors could become distinctly

noticeable, and other movements are less likely to

mask tremor signals in acceleration data.

This work proposes the use of a deep network for

both feature learning and tremor detection in a

realistic scenario and aims to analyse the effect of

different factors to develop a proper PD detection

system, such as the subjects’ posture or the test time

required per subject, rather than focusing solely on

obtaining the best detection performance. The

selection of an appropriate type and amount of data

collection could improve the overall assessment

during medical visits.

3 MATERIALS AND METHODS

This section includes information about the dataset

used in this work, the signal processing applied, and

the deep neural network used in the PD detection

system and the followed evaluation methodology.

3.1 Dataset

The PD-BioStampRC21 dataset (Adams et al., 2021;

Adams et al., 2017) comprises tri-axial

accelerometer data obtained from five wearable

sensors, encompassing participants with both

Parkinson’s Disease Detection Through Inertial Signals and Posture Insights

1145

Parkinson's disease (PD) and healthy controls. The

data collection utilized lightweight MC 10 BioStamp

RC sensors, with each participant wearing five

sensors affixed to specific body parts—chest, left

anterior thigh, right anterior thigh, left anterior

forearm, and right anterior forearm, as depicted in

Figure 1. The samples were acquired at a sampling

rate of 31.25 Hz. The dataset encompasses

recordings from 34 subjects: 17 healthy controls and

17 PD participants. Upon analysis, it was observed

that some sensors from control participants with IDs

007, 014, and 060 had missing data, prompting their

exclusion from the study.

Figure 1: A study participant wearing the sensors at five

different locations on the chest and each limb (Adams et

al., 2017).

3.2 Signal Pre-Processing

In this work, we used the information from each

inertial sensor isolated or from all together using two

possible input formats to feed the deep neural

networks: Raw data and Fast Fourier Transform

(FFT) magnitude coefficients. Moreover, we

analysed the amount of data from each user that we

need to properly assess his PD status: 1, 5, 10, and

15 minutes for each participant along with their

status in order to feed the classification system.

Initially, the recordings were segmented into

overlapping windows, with a shift equal to half the

window size between consecutive windows. All

windows from each participant were labelled as

either healthy control or PD based on the respective

participant's health status. The classification system

then categorizes each window as either belonging to

a healthy control or a person with PD. In this work,

we evaluated the classification performance when

considering a window size of 3.2 seconds

corresponding to 100 time samples. We obtained the

best classification performance using this window

size over this dataset in preliminary studies.

Next, for each window, we analysed time and

frequency domain signals as inputs for a deep neural

network, incorporating two distinct preprocessing

approaches based on the signal domain. For Raw

data, the original signal suffered no preprocessing,

and the inputs for the deep neural network consisted

directly of the time samples encompassed within

each window. For the FFT, the inputs comprised the

coefficients of the FFT magnitude. These

coefficients were computed in advance for each

analysis window, representing the spectrum from 0

Hz to 15.625 HZ (half of the sampling frequency in

the PD-BioStampRC21 dataset). We decided to

compute this input format because the energy in

tremor motion mostly concentrates in low

frequencies (M. Gil-Martin, Montero, & San-

Segundo, 2019). This paper analyses and compares

both alternatives for tremor modelling and detection.

In addition, this work is focused on analysing the

effect of the posture performed during the motion

recording in order to study which activity is better to

detect the tremor and generalize to new recordings.

We labelled the 3.2-second windows as ‘lying’ or

‘sitting’ using the information from the chest and

thigh sensors (Adams et al., 2017). For each

window, we determined the dominant axis for each

sensor (the axis direction along which the mean

acceleration was largest) and labelled the window

considering the orientation and location of the

sensors.

3.3 Deep Learning Architecture

The deep learning architecture used in this study is a

Convolutional Neural Network (CNN) consisting of

two main components: a feature learning subnet and

a classification subnet. The first subnet acquires

insights from raw data or FFT magnitude

coefficients extracted from inertial signals through

two convolutional layers (32 kernels of dimensions

(1, 5)) and two max-pooling layers (kernels of

dimensions (1, 2)). The second subnet uses fully

connected layers to categorize the learned features

into the predicted classes: a healthy person or a PD

patient. Dropout layers (0.3) were incorporated after

max-pooling and fully connected layers to prevent

overfitting during training. The final layer employes

a SoftMax activation function to provide predictions

for each class in every analysis frame, while

intermediate layers used ReLU to mitigate the

gradient vanishing effect. Categorical cross-entropy

serves as the loss metric, and the Adaptive Moment

Estimation (Adam) optimizer dynamically adjusted

the learning rate during training. The deep learning

structure was trained during 30 epochs and a batch

size of 100. Figure 2 illustrates the architecture

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1146

utilized in this study for modelling and classifying

analysis windows as either belonging to a healthy

person or a PD patient.

As depicted in the figure, the inputs of the CNN

are structured in a 2D matrix with dimensions N x

M. N represented the number of input signals, which

is 3 when utilizing a single sensor (X, Y, and Z

signals) or 15 when incorporating all five sensors

available in the dataset (3 x 5). M denotes the

number of analysed samples from each sensor

signal, depending on the signal domain in each

experiment. When using raw data as input, M is

equal to the size of the analysis window (100).

However, in the frequency domain, M represents the

number of FFT coefficients obtained from each

window, equating to half the window size (50).

3.4 Evaluation Methodology

In this work, a specific data distribution has been

used to create the most realistic scenario for a PD

detection system: the Leave-One-Subject-Out

(LOSO) Cross Validation (CV) strategy.

This strategy is a specific type of K-fold CV

where the system is evaluated with the data from one

subject and is trained with the data from the rest of

the K-1 subjects. In this case, the process is repeated

several times leaving a different subject for testing

and the results are also the average of the partial

results obtained for all repetitions.

This strategy avoids using recordings from the

same subjects in both training and testing subsets,

which pursues a more realistic scenario where a new

unseen patient’s data will be modelled without using

data from that subject. LOSO CV allows

generalizing to new, unseen subjects, while

capturing a wide variability of tremor motions from

the training subjects.

As evaluation metric, we used accuracy, which

defines the ratio between the number of correctly

classified examples and the number of total

examples. This way, for a classification problem

with N testing examples and C classes, accuracy is

defined in Equation (1).

Accurac



P











(1)

In addition, we used confidence intervals, which

include plausible values for a specific metric, to

show assure a significant difference between results

of two experiments (when their confidence intervals

do not overlap). Equation (2) represents the

computation of confidence intervals attached to a

specific metric value and N samples when the

confidence level is 95%.



95%



1.96



metric  100  metric 

(2)

In this study, we characterized PD tremor at the

window level, as the input examples for the deep

neural architecture were based on windows.

Nonetheless, we also presented performance at the

user level, where the prediction for an individual

was determined by the mode of predictions across

all the windows associated with that subject. This

methodology enables the incorporation of

information from all windows into a unified

prediction, offering a more holistic perspective from

a medical standpoint. Such an approach facilitates

the identification of overarching health patterns

instead of solely concentrating on the existence or

non-existence of tremors during short time intervals,

thereby mitigating the potential for incomplete or

inaccurate assessments.

Figure 2: Convolutional Neural Network Architecture for PD detection where N denotes the number of input signals (3 or

15) and M denotes the number of samples for each analysis window or example (100 or 50).

Parkinson’s Disease Detection Through Inertial Signals and Posture Insights

1147

4 RESULTS AND DISCUSSION

This section contains details about the experiments

performed in this work, including results and

discussion about the posture performed, the required

time for testing a subject and the possibility of

training and testing the system using recordings

from different postures.

4.1 Posture Insights

Regarding the posture insights, we decided to

evaluate the PD detection performance of the system

when using Lying and Sitting activities windows and

different sensors separately. Moreover, we analysed

the effect of the signal domain (Raw or FFT) over

this detection. Figure 3 shows a comparison of

performance at window-level when using 15

minutes per subject for different input signal

domains, performed activity, and sensor(s).

We observed a significant increment of

performance when using signals in the frequency

domain for most of the sensors when using each of

the activities. An increase in visibility of PD tremor

may be attributed to its intensified presence in the

frequency domain. Information regarding the energy

associated with the tremor frequency (between 3–9

Hz (Deuschl, Fietzek, Klebe, & Volkmann, 2003;

M. Gil-Martin et al., 2019)) and its harmonics can be

observed in the spectrum of the X, Y, and Z signals

recorded by the inertial sensor. Consequently, using

a CNN with FFT magnitude coefficients as inputs

has proven to achieve superior results when

compared to employing raw data samples directly.

Comparing the activities performed while

recording the data, we observed an increment of

performance when using sitting windows compared

to using lying windows for both input signal

domains (employing the same amount of data, i.e. 15

minutes per subject for these experiments). Sitting

activity emerges as a potentially more helpful setting

for detecting PD tremor using inertial sensors. This

may be attributed to the muscle engagement

necessary for maintaining an upright sitting position,

making tremors more pronounced, compared to a

relaxed lying posture. Furthermore, the sitting

posture offers a consistent and distinctive structure

across various subjects. Individuals tend to sustain

relatively fixed sitting positions, ensuring a uniform

and easily recognizable posture. In contrast, lying

down introduces, especially during sleeping,

postural changes, leading to significant alterations in

the representations along the x, y, and z axes of

inertial sensors. The standardization of sitting

posture stands in contrast to the variability in lying

posture, where alterations in body orientation during

sleep could hinder the maintenance of consistent

sensor data representations. When using 15 minutes

per subject, we obtained a maximum performance of

73.92 ± 0.65 % employing the FFT of sitting activity

and all the sensors, compared to 60.39 ± 0.73 %

when using the lying activity in the same setup.

Moreover, the exploration of isolated sensors,

both in the upper and lower limbs, presents a

promising avenue for the creation of biomarkers

associated with tremors manifesting in distinct parts

of the body. This nuanced analysis allows for a more

granular understanding of the tremor patterns

specific to each limb, potentially leading to the

development of targeted biomarkers. Such

biomarkers could offer valuable insights into the

severity and characteristics of tremors across

different body regions, as the Unified Parkinson's

Disease Rating Scale (UPDRS) assessment. Figure 3

also informs that chest sensor is the most

informative location to detect PD but the rest sensors

also achieve reasonable performance for the

classification task. However, since we obtained

better performance with all the sensors, we decided

to use all of them for the rest of experiments of this

study.

Figure 3: Accuracy at window-level using 15 minutes per

subject depending on the input signal domain, the activity

performed while recording the tremor and the sensor(s)

used.

4.2 Required Time for Testing a

Subject

Concerning the test time used from each subject to

build a proper PD detection biomarker, we decided

to analyse how much we could reduce the test time

used from each subject without losing significant

performance. We analysed 1, 5, 10, and 15 minutes

from each subject for testing the system. We kept

the 15 minutes per subject for training the system

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1148

(which corresponds to 465 minutes considering the

remaining 31 subjects).

Figure 4 shows a comparison of performance at

window-level when using 15 minutes per subject for

training for different input signal domains and

activity performed during recording when evaluating

unseen subjects using different amount of data (1, 5,

10 and 15 minutes). Figure 5 shows the same

comparison of performance at user-level. These

figures show that using 5 minutes from each subject

at testing stage would be enough to properly assess

unseen subjects’ PD status (accuracies of 77.71 ±

1.07 % at window-level and 87.10 ± 11.80 % at

user-level using the FFT while sitting). Despite of

the fact that assessing 1 minute could be sufficient

(accuracies of 74.72 ± 2.52 % at window-level and

70.97 ± 15.98 % at user-level using the FFT while

sitting) since there is no significant difference

between both results, using 5 minutes could offer a

more robust solution since 1 minute could be a short

interval during a patient could not manifest a PD

tremor.

Clinical visits intended to assess the progression

of PD are often constrained by their brevity, making

it challenging to draw conclusive and accurate

insights into PD detection. Contrary to this common

limitation, the findings presented in this paper

underscore a notable advance: 5 minutes of

recording proves sufficient for achieving a robust

PD detection. The results indicate that extending the

recording time from a new subject beyond this

threshold does not yield discernible improvements in

classification performance. This revelation

challenges the conventional notion of requiring long

recording periods, emphasizing the capability of a

concise data collection approach for accurate PD

detection.

Figure 4: Accuracy at window-level using 15 minutes per

subject for training depending on the input signal domain,

the activity performed and the time per subject used for

testing.

Figure 5: Accuracy at user-level using 15 minutes per

subject for training depending on the input signal domain,

the activity performed and the time per subject used for

testing.

4.3 Transfer Knowledge Between

Postures

In the pursuit of refining the robustness and

generalizability of a PD detection system, this work

also explored the idea of training the system with

data collected in a lying posture and subsequently

testing it with recordings from a sitting posture, and

vice versa. This way, we could inspect the capacity

of generalization across distinct postures by

knowledge transferring between lying and sitting

postures.

Figure 6 shows a comparison of performance

when using lying and sitting activities to train the

system (X axis) and to evaluate it (legend). In these

experiments, 15 minutes per subject were used for

training and 5 minutes per subject were used for

testing. In this figure, the columns of the same

colours are directly comparable because the testing

data are exactly the same. As a general comment, we

can say that there is not huge degradation in

performance. That means that the tremor appears in

the limbs involuntarily in different positions, but

there are significant differences. This way, we could

observe that when training a system with lying and

testing with sitting (green columns of left bars) the

performance drops compared to the scenario of also

training with sitting data (green columns of right

bars) for both input signal domains. This aspect

reflects that lying activity does not incorporate

sufficient information to generalize to sitting

instances. However, training a system with sitting

and testing with lying (yellow columns of right bars)

the performance remains similar compared to the

scenario of also training with lying data (yellow

columns of left bars). This aspect reveals that sitting

activity incorporates valuable information to

generalize to lying instances.

Parkinson’s Disease Detection Through Inertial Signals and Posture Insights

1149

Figure 6: Accuracy at window-level using 15 minutes per

subject for training and 5 minutes per subject for testing

depending on the input signal domain, the activities

performed during training (X axis) and testing (legend).

5 CONCLUSIONS

A broad data analysis in realistic scenarios is

necessary when detecting PD through a deep

learning system using inertial sensors to highlight

key factors to the refinement of tremor detection.

This work uses the PD-BioStampRC21 dataset

including healthy control and PD participants

wearing five inertial sensors to make an exhaustive

study concerning the posture performed during the

data collection.

Ensuring an appropriate distribution of data is

crucial in PD detection to prevent data overlap

between training and testing subsets and create

systems that could generalize to unseen subjects.

The LOSO CV technique emerges as a robust

solution, achieving model generalizability.

Sitting activity becomes a crucial recording

setting for detecting PD tremor using inertial

sensors. The standardization of sitting activity

among different subjects compared to lying activity,

especially during sleeping, could benefit the tremor

detection for unseen subjects. The proposed system

obtained an accuracy of 73.92 ± 0.65 % when using

15 minutes per subject from all sensors and the FFT

of sitting activity compared to 60.39 ± 0.73 % when

using lying activity.

Concerning the required amount of data from a

testing subject, we observed that using 5 minutes

while sitting could be sufficient to provide a robust

solution. This way, it is not necessary to record a

large amount of data from a patient to properly

assess his PD status.

When training a system with lying and testing

with sitting, there is a significant decrease in

performance compared to training and testing with

sitting data. This suggests that lying activity lacks

sufficient information to generalize to sitting

instances. However, when training with sitting and

testing with lying, the performance remains similar

to the scenario of training with lying data. This

indicates that sitting activity contains valuable

information that allows for effective generalization

to lying instances.

As future work, there is potential for further

refinement in the data analysis. More specifically,

enhancing the selection of windows characterized by

high energy levels could prove helpful in identifying

examples where tremors are more noticeable,

thereby enhancing the overall performance of PD

detection. Moreover, the creation of a regression

system capable of precisely estimating UPDRS

scores could provide valuable insights into the

disease progression. The incorporation of these

aspects could contribute to the development of more

effective diagnostic and monitoring tools for PD.

Regarding the limitations of this study, it is

relevant to remark that the PD detection proposed is

based on motion symptoms. Although these

symptoms appear in many patients, they do not

appear with the same intensity. The system proposed

can be completed with other AI-based system

extracting information from other signals like EEG.

ACKNOWLEDGEMENTS

The work was supported by the project “TremorDetect -

Detección de la enfermedad de Parkinson a través de

señales inerciales”, funded by “Primeros Proyectos” call

from ETSIT, UPM, by projects AMIC-PoC (PDC2021-

120846-C42), GOMINOLA (PID2020-118112RB-C21

and PID2020-118112RB-C22) and BeWord (PID2021-

126061OB-C43), supported by the Spanish Ministry of

Science and Innovation (MCIN/AEI/10.13039/

501100011033) and by the European Union

“NextGenerationEU/PRTR”, and ASTOUND (101071191

HORIZON-EIC-2021-PATHFINDERCHALLENGES-01)

funded by the European Commission.

REFERENCES

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

Tarolli, C. G., Sharma, S., . . . Sharma, G. (2021). A

real-world study of wearable sensors in Parkinson’s

disease. npj Parkinson's Disease, 7(1), 106.

doi:10.1038/s41531-021-00248-w

Adams, J. L., Dinesh, K., Xiong, M., Tarolli, C. G.,

Sharma, S., Sheth, N., . . . Sharma, G. (2017). Multiple

Wearable Sensors in Parkinson and Huntington

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

1150

Disease Individuals: A Pilot Study in Clinic and at

Home. Digital biomarkers, 1(1), 52-63.

doi:10.1159/000479018

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

(2021). A-WEAR Bracelet for Detection of Hand

Tremor and Bradykinesia in Parkinson's Patients.

Sensors, 21(3). doi:10.3390/s21030981

Cole, B. T., Roy, S. H., De Luca, C. J., Nawab, S. H., &

Ieee. (2010, 2010

Aug 30-Sep 04). Dynamic Neural Network Detection of

Tremor and Dyskinesia from Wearable Sensor Data.

Paper presented at the 32nd Annual International

Conference of the IEEE Engineering-in-Medicine-and-

Biology-Society (EMBC 10), Buenos Aires,

ARGENTINA.

Dai, H., Zhang, P., & Lueth, T. C. (2015). Quantitative

Assessment of Parkinsonian Tremor Based on an

Inertial Measurement Unit. Sensors, 15(10), 25055-

25071. doi:10.3390/s151025055

Deuschl, G., Fietzek, U., Klebe, S., & Volkmann, J.

(2003). Chapter 24 Clinical neurophysiology and

pathophysiology of Parkinsonian tremor. In M. Hallett

(Ed.), Handbook of Clinical Neurophysiology (Vol. 1,

pp. 377-396): Elsevier.

Garcia-Magarino, I., Medrano, C., Plaza, I., & Olivan, B.

(2016). A smartphone-based system for detecting hand

tremors in unconstrained environments. Personal and

Ubiquitous Computing, 20(6), 959-971. doi:10.

1007/s00779-016-0956-2

Gil-Martin, M., Montero, J. M., & San-Segundo, R.

(2019). Parkinson's Disease Detection from Drawing

Movements Using Convolutional Neural Networks.

Electronics, 8(8), 10. doi:10.3390/electronics8080907

Gil-Martin, M., San-Segundo, R., Fernandez-Martinez, F.,

& Ferreiros-Lopez, J. (2020). Improving physical

activity recognition using a new deep learning

architecture and post-processing techniques.

Engineering Applications of Artificial Intelligence, 92.

doi:10.1016/j.engappai.2020.103679

Gil-Martin, M., San-Segundo, R., Fernandez-Martinez, F.,

& Ferreiros-Lopez, J. (2021). Time Analysis in

Human Activity Recognition. Neural Processing

Letters. doi:10.1007/s11063-021-10611-w

Gil-Martín, M., San-Segundo, R., Fernández-Martínez, F.,

& de Córdoba, R. (2020). Human activity recognition

adapted to the type of movement. Computers &

Electrical Engineering, 88, 106822. doi:https://doi.

org/10.1016/j.compeleceng.2020.106822

Hathaliya, J. J., Modi, H., Gupta, R., Tanwar, S., Sharma,

P., & Sharma, R. (2022). Parkinson and essential

tremor classification to identify the patient?s risk

based on tremor severity. Computers & Electrical

Engineering, 101. doi:10.1016/j.compeleceng.

2022.107946

Jankovic, J. (2008). Parkinson’s disease: clinical features

and diagnosis. Journal of Neurology, Neurosurgery &

Psychiatry, 79(4), 368-376. doi:10.1136/jnnp.

2007.131045

José, L. C.-V. a. G. E. S. (1995). Effects of parkinsonism

on motor control. Life Sciences, 58(3), 165-176.

doi:https://doi.org/10.1016/0024-3205(95)02237-6

Lang, M., Pfister, F. M. J., Frohner, J., Abedinpour, K.,

Pichler, D., Fietzek, U., . . . Hirche, S. (2019). A

Multi-Layer Gaussian Process for Motor Symptom

Estimation in People With Parkinson's Disease. Ieee

Transactions on Biomedical Engineering, 66(11),

3038-3049. doi:10.1109/tbme.2019.2900002

Portugal, B., Artaud, F., Degaey, I., Roze, E., Fournier, A.,

Severi, G., . . . Elbaz, A. (2023). Association of

Physical Activity and Parkinson Disease in Women.

Long-term Follow-up of the E3N Cohort Study,

101(4), e386-e398. doi:10.1212/wnl.00000000002074

Rigas, G., Tzallas, A. T., Tsipouras, M. G., Bougia, P.,

Tripoliti, E. E., Baga, D., . . . Konitsiotis, S. (2012).

Assessment of Tremor Activity in the Parkinson's

Disease Using a Set of Wearable Sensors. Ieee

Transactions on Information Technology in

Biomedicine, 16(3), 478-487. doi:10.1109/titb.2011.

2182616

San-Segundo, R., Manuel Montero, J., Barra-Chicote, R.,

Fernandez, F., & Manuel Pardo, J. (2016). Feature

extraction from smartphone inertial signals for human

activity segmentation. Signal Processing, 120, 359-

372. doi:10.1016/j.sigpro.2015.09.029

San-Segundo, R., Navarro-Hellin, H., Torres-Sanchez, R.,

Hodgins, J., & De la Torre, F. (2019). Increasing

Robustness in the Detection of Freezing of Gait in

Parkinson's Disease. Electronics, 8(2). doi:10.

3390/electronics8020119

Vanrell, S. R., Milone, D. H., & Rufiner, H. L. (2018).

Assessment of Homomorphic Analysis for Human

Activity Recognition From Acceleration Signals. Ieee

Journal of Biomedical and Health Informatics, 22(4),

1001-1010. doi:10.1109/jbhi.2017.2722870

Zhang, S., Wei, Z., Nie, J., Huang, L., Wang, S., & Li, Z.

(2017). A Review on Human Activity Recognition

Using Vision-Based Method. Journal of Healthcare

Engineering, 2017. doi:10.1155/2017/3090343

Parkinson’s Disease Detection Through Inertial Signals and Posture Insights

1151