A Hybrid Model based on Convolutional Neural Networks and Long

Short-term Memory for Rest Tremor Classiﬁcation

Jihen Fourati

1,2 a

, Mohamed Othmani

and Hela Ltiﬁ

1,4

National Engineering School of Sfax, University of Sfax, BP 1173, Sfax, Tunisia

Research Lab: Technology, Energy, and Innovative Materials Lab, Faculty of Sciences of Gafsa, University of Gafsa,

Tunisia

Faculty of Sciences of Gafsa, University of Gafsa, BP 2100,Gafsa, Tunisia

Research Groups in Intelligent Machines Lab ,BP 3038, Sfax, Tunisia

Keywords:

Resting Tremor, Deep Learning, Long-short Term Memory, Convolutional Neural Network, Parkinson’s

Disease.

Abstract:

Parkinson’s disease is a neurodegenerative disease, in which tremor is the main symptom. Deep brain stimula-

tion can help manage a broad range of neurological ailments such as Parkinson’s disease. It involves electrical

impulses delivered to speciﬁc targets in the brain, with the purpose of altering or modulating neural function-

ing. Security is playing a vital role in protecting healthcare gadgets from unauthorized access or modiﬁcation.

Our purpose is to adopt deep learning methodologies to classify resting tremors. To achieve this purpose,

a novel approach for resting tremor classiﬁcation in patients with Parkinson’s disease using a hybrid model

based on convolutional neural networks and long short-term memory is proposed. This research exploits the

high-level feature extraction of the convolutional neural network model and the potential capacity to capture

long-term dependencies of the long short-term memory model. The performed experiments demonstrate that

our proposed approach outperforms the best result for other state-of-the-art methods.

1 INTRODUCTION

Parkinson’s disease (PD) is a neurodegenerative

disorder, in which patients suffer different symp-

toms: resting tremor, akinesia and rigidity (LeesAJ,

2009)(Parkinson, 1817). Resting tremor (RT) is a

rhythmic and oscillatory involuntary movement that

appears in a body part (Abdo et al., 2010). It is the

roughest manifested symptom of Parkinson disease

tremors, which occurs at a frequency band between

4 and 6 Hz (Lyons and Pahwa, 2005) and disappears

when a voluntary movement is performed. Deep brain

stimulation (DBS) is an exceedingly used therapy op-

tion to lessen the motor signs of advanced PD. How-

ever, it has crucial security-related issues. The ability

to manipulate the pacemaker-like gadget enables per-

forming numerous randomized trials to assess the efﬁ-

ciency of the device (Rathore et al., 2019). Indeed, an

attacker can stop required stimulation or induce some

needless shocks in the cerebrum by fake signals pro-

duction (Choi et al., 2018). Deep brain stimulation

affects both tremor amplitude and tremor frequency

https://orcid.org/ 0000-0002-5499-5248

(Beuter et al., 2001). To quantify the tremor level for

PD subjects, (Pedrosa et al., 2018) developed two pre-

dictive models to classify Parkinson’s disease’s rest

tremor between high or low frequencies. The pro-

posed models have reached a classiﬁcation accuracy

of 92.8%. Furthermore, (Perumal and Sankar, 2016)

have studied the impact of using both gait and tremor

features for the early detection and monitoring of PD

by the use of statistical analysis and machine learning

techniques.

To enhance resting tremor detection, a multi-

characteristic classiﬁcation approach depend on the

characteristics of the local ﬁeld potentials (LFP) has

been used to recognize tremor-related features in

PD patients (Bakstein et al., 2012), and shows that

LFPs supplied enough information for detecting rest-

ing tremor. The authors of (Shah et al., 2018) pro-

posed a method base on frequency and time domain

combined with a logistic regression classiﬁer to de-

tect Parkinsonian rest tremors. Despite that, the de-

lay of tremor detection was not stated and is a sub-

stantial parameter for closed-loop DBS implementa-

tion. Additionally, (L

opez-Blanco et al., 2019) pro-

Fourati, J., Othmani, M. and Ltiﬁ, H.

A Hybrid Model based on Convolutional Neural Networks and Long Short-term Memory for Rest Tremor Classiﬁcation.

DOI: 10.5220/0010773600003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 75-82

ISBN: 978-989-758-547-0; ISSN: 2184-433X

posed an android application for PD tremor analysis

on a smartwatch. Their application experiments show

that this hardware has the prospect to quantify PD

subjects’ tremors impartially in the room of consul-

tation. The precise detection of tremor onset in PD

is crucial to the success of DBS therapy (Yao et al.,

2020). For that, (Yao et al., 2020) proposed a method

for detecting tremors during rest by the use of perti-

nent characteristics combined with machine learning

and Kalman. Moreover, (Hssayeni et al., 2019) de-

veloped two methods based on deep learning and gra-

dient boosting decision tree model used with wear-

able sensors to evaluate overall tremor parkinson. As

well, (Oktay and Kocer, 2020) presented a method for

the classiﬁcation of two types of tremors (Essential

tremor and Parkinson tremor) based on convolutional

long short-term memory. The experiments showed

that convolutional long short-term memory provides

successful results for differentiation of tremors. The

model reached a testing accuracy of 90%. In addition,

(Patel et al., 2009) developed a system to measure the

gravity of tremor, bradykinesia, and dyskinesia using

a wearable sensor platform. Further, (Salarian et al.,

2007) presented a method to detect tremors and com-

pare the tremor amplitude measurement to the cor-

responding uniﬁed parkinson’s disease rating scale

(UPDRS) score. Moreover, Edwards et al.(Edwards

and Beuter, 1999) used tremor features like the am-

plitude, frequency, and spectral power to differentiate

PD tremors. Besides, they combined a set features

into one variable to recognize a PD from abnormal

tremors effectively (Goldberger et al., 2000).

Although several studies have been conducted on the

resting tremor classiﬁcation, not much research has

been published about classifying different attack pat-

terns for deep brain implants. Until recently, only two

past attempts have been made to classify different at-

tack patterns for brain implants: (Rathore et al., 2019)

proposed a deep learning methodology for predict-

ing and forecasting different signal patterns of deep

brain stimulation. Typically, the rest tremor veloc-

ity is analyzed for evaluating the Parkinson tremor

intensity. Various attacks have been introduced in

the DBS context to simulate and distinguish between

false and authentic stimulations. Moreover, referring

to the security of signals from DBS, (Abdaoui et al.,

2020) designed a monitoring system for distinguish-

ing false alarms from legitimate ones and classiﬁed

the attacks using Raspberry Pi3 and deep learning.

They achieved an accuracy of 97% for predicting fake

signals. In this paper, we propose a novel classiﬁ-

cation approach, for those who are receiving DBS to

relieve tremors, using a hybrid model based on convo-

lutional neural networks and long short-term memory.

This paper studies the pattern of rest tremor velocity

(a type of feature observed to evaluate the intensity

of neurological disorders) based on the pattern of in-

troduced attack strategies. For this, we studied and

examined RTV values to design and train the neural

network.

The main contribution of this work is predicting

whether the signal is an attack or a genuine signal for

deep brain implants.

2 MATERIALS AND METHODS

2.1 Dataset

For tremor classiﬁcation, rest tremor in subjects with

Parkinson’s disease receiving chronic high frequency

electrical deep brain stimulation (DBS) was recorded

continuously throughout switching the deep brain

stimulator on (at an effective frequency) and off. Data

from Physionet online database (Goldberger et al.,

2000) were utilized, consisting of readings from the

experiments conducted on a group of 16 subjects with

PD. Neurophysiological data were acquired by em-

ploying a low-intensity laser that was directed to a

reﬂective piece of paper in the subject’s ﬁnger for a

time period of 60 seconds for data acquisition (Beuter

et al., 2001).

2.2 Data Acquisition

Data were collected using the MacLab data acquisi-

tion system and sampled at 100 Hz. Raw data were

exported to S-Plus for analysis and converted from

volts to mm/s (Beuter et al., 2001).

The data acquisition are carried out as the following

steps:

- The laser was placed at about 30 cm from the sub-

ject’s index ﬁnger tip.

- The laser beam was directed to a piece of reﬂective

tape placed on the ﬁnger tip.

- The velocity-transducer laser captures raw values

outputting voltage proportional to the velocity of the

ﬁnger for 60 seconds.

2.3 Data Preprocessing

Before dealing with the data, some signal manipu-

lations were needed. The ﬁles of the dataset (Gold-

berger et al., 2000) vary in tremor velocity units be-

tween patients. A few patients presented their data

in meters per second, while others presented in mil-

limeters per second. Thus, it is necessary to perform

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

normalization of the data, such that the posterior pro-

cessing and obtained results are identical among all

patients’ data. A simple formula was used to normal-

ize all data in millimeters per second.

mm/s = m/s × 1,000 (1)

As described by (Rathore et al., 2016), different types

of attack strategies can be employed by the attacker.

For that, we generated a new dataset by emulat-

ing different attack strategies along with modulating

learned stimulation patterns. The generated dataset

is composed of 4096 genuine and 4096 attacked se-

quences collected from real measurements. Each se-

quence contains 300 samples and one label indicating

whether the sequence is genuine or attacked.

3 PROPOSED APPROACH

In this section, we present the details of our proposed

approach, which contains convolutional and recurrent

neural networks. First, we give a general overview

of the approach. After that, we discuss each phase in

detail.

3.1 Overview

Our main objective is to address the problem of rest

tremor classiﬁcation in patients with Parkinson’s dis-

ease using a combination of convolutional neural net-

work (CNN) and long short-term memory (LSTM)

based on hybrid deep learning model. The ﬁrst step of

our proposed approach is to improve the data as men-

tioned in section 2. Then, we extract the time-domain

features of the RT data through a one-dimensional

convolutional neural network. Next, these features

are feeding into LSTM to extract the best represen-

tative features. For the last step, the classiﬁer predicts

whether the signal is an attack or a genuine signal.

The block diagram of our hybrid approach is depicted

in Fig.1.

Figure 1: Block diagram of CNN-LSTM hybrid model-

based RT classiﬁcation.

3.2 Convolutional Neural Network

Model

CNN is a well-known deep learning architecture in-

spired by the human neural system (Ko, 2018). It

identiﬁes automatically the relevant features without

any human supervision (Gu et al., 2018). A typi-

cal CNN architecture consists of alternating convo-

lutional layers and pooling layers, followed by one or

more fully connected layers. The convolutional layers

consist of multiple kernels stacked together that are

convolved with their input. It extracts the high-level

features from the input signal by a sliding-window

technique that outputs feature maps (Pak et al., 2018).

The pooling layer provides a typical downsampling

operation by applying the pooling operator to aggre-

gate information inside each small region of the input

feature channels and then select the most signiﬁcant

feature (Yamashita et al., 2018). These features are

fed to the fully connected layer which generates the

CNN model’s output data.

3.3 Long Short-term Memory Model

LSTM, a sophisticated version of recurrent neural

network which is able to learn long-term dependen-

cies, is designed to solve the long-term dependency

problem by means of short-term memory (Pak et al.,

2018). LSTM is able to process even the longest se-

quence data without vanishing the gradient (Pak et al.,

2018). Each LSTM unit is composed of a memory

cell and three main gates: input, output and forget.

The memory cell is designed to selectively add or re-

move information into/from this cell under the control

of these three gates (Bai and Tahmasebi, 2021). The

input gate is mathematically represented as following:

= σ(W

∗ [h

t−1

] + b

) (2)

The operator ‘*’ represents the element-wise multi-

plication of the vectors.

The information to be neglected from the previous

memory is controlled by forget gate which is math-

ematically deﬁned as following:

= σ(W

∗ [h

t−1

] + b

) (3)

The cell state is updated by the update gate, expressed

mathematically by the following equations:

ˇc

= tanh(W

∗ [h

t−1

] + b

) (4)

= f

∗ c

t−1

+ i

∗ ˇc

(5)

The hidden layer of the previous time step is updated

by the output gate which is also responsible for the

updating the output as it is given by:

= σ(W

∗ [h

t−1

] + b

) (6)

= o

∗ tanhc

(7)

A Hybrid Model based on Convolutional Neural Networks and Long Short-term Memory for Rest Tremor Classiﬁcation

3.4 Convolutional Neural Network and

Long Short-term Memory for Rest

Tremor Classiﬁcation

Artiﬁcial intelligence has caught the attention of the

scientiﬁc community in diverse ﬁelds such as intelli-

gent decision support systems (IDSS) (Ellouzi et al.,

2017) (Ellouzi et al., 2015), wavelets neural network

(Bellil et al., 2008), pattern recognition, function ap-

proximation optimization, etc. Deep learning is a sub-

set of machine learning where artiﬁcial neural net-

works, algorithms inspired by the human brain, learn

from large amounts of data. The proposed model, i.e.,

hybrid deep learning strategy, is inspired by biologi-

cal artiﬁcial neural networks.

The proposed model exploits the efﬁciency of

convolutional neural networks for extracting higher-

level features and learning the internal representation

of sequence data as well as the effectiveness of long

short-term memory layers for identifying short-term

and long-term dependencies. The main purpose of

our proposed model is to efﬁciently combine the ad-

vantages of these deep learning techniques. In this

study, a hybrid approach of CNN and LSTM has been

successfully used as a rest tremor classiﬁer. Our train-

ing phase inputs consist of the whole signals (normal

and under attack). The inputs go through one convo-

lution layer, one max-pooling layer followed by an-

other convolution layer. The convolution layers (1,

2) are convolved with their respective kernel size and

their ﬁlter number (2, 32). Between the block of con-

volution layers, the max-pooling layer is applied to

the feature maps. Thus, it was employed to reduce

the number of parameters to learn and to minimize

the amount of computation performed in the network.

After the block of convolution layers and the max-

pooling layer, two LSTM layers were applied with

recurrent activation function with a dropout and re-

current dropout of 0.2. The total unit of each LSTM

layer is 32. Nonlinearity is presented in the model

by offering some layers with the hyperbolic tangent

(tanh) activation function. The tanh function is the

updated version of the sigmoid function on the range,

which is a symmetric function centred on zero. Its

output is bounded, and it brings nonlinearity to the

neural network (Wang et al., 2020). In both LSTM

layers, we applied the “tanh” activation function. The

mathematical form of tanh function is as follows:

f (x) =

1 − e

−2x

1 + e

−2x

(8)

Where x is the input to a neuron. The details of each

layer parameter of the proposed rest tremor classiﬁca-

tion model are presented in Table 1. For the training

parameters of the proposed network, we adopted 80%

of samples for training and the remaining 20% for

testing. Figure 2 shows the proposed network model.

As we are facing a binary classiﬁcation problem (nor-

mal or under attack) and the output of our model is

a probability (In the end, we applied the classiﬁca-

tion layer using a sigmoid function.), the best choice

is to conﬁgure the model with adaptive moment es-

timation (ADAM) optimizer and the binary crossen-

tropy loss function. ADAM algorithm ensures that the

learning steps, during the training process, are scale-

invariant relative to the parameter gradients (Livieris

et al., 2020). Finally, the training and testing of the

proposed approach is done in 300 epochs and a batch

size of 64 samples. A hyperparameter tuning was per-

formed till the optimum performance was reached.

4 EXPERIMENTAL RESULTS

Our model was evaluated on a laptop computer with

a 4 GHz CPU, 64 GB of memory. All methods were

executed using Jupyter Notebooks (Python 3.7.10).

The proposed approach was tested on the Phys-

ionet dataset. Physionet ”tremor-DB” is made up of

the original rest tremor velocity signals, the ground

truths for benchmark the RT classiﬁcation. All data

recordings are considered genuine measurements. In

order to study the effect of deep brain stimulation on

amplitude and frequency characteristics of rest tremor

in Parkinson’s disease (Beuter et al., 2001), different

attack strategy was carried out. We have emulated dif-

ferent attack strategies introduced in the DBS frame-

work by changing the learnt stimulation pattern. To

improve the computational efﬁciency of the proposed

model, we generate a new dataset by creating faulty

signals that correspond to the attack patterns. The

dataset is composed of 4096 genuine and 4096 at-

tacked sequences collected from real measurements.

Each sequence contains 300 samples and one label

indicating if the sequence is genuine (label 0) or at-

tacked (label 1).

Cross-validation (CV) is a standardized test com-

monly used to test the ability of the classiﬁcation sys-

tem using various combinations of the testing and

training datasets. The proposed approach was trained

with the k-fold cross-validation procedure (k=10) by

using train data. This procedure divides arbitrarily

the set of observations into 10 approximately equal

folds. The ﬁrst fold is used as a validation set, and the

method ﬁts on the remaining k-1 folds.

We opted for using 10-fold cross-validation tech-

nique since this is one of the most using method for

estimating accuracy due to its relatively low bias and

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

Table 1: The architecture of our proposed model.

Layers No. Type Kernel size No. kernels/Units Output shape

Layer 1 Conv1D 2 32 2 × 32

Layer 2 MaxPooling1D 2 2 × 32

Layer 3 Conv1D 2 32 2 × 32

Layer 4 LSTM 32 32

Layer 5 LSTM 32 32

Layer 6 Dense 1 1

Figure 2: The proposed CNN-LSTM model.

Figure 3: Loss and accuracy curves (training and validation).

variance.

The cross-validation process was repeated until

every tremor dataset were included in the testing

dataset with 10-fold. The models obtained from the

training phase are tested section-by-section and then

averaged of recall, precision, accuracy, and F1-score

(given in Table 5). The 10-fold cross-validation pro-

cess results are shown in Table 2. The mean and stan-

dard values of the training accuracy and validation ac-

curacy after 10-fold cross-validation are evaluated as

97.01% (±0,0119898) and 94.90%(±0,0193959), re-

spectively. In Table 3, we summarize the average

and standard deviation for the calculated k-fold cross-

validation results process.

The best model results are obtained in epoch number

287. It takes 109 ms/step and leads to training loss

= 0.1760, training binary accuracy = 0.9701, valida-

tion loss = 0.1760, and validation binary accuracy =

0.9593. as shown in Fig. 3.

To compare our experiments with previous works,

we deﬁned four metrics: accuracy, recall, precision ,

and F1-score. The most frequent classiﬁcation evalu-

ation metric is accuracy. However, it can be mislead-

ing when the target classes are imbalanced. In such

A Hybrid Model based on Convolutional Neural Networks and Long Short-term Memory for Rest Tremor Classiﬁcation

Table 2: 10-fold cross-validation results.

Experimental trials Training accuracy Training loss Validation accuracy Validation loss

1 0.9630 0.1494 0.9277 0.2369

2 0.9690 0.1247 0.9444 0.1619

3 0.9815 0.1144 0.9628 0.1566

4 0.9690 0.1317 0.9630 0.1513

5 0.9444 0.2487 0.9442 0.2041

6 0.9815 0.1020 0.9649 0.1483

7 0.9815 0.0848 0.9669 0.1531

8 0.9815 0.1138 0.9628 0.1805

9 0.9669 0.1379 0.9074 0.3349

10 0.9630 0.1574 0.9463 0.1856

Table 3: Statistical analysis of 10-fold cross-validation pro-

cess.

Metrics Mean Standard deviation

Training accuracy 0.97013 0.01198

Training loss 0.13648 0.04499

Validation accuracy 0.94904 0.01936

Validation loss 0.19132 0.05772

Table 4: Confusion matrix of binary rest tremor classiﬁca-

tion problem.

Predicted class

Normal Under Attack

Actual Normal TP FN

class Under Attack FP TN

cases, other evaluation metrics should be considered

in addition to the accuracy. Recall , precision and F1-

score are excellent quantiﬁcation measures for binary

classiﬁcation.

The model classiﬁcation performance is best pre-

sented in a two-class confusion matrix consisting of a

2x2 matrix, with True Positives (TP), False Negatives

(FN), False Positives (FP), and True Negatives (TN)

described in Table 4.

Accuracy =

T P + T N

T P + T N + FP + FN

(9)

Recall =

T P

T P + FN

(10)

Precision =

T P

T P + FP

(11)

F1 − score = 2 ×

Recall × Precision

Recall + Precision

(12)

Table 5 illustrates the comparison between our

proposed approach and other previous studies. The

proposed approach achieved an accuracy, a recall, a

Figure 4: ROC curve. Class 0: Normal, Class 1: Under

attack.

precision, f1-score, and area under the receiver oper-

ating characteristic curve (AUC) of 94.9%, 96.42%,

90.0%, 93.0%, and,93.75% respectively. It can be

noted that the best accuracy, recall, and F1-score

were achieved by the proposed method, i.e., 94.9%,

96.42%, and 93.0%, which demonstrates its excellent

capacity for addressing the challenge. The area under

the curve reaches a high value.

Figure 4 shows the receiver operating characteris-

tic curve (ROC curve) of the classiﬁcation with rest-

ing tremor. The ROC curve plots between the val-

ues of the true positive rate (sensitivity) to the false

positive rate (1-speciﬁcity) at various classiﬁcation

thresholds.

The area under the curve was achieved as 93.75%,

which quantiﬁes the overall ability of the algorithm to

distinguish between a subject with normal rest tremor

velocity and a subject under attack.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

Table 5: Comparison results on Physionet dataset.

Methods Accuracy Precision Recall F1-score

LDA (Perumal and Sankar, 2016) 0.869 - - -

KNN+SVM (Pedrosa et al., 2018) 0.928 1 0.833 0.909

Ours 0.949 0.90 0.964 0.93

5 CONCLUSIONS

In this paper, a novel approach based on convolutional

neural networks and a long short-term memory hybrid

model for resting tremor classiﬁcation is presented.

The aim of this study was to exploit the high-level

feature extraction of the convolutional neural network

model and the potential capacity to capture long-term

dependencies of the long short-term memory. A com-

parison study is reported to demonstrate the perfor-

mance and the effectiveness of the novel proposed ap-

proach among the methods in previous literature. As

exhibited in experiments, the proposed approach out-

performs state-of-the-art methods in terms of recall,

accuracy, and F1-score.

REFERENCES

Abdaoui, A., Al-Ali, A., Riahi, A., Mohamed, A., Du, X.,

and Guizani, M. (2020). Secure medical treatment

with deep learning on embedded board. In Energy

Efﬁciency of Medical Devices and Healthcare Appli-

cations, pages 131–151. Elsevier.

Abdo, W. F., Van De Warrenburg, B. P., Burn, D. J., Quinn,

N. P., and Bloem, B. R. (2010). The clinical approach

to movement disorders. Nature Reviews Neurology,

6(1):29–37.

Bai, T. and Tahmasebi, P. (2021). Efﬁcient and data-driven

prediction of water breakthrough in subsurface sys-

tems using deep long short-term memory machine

learning. Computational Geosciences, 25(1):285–

297.

Bakstein, E., Burgess, J., Warwick, K., Ruiz, V., Aziz, T.,

and Stein, J. (2012). Parkinsonian tremor identiﬁca-

tion with multiple local ﬁeld potential feature classiﬁ-

cation. Journal of neuroscience methods, 209(2):320–

330.

Bellil, W., Othmani, M., Ben Amar, C., and Alimi, A.

(2008). A new algorithm for initialization and train-

ing of beta multi-library wavelets neural network. Ad-

vances in robotics, automation and control, I-Tech Ed-

ucation and Publishing, pages 199–220.

Beuter, A., Titcombe, M., Richer, F., Gross, C., and Guehl,

D. (2001). Effect of deep brain stimulation on am-

plitude and frequency characteristics of rest tremor in

parkinson’s disease. Thalamus & Related Systems,

1(3):203–211.

Choi, W., Lee, Y., Lee, D., Kim, H., Park, J. H., Kim, I. S.,

and Lee, D. H. (2018). Energy-aware key exchange

for securing implantable medical devices. Security

and Communication Networks, 2018.

Edwards, R. and Beuter, A. (1999). Indexes for identiﬁca-

tion of abnormal tremor using computer tremor eval-

uation systems. IEEE Transactions on biomedical en-

gineering, 46(7):895–898.

Ellouzi, H., Ltiﬁ, H., and Ayed, M. B. (2015). New multi-

agent architecture of visual intelligent decision sup-

port systems application in the medical ﬁeld. In 2015

IEEE/ACS 12th international conference of computer

systems and applications (AICCSA), pages 1–8. IEEE.

Ellouzi, H., Ltiﬁ, H., and Ben Ayed, M. (2017). Multi-agent

modelling of decision support systems based on visual

data mining. Multiagent and Grid Systems, 13(1):31–

45.

Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff,

J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,

G. B., Peng, C.-K., and Stanley, H. E. (2000). Phys-

iobank, physiotoolkit, and physionet: components of

a new research resource for complex physiologic sig-

nals. circulation, 101(23):e215–e220.

Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai, B.,

Liu, T., Wang, X., Wang, G., Cai, J., et al. (2018). Re-

cent advances in convolutional neural networks. Pat-

tern Recognition, 77:354–377.

Hssayeni, M. D., Jimenez-Shahed, J., Burack, M. A., and

Ghoraani, B. (2019). Wearable sensors for estima-

tion of parkinsonian tremor severity during free body

movements. Sensors, 19(19):4215.

Ko, B. C. (2018). A brief review of facial emotion recogni-

tion based on visual information. sensors, 18(2):401.

LeesAJ, H. (2009). Reveszt. parkinson’sdisease. Lancet,

373(9680):2055r2066.

Livieris, I. E., Pintelas, E., and Pintelas, P. (2020). A

cnn–lstm model for gold price time-series forecasting.

Neural computing and applications, 32(23):17351–

17360.

opez-Blanco, R., Velasco, M. A., M

endez-Guerrero, A.,

Romero, J. P., Del Castillo, M. D., Serrano, J. I., Ro-

con, E., and Benito-Le

on, J. (2019). Smartwatch for

the analysis of rest tremor in patients with parkinson’s

disease. Journal of the neurological sciences, 401:37–

42.

Lyons, K. E. and Pahwa, R. (2005). Handbook of essential

tremor and other tremor disorders. CRC Press.

Oktay, A. B. and Kocer, A. (2020). Differential diagno-

sis of parkinson and essential tremor with convolu-

A Hybrid Model based on Convolutional Neural Networks and Long Short-term Memory for Rest Tremor Classiﬁcation

tional lstm networks. Biomedical Signal Processing

and Control, 56:101683.

Pak, U., Kim, C., Ryu, U., Sok, K., and Pak, S. (2018).

A hybrid model based on convolutional neural net-

works and long short-term memory for ozone concen-

tration prediction. Air Quality, Atmosphere & Health,

11(8):883–895.

Parkinson, J. (1817). An essay on the shaking palsy Whit-

tingham and Rowland for Sherwood. London: Needly

and Jones.

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

J., Standaert, D., Akay, M., Dy, J., Welsh, M., and

Bonato, P. (2009). Monitoring motor ﬂuctuations in

patients with parkinson’s disease using wearable sen-

sors. IEEE transactions on information technology in

biomedicine, 13(6):864–873.

Pedrosa, T.

I., Vasconcelos, F. F., Medeiros, L., and Silva,

L. D. (2018). Machine learning application to quan-

tify the tremor level for parkinson’s disease patients.

Procedia computer science, 138:215–220.

Perumal, S. V. and Sankar, R. (2016). Gait and tremor as-

sessment for patients with parkinson’s disease using

wearable sensors. Ict Express, 2(4):168–174.

Rathore, H., Al-Ali, A. K., Mohamed, A., Du, X., and

Guizani, M. (2019). A novel deep learning strategy

for classifying different attack patterns for deep brain

implants. IEEE Access, 7:24154–24164.

Rathore, H., Badarla, V., and Shit, S. (2016). Consensus-

aware sociopsychological trust model for wireless

sensor networks. ACM Transactions on sensor net-

works (TOSN), 12(3):1–27.

Salarian, A., Russmann, H., Wider, C., Burkhard, P. R.,

Vingerhoets, F. J., and Aminian, K. (2007). Quan-

tiﬁcation of tremor and bradykinesia in parkinson’s

disease using a novel ambulatory monitoring sys-

tem. IEEE Transactions on biomedical engineering,

54(2):313–322.

Shah, S. A., Tinkhauser, G., Chen, C. C., Little, S., and

Brown, P. (2018). Parkinsonian tremor detection from

subthalamic nucleus local ﬁeld potentials for closed-

loop deep brain stimulation. In 2018 40th Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC), pages 2320–

2324. IEEE.

Wang, Y., Li, Y., Song, Y., and Rong, X. (2020). The inﬂu-

ence of the activation function in a convolution neural

network model of facial expression recognition. Ap-

plied Sciences, 10(5):1897.

Yamashita, R., Nishio, M., Do, R. K. G., and Togashi, K.

(2018). Convolutional neural networks: an overview

and application in radiology. Insights into imaging,

9(4):611–629.

Yao, L., Brown, P., and Shoaran, M. (2020). Improved de-

tection of parkinsonian resting tremor with feature en-

gineering and kalman ﬁltering. Clinical Neurophysi-

ology, 131(1):274–284.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence