Computers in Biology and Medicine 174 (2024) 108462

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

An efficient Parkinson’s disease detection framework: Leveraging

time-frequency representation and AlexNet convolutional neural network
Siuly Siuly a, b, *, Smith K. Khare c, Enamul Kabir d, Muhammad Tariq Sadiq e, Hua Wang a
a
Institute for Sustainable Industries & Liveable Cities, Victoria University, Melbourne, Australia
b
Centre for Health Research, University of Southern Queensland, Toowoomba, Australia
c
Mærsk Mc-Kinney Møller Institute, Faculty of Engineering, University of Southern Denmark, Denmark
d
School of Mathematics, Physics and Computing, University of Southern Queensland, Toowoomba, Australia
e
School of Computer Science and Electronic Engineering, University of Essex, Wivenhoe Park, Colchester, CO4 3SQ, United Kingdom

A R T I C L E I N F O A B S T R A C T

Keywords: Parkinson’s disease (PD) is a progressive neurodegenerative disorder affecting the quality of life of over 10
Parkinson’s disease detection million individuals worldwide. Early diagnosis is crucial for timely intervention and better patient outcomes.
Electroencephalogram signals Electroencephalogram (EEG) signals are commonly used for early PD diagnosis due to their potential in moni­
Time-frequency representation
toring disease progression. But traditional EEG-based methods lack exploration of brain regions that provide
Wavelet scattering transform
essential information about PD, and their performance falls short for real-time applications. To address these
AlexNet CNN
Feature extraction limitations, this study proposes a novel approach using a Time-Frequency Representation (TFR) based AlexNet
Convolutional Neural Network (CNN) model to explore EEG channel-based analysis and identify critical brain
regions efficiently diagnosing PD from EEG data. The Wavelet Scattering Transform (WST) is employed to
capture distinct temporal and spectral characteristics, while AlexNet CNN is utilized to detect complex spatial
patterns at different scales, accurately identifying intricate EEG patterns associated with PD. The experiment
results on two real-time EEG PD datasets: San Diego dataset and the Iowa dataset demonstrate that frontal and
central brain regions, including AF4 and AFz electrodes, contribute significantly to providing more represen­
tative features compared to other regions for PD detection. The proposed architecture achieves an impressive
accuracy of 99.84% for the San Diego dataset and 95.79% for the Iowa dataset, outperforming existing EEG-
based PD detection methods. The findings of this research will assist to create an essential technology for effi­
cient PD diagnosis, enhancing patient care and quality of life.

1. Introduction symptoms become increasingly severe, significantly impacting the daily

life and work of patients. The worsening symptoms lead to higher rates
Parkinson’s disease (PD) is a rapidly growing progressive neurode­ of disability and increased need for care [4]. Many people with PD also
generative condition that leads to significant disability and mortality develop dementia during the course of their disease.
rates. PD stands as the second most common neurodegenerative disor­ Although the exact causes of PD are not yet entirely known, genetics
der, following Alzheimer’s disease and the most prevalent movement [5], environment [6], ageing [7], and additional factors like inflam­
disorder in the world [1]. The condition primarily impacts the central mation, oxidative stress, and mitochondrial dysfunction [8] have all
nervous system, particularly the regions of the brain responsible for been linked to the condition’s onset. Age is a significant risk factor,
motor control. PD manifests with a diverse range of motor and being more frequent in individuals over 60 years old and men are more
non-motor symptoms. The motor symptoms include resting tremors, susceptible to PD than women. According to the Parkinson’s Founda­
bradykinesia (slowness of movement), rigidity (muscle stiffness), and tion, over 10 million people worldwide are estimated to live with Par­
postural instability leading to balance issues. Meanwhile, non-motor kinson’s disease, and this prevalence is projected to double by 2040 [4,
symptoms can encompass depression, anxiety, sleep disturbances, con­ 9]. As there is currently no cure for Parkinson’s disease, and it persists as
stipation, and cognitive changes [2,3]. When the disease progresses, the a lifelong condition, early detection, prompt diagnosis, and appropriate

* Corresponding author. Institute for Sustainable Industries & Liveable Cities, Victoria University, Melbourne, Australia.
E-mail address: [email protected] (S. Siuly).

https://doi.org/10.1016/j.compbiomed.2024.108462
Received 13 October 2023; Received in revised form 7 April 2024; Accepted 7 April 2024
Available online 9 April 2024
0010-4825/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

treatment are vital in enabling affected individuals to lead fulfilling under the curve (AUC) of 91%, Koch et al. [22] developed a Random
lives. These interventions enable individuals to maintain their regular Forest Classifier for PD detection using clinical and automated charac­
daily lives to the best extent possible. Therefore, researchers are actively teristics from EEG data. Two hybrid models for diagnosing PD were
exploring early detection methods to identify vulnerable patients and introduced by Shi et al. [23] with the highest accuracy of 82.89%. With
commence treatment as soon as possible, allowing them to lead as 96.9% accuracy. Lee et al. [24] developed a hybrid model combining
normal a life as possible within society. CNN and LSTM to take advantage of the spatial and temporal charac­
There is no definitive test for PD recognition. Diagnosis is typically teristics of EEG. Discrete wavelets transform (DWT)-based approach [3],
based on clinical observation of symptoms and a patient’s medical his­ smoothed pseudo-Wigner Ville distribution (SPWVD) coupled with CNN
tory. Neurological examinations and brain imaging techniques, such as based approach [6], deep-learning approach based on layer convolu­
Magnetic Resonance Imaging (MRI) and Positron emission tomography tional neural network [1] are some additional techniques that use PD
(PET) scans, Electroencephalography (EEG) may be used to rule out detection using EEG. Several popular deep learning models have been
other possible causes of symptoms [2]. Out of these, EEG holds signifi­ introduced in Refs. [25–29] to enhance performance, stability, and
cant importance in PD diagnosis as it provides a non-invasive means of decrease false discovery rate.
studying changes in brain activity associated with the condition [3]. The existing literature highlights certain gaps in the current methods
This valuable tool allows researchers to gain insights into the underlying for PD identification using EEG signals. Notably, the combination of
mechanisms of Parkinson’s disease and explore innovative treatments to time-frequency representation (TFR) with deep learning models, which
improve the quality of life for those affected by the disorder. In in­ is crucial for efficiently identifying PD, has not been extensively
dividuals with PD, EEG can reveal certain abnormalities in brain wave explored. EEG signals exhibit complexity and diversity, being non-
patterns. These abnormalities are indicative of disrupted brain function periodic, non-stationary, and non-linear, making TFR and deep
and can provide valuable clues about the presence and severity of the learning integration highly relevant. Additionally, the current research
disease. Abnormal EEG patterns may also help in tracking the progres­ has overlooked the investigation of the individual contributions of
sion of PD over time [2,3]. different EEG channels and the specific brain regions responsible for
In the literature, a variety of techniques for the automated diagnosis detecting PD as potential biomarkers. Understanding the involvement of
of PD using EEG signals have been proposed. These methods examined specific brain regions could aid in refining diagnostic accuracy.
computer-aided diagnostic systems that might automatically distinguish Furthermore, many studies in the field of EEG-based PD detection are
PD patients from healthy controls and learn the EEG characteristics that based on relatively small sample sizes and may lack external validation
distinguish PD patients from healthy controls. Li et al. [10] introduced on larger and more diverse datasets. In summary, the current literature
hybrid deep neural networks (DNNs), which construct parallel and series indicates the need for further research to address these limitations by
combination models to integrate convolutional neural networks (CNN) exploring TFR in combination with deep learning models, investigating
with long short-term memory (LSTM). EEG data from 25 patients with EEG channel contributions and brain region involvement, and validating
PD and 30 healthy controls were used to test the approach and their findings on more extensive and diverse datasets for enhanced PD
proposed parallel model achieved 97.6% specificity, 97.1% sensitivity, detection and diagnostic accuracy.
and 98.6% accuracy. A graph-based aspirin model for automatic PD To address the existing issues, this study intends to design a Time-
identification using EEG signals was presented by Baura et al. [11]. The Frequency Representation (TFR) based AlexNet network model for PD
suggested method was evaluated on 16 healthy patients (9 Female, 7 detection using EEG signals with the aim of achieving the following
Male), 15 PD (8 Female, 7 Male), and achieved the best classification three objectives: (1) Develop a computationally efficient and straight­
performance of 95.48%. In order to automatically decompose PD uti­ forward model for PD classification using EEG; (2) Investigate the effect
lizing EEG from 16 healthy controls (HC) and 15 PD (ON and OFF of different EEG channels in PD detection to localize the brain region
medication) subjects, Khare et al. [12] presented an automated tuneable responsible for selecting representative channels; (3) Enhance the per­
Q wavelet transform (A-TQWT) by extracting five significant charac­ formance of the proposed model compared to state-of-the-art methods.
teristics. The suggested technique correctly classified HC vs PD OFF and For the first time, this study is to introduce the Wavelet Scattering
HC vs PD ON medications with accuracy rates of 96.13% and 97.65%, Transform (WST), a method tailored to handle the oscillatory nature of
respectively. EEG signals effectively, in conjunction with the AlexNet model for ac­
A random-forest classifier was developed by De Oliveira et al. [13] to curate PD detection. In the proposed framework, WST is used to present
offer a machine-learning technique for PD identification that has a TFR for handling the oscillatory nature of EEG signals more precisely.
classification accuracy of 99.22%. Using EEG data from many research Additionally, the AlexNet model is utilized effectively in detecting PD
institutes, including healthy subjects and PD patients, Anjum et al. [14] leveraging the TFR information obtained from EEG data.
developed a system for identifying PD and achieved 85.40% accuracy. The proposed approach offers significant contributions, which can be
Using 20 healthy cases and 20 PD patients, Balestrino & Schapira [15] summarized as follows:
proposed a method based on the RUSBoosted trees classifier for PD
identification and achieved 87% accuracy. In order to extract high-order • We have designed a new framework for PD detection, integrating
features from EEG data for PD identification, Yuvaraj et al. [16] used a TFR and AlexNet AlexNet-CNN algorithms. This framework effi­
machine-learning strategy based on an SVM classifier and attained a ciently distinguishes between healthy control and PD subjects using
classification accuracy of 99.62%. Using a combination of magnetoen­ real-time EEG datasets. Notably, we are the first to employ the WST-
cephalography and EEG signals, Naghsh et al. [17] developed an auto­ based TFR in combination with AlexNet-CNN, and our approach has
mated method for diagnosing PD on 20 participants, 10 of whom had the been successfully tested on these two datasets.
disease and 10 of whom did not. They reached a 95% accuracy rate. A • Through our research, we have conducted a thorough channel-wise
13-layer CNN was developed by Oh et al. [18] using resting-state EEG, analysis of EEG data, leading to the localization of the brain region
and it successfully recognized de novo PD with an accuracy of 88.25%. A responsible for selecting the representative channel. This investiga­
tuneable Q wavelet transform, and a probabilistic neural network were tion provides valuable insights into the neural mechanisms under­
reported to be used by Murugappan et al. [19] to recognise PD. Wagh lying PD detection through EEG signals.
et al. [20] proposed an 8-layer graph-CNN that could identify PD and • We investigated the effectiveness of the AlexNet model by comparing
other neurological conditions with an accuracy of 85%. it with three widely used CNN models, VGG16, DarkNet19 and
For the purpose of PD identification, Xu et al. [21] developed a ResNet18, within the proposed framework to establish a robust PD
pooling based deep RNN technique and reported, in that order, 91.81% detection scheme.
specificity, 84.84% sensitivity, and 88.31% accuracy. With an area

2
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

• The proposed AlexNet based approach exhibits superior performance pre-processing, please refer to Refs. [3,31,32].
in PD detection from EEG data compared to existing state-of-the-art
EEG-based PD detection methods. The model’s improved accuracy
and efficacy represent a significant advancement in the field of EEG- 2.2. Proposed methodology
based PD diagnosis.
In this study, a novel framework is presented for identifying PD from
The structure of this paper is outlined as follows: Section 2 offers a EEG data, which involves a combination of Wavelet Scattering Trans­
comprehensive account of the EEG datasets utilized in this study and an form (WST) and AlexNet-based Convolutional Neural Network (CNN)
overview of the proposed PD detection framework. In Section 3, detailed model. The WST is used to obtain Time frequency representation (TFR)
information about the experimental setup, experimental results, and of EEG signal data. The proposed method comprises several essential
corresponding discussions is presented. Section 4 presents a discussion steps, including pre-processing, time-frequency representation, hidden
about the study and its findings. Lastly, Section 5 concludes the paper feature extraction, and classification. Fig. 1 visually demonstrates the
and includes insights into future research directions. design concept of the proposed framework in this research. Detailed
explanations of each phase within the proposed framework are provided
2. Material and methods in the following sections.

2.1. Datasets 2.2.1. Phase 1: time frequency representation using wavelet scattering
transform (WST)
Dataset 1 (Iowa dataset): This dataset was obtained from studies Time-frequency representation (TFR) is a valuable method for
conducted at the University of Iowa (UI) in Iowa City, Iowa [14,30]. This analyzing EEG signals, as it allows for simultaneous examination of both
dataset consists of EEG recordings from 14 individuals with PD and 14 the time and frequency domains. This capability is essential in detecting
healthy control (HC) subjects. The participants were not subjected to PD because the disease exhibits distinct temporal and spectral charac­
any exclusion criteria. However, the HC participants were carefully teristics that may not be readily discernible when studying each domain
matched with PD patients in terms of age and sex, and no significant independently. In this research, the Wavelet Scattering Transform
differences were observed in education or premorbid intelligence (WST) tool is employed to obtain the TFR of EEG signals, which effec­
(Table 1). During the experimental protocol, EEG recordings in the Iowa tively captures both local and global structures within the data. By
dataset were obtained only under eyes-open conditions. The EEG data leveraging TFR, we can better capture and identify the unique patterns
were captured using 0.1–100 Hz sintered Ag/AgCl electrodes on a associated with PD, enhancing our ability to diagnose and understand
64-channel Brain Vision system. The sampling rate used was 500 Hz, and the condition.
each subject’s EEG was recorded for at least 2 min. An online reference The WST builds upon the principles of wavelet analysis belonging the
was set to channel Pz as the baseline. Further details about these datasets family of wavelet transforms and operates on scattering coefficients
can be found in Ref. [14]. derived from wavelet transforms, modulus nonlinearities, and aver­
Dataset 2 (San Diego dataset): Dataset 2 is an openly accessible aging. It involves iteratively applying wavelet transforms to a signal at
dataset collected at the University of San Diego, California [31]. This various scales, computing the modulus of the resulting coefficients at
dataset comprises data from 16 healthy control participants (7 males each scale, and then recombining them to create a new representation of
and 9 females) with an average age of 63.5 ± 9.6 years, and 15 in­ the signal [33]. This resultant representation offers insights into the
dividuals diagnosed with PD (8 females and 7 males) with an average signal’s structure at different scales, making it advantageous for
age of 63.2 ± 8.2 years. The PD patients were carefully matched to the analyzing signals with multi-scale characteristics. This transform is
healthy control group in terms of right-handedness, gender, age, and particularly valuable when examining signals exhibiting intricate and
cognitive abilities, as assessed by the Mini-Mental State Exam and North multi-scale structures, as it enables the extraction of information at
American Adult Reading Test. All PD patients in the dataset had mild to multiple scales and resolutions.
moderate disease (Stage II and III on the Hoehn and Yahr scale), with an Instead of traditional WT, this study considers WST for EEG-based PD
average disease duration of 4.5 ± 3.5 years. The data collection involved detection for these reasons: (1) Better Time-Frequency Representation:
recordings on different days, with a counterbalanced order for PD pa­ It provides a more accurate representation of EEG signals, capturing
tients with medication ON (PDSO) and OFF (PDSF) conditions. For high and low-frequency components effectively, crucial for identifying
PDSO recordings, the subjects maintained their normal medication relevant patterns in PD-related EEG data. (2) Multiscale Analysis: It
schedule, while for PDSF recordings, the patients refrained from taking performs hierarchical decomposition across multiple scales, enabling
medication for at least 12 h. Healthy control subjects were tested only detection of subtle changes and abnormalities in EEG data associated
once. During data collection, the subjects were seated comfortably and with PD. (3) Reduced Computational Complexity: Compared to tradi­
asked to relax while fixating their eyes on a cross displayed on a screen. tional WT, WST has lower computational complexity, making it more
EEG data were recorded for a minimum of 3 min using a 32-channel efficient for processing large EEG datasets common in PD detection
Biosemi ActiveTwo EEG system at a sampling frequency of 512 Hz. studies.
For a more detailed description of the dataset, data acquisition, and The stages involved in the WST for generating scattering coefficients
are visually depicted in Fig. 2. Fig. 3 presents a tree overview diagram
illustrating the zero-order scattering coefficients obtained through
Table 1 averaging. In Fig. 3, f represents the input EEG signal, φJ is the scaling
Demographic information of PD and healthy participants. function, and {ψ j,k}are the wavelets, respectively. The features are
Category Dataset 1: Iowa dataset Dataset 2: San Diego iteratively generated by wavelet scattering transform. Firstly, input EEG
dataset data is convolved with scaling function (f*φJ) to generate the zeroth-
HC PD HC PD order scattering coefficients, S[0]. As a next step, following process is
repeated for each node [34].
Number of Participants 14 14 16 15
Gender (Male/female) 6/8 6/8 7/9 8/7
Age (years) 70.5 ± 8.7 70.5 ± 8.7 63.5 ± 9.6 63.2 ± 8.2 • Perform the WT of EEG data with each wavelet filter in the first filter
Disease duration (years) – 5.6 ± 3.2 – 4.5 ± 3.5 bank.
Sampling Frequency (Hz) 500 500 512 512 • Next, the modulus of each filtered output should be calculated. The
Used number of channel 64 64 32 32
nodes are the scalogram, U [1].

3
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Fig. 1. Overall framework of our proposed model for efficient identification of PD using EEG data.

• Utilizing the scaling filter, average each modulus. The outcomes are
the first-order scattering coefficients, S [1].

The details descriptions of WST are available in Refs. [33,34]. The

WST based TFR allows for the extraction of time-varying frequency in­
Fig. 2. Stages involved in scattering wavelet transform.
formation from EEG signals, enabling the detection of subtle changes
and patterns associated with PD.
Figs. 4 and 5 display the TFR acquired through the use of WST on
Dataset 1 and Dataset 2, respectively. By observing these figures, it

Fig. 3. Tree view of zeroth-order scattering coefficients computed by simple averaging of the input (https://au.mathworks.com/help/wavelet/ug/wavelet-scatte
ring.html).

Fig. 4. Exemplary TFR obtained using WST on Dataset 1 (a) HC and (b) PD.

4
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Fig. 5. Exemplary TFR obtained using WST on Dataset 2 (a) HC, (b) PDSF, and (c) PDSO.

becomes apparent that the TFR of healthy controls (HC) and individuals some cases, they are followed by max-pooling layers. The remaining
with PD on Dataset 1 exhibit distinct features. Similarly, the TFR of HC, three layers were fully connected. To facilitate parallel processing, the
individuals with PD in the “on” state (PDSF), and individuals with PD network is split into two copies, with each copy running on one GPU
in the “off” state (PDSO) on Dataset 2 also exhibit representative char­ [35]. The overall can be summarized as follows in Fig. 6:
acteristics. These findings highlight the ability of TFR to capture intri­ The five convolutional layers of the AlexNet CNN model are used to
cate details of EEG signals during PD, enabling the generation of extract the input image’s deep information [35,36]. The number and
distinguishable patterns for differentiating between HC and PD EEG size of the filters in the convolutional layer varies. For the study of deep
signals. features, the filter is moved around as a tensor or an image. The sub­
sequent input receives the extracted deep features and continues the
2.2.2. Phase 2: model architecture: AlexNet based CNN model extraction of deep features there. Three completely connected layers
Automated PD detection relies heavily on effective decision-making. that reduce the two-dimensional feature matrix to one dimension form
In this study, we propose an architecture for extracting important hid­ the CNN model’s final layer. Softmax layer, which allocates the class
den features from EEG data for PD detection that involves a TFR-based based on a probabilistic approach, is used to end the model. Fig. 7 de­
AlexNet CNN model. The main motivation behind choosing the AlexNet- picts the AlexNet model’s schematic structure. Table 2 provides an
based CNN model for EEG-based PD detection lies in its simplicity and overview of the CNN model’s tuning parameters for this proposed
proficiency in capturing complex spatial patterns within the EEG data. model.
Specific features of AlexNet, such as its deep architecture and convolu­
tional layers, make it well-suited for extracting hierarchical features 2.2.3. Phase 3: predicted outcome and decision-making
from EEG signals. Further details outlining the reasons are provided as In this phase, predicted outcomes (e.g., PD or HC classification) are
follows: (1) AlexNet presents a deeper architecture compared to other obtained through two fully connected layers, which constitute the final
CNN models, facilitating the capture of more intricate features within layer of the AlexNet CNN model and produce a probability distribution
the data. This increased depth empowers the model to acquire hierar­
chical representations, essential for comprehending the complex pat­
terns inherent in EEG signals associated with PD. (2) AlexNet
incorporates local response normalization, a technique enhancing the
model’s generalization by normalizing responses across various chan­
nels. This normalization mechanism enhances the model’s robustness to
variations and assists in capturing pertinent patterns within the EEG
signals. (3) The convolutional filters utilized in AlexNet are constructed
to identify spatial patterns at diverse scales. This characteristic proves
advantageous for capturing both local and global patterns within EEG
signals, a crucial aspect for precise identification of PD-related
abnormalities.
AlexNet is a pioneering convolutional neural network architecture. It
consists of eight layers, including five convolutional layers, and three
fully connected layers. The initial five layers are convolutional, and in
Fig. 7. Schematic diagram of AlexNet CNN model.

Fig. 6. Working steps of AlexNet model.

Where, CNN = convolutional layer (with ReLU activation); RN = local response normalization; MP = maxpooling; FC = fully connected layer (with ReLU activation);
Linear = fully connected layer (without activation); DO = dropout.

5
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Table 2
Summary of the tuning parameters of the proposed Alexnet model.
Name No. of filters/ Size of Stride Padding Size of feature Activation Weights Biases No. of learnable
neurons filter maps parameters

Input – – – – 227 x 227 x 3 – – – –

Convolution Layer 1 96 11 × 11 4 – 55 × 55 × 96 ReLU 34848 96 34944
MaxPooling 1 – 3×3 2 – 27 × 27 × 96 – 0 0 0
Convolution Layer 2 256 5×5 1 2 27 × 27 × 256 ReLU 614400 256 614656
MaxPooling 2 – 3×3 2 – 13 × 13 × 256 – 0 0 0
Convalution Layer 3 384 3×3 1 1 13 × 13 × 384 ReLU 884736 384 885120
Convalution Layer 4 384 3×3 1 1 13 × 13 × 384 ReLU 1327104 384 1327488
Convalution Layer 5 256 3×3 1 1 13 × 13 × 256 ReLU 884736 256 884992
MaxPooling 3 – 3×3 2 – 6 × 6 × 256 – 0 0 0
Dropout 1 Rate = 0.5 – – – 6 × 6 × 256 – 0 0 0
Fully connected layer 1 – – – – 4096 ReLU 37748736 7096 37752832
Dropout 2 Rate = 0.5 – – – 4096 – 0 0 0
Fully connected layer 2 – – – – 4096 ReLU 16777216 4096 16781312
Fully connected layer 3 – – – – 2 ReLU 8192 2 8194
Total learnable 58289538
parameters

over all possible categories for the given input (refer to Fig. 6). The final the most effective electrodes for the task. Previous works have typically
layer employs a softmax function, yielding output probabilities ranging explored the analysis of all channels, making it difficult to localize
from 0 to 1 for each class in the classification task. These output prob­ specific brain regions accurately. In this study, we propose a novel
abilities represent the predicted likelihood of the input signal belonging approach that focuses on individual channel analysis to improve brain
to a specific class, where higher probabilities indicate greater confidence region localization. Our methodology involves using the WST for time-
in the classification. Ultimately, based on these probabilities, decisions frequency analysis, allowing us to extract simultaneous information
are made regarding the number of subjects classified as PD and the about time and frequency characteristics. These time-frequency char­
number classified as HC. acteristics obtained from the WST are fed into the AlexNet CNN model,
which facilitates automatic extraction and classification of deep fea­
2.2.4. Phase 4: performance evaluation tures. To ensure effective analysis, we maintain a common experimental
The consistency of a decision-making model is determined by its setup.
overall performance. In the context of a medical expert decision-making For the WST, we set the quality factor to Refs. [1,2,4]. In this study,
system, multiple evaluations are conducted to assess the system the CNN models are evaluated using train-test split validation on the
comprehensively. To gauge the overall consistency of our developed dataset, where the data is divided into training and testing sets to assess
model, we have assessed four performance measures in addition to ac­ the model. we allocated 80% of the data for training, and the remaining
curacy. These measures include sensitivity (SEN), specificity (SPE), portion designated for testing or validation. To mitigate overfitting and
precision (PPV), and F-1 score. The mathematical expressions for these potential bias, we validated the model several times using ‘holdout
performance measures are defined as follows: validation’ technique. In this method, the dataset is randomly parti­
tioned into training and validation (or test) sets. The training set is
Pt + N t Pt
Accuracy = ; Sensitivity = ; utilized for model training, while the validation set (or test set) is
Pt + Pf + N t + N f Pt + N f
reserved for evaluating its performance. This holdout validation process,
Nt Pt involving random splits into train-test sets, is performed iteratively to
Specificity =
Pf + Nt
; Precision =
Pt + Pf enhance robustness against overfitting and bias. The model undergoes
five iterations with random splits, and the final results are obtained by
2 ∗ Precision ∗ Sensitivity averaging the testing accuracy across the five holdout validations. To
F − 1 score =
Precision + Sensitivity fine-tune the CNN model, we set the bias and weight learn factors to 20,
use a batch size of 16, and employ a learning rate of 1e-4. The maximum
where Pt , Pf, Nt, and Nf are true positive, false positive, true negative, epoch size is set to 15 to ensure optimal training. Through this approach,
and false negative values. we aim to overcome the challenges posed by EEG’s non-stationary na­
ture and limited spatial resolution while improving the accuracy of brain
3. Results region localization.

This study assessed the proposed TFR-based AlexNet CNN model

using two real-time PD EEG datasets: the Iowa dataset (dataset 1) and 3.2. Experimental results
the San Diego dataset (dataset 2). In this section, firstly we present a
brief overview of the experimental setup for the model implementation. PD is a complex neurological disorder, and its detection requires a
Subsequently, we present the results obtained from the experiments. comprehensive analysis of multiple brain regions and their EEG signals.
Finally, we provide a comparative analysis with existing methods in the Table 3 presents the channel-wise accuracy achieved on dataset 1 (Iowa
field, followed by a discussion of the findings. dataset) for detecting PD vs HC. Fig. 8 provides a comprehensive view of
the variations in channel-wise performance across each channel on
dataset 1. The findings reveal that channel AFz attained the highest
3.1. Experimental set up accuracy of 95.79%. Additionally, our developed model successfully
generated representative features in ‘Anterior Frontal midline regions’
The analysis of EEG signals poses challenges due to their non- (situated on the midline of the forehead), while the parietal region also
stationary nature. Despite its excellent temporal resolution, EEG suf­ exhibited significant deep features, contributing to its higher accuracies.
fers from limited spatial resolution. To address this limitation, many As mentioned before, Dataset 2 (San Diego dataset) consists of three
electrodes are used for analysis, but this raises the issue of determining conditions: PD patients with ON (PDSO) and OFF medication (PDSF), as

6
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Table 3 Table 4
Channel-Wise Accuracy (%) Achieved with the Proposed Model For HC vs PD On Channel-Wise Accuracy (In Percentage) Achieved with the Proposed Model For
Dataset 1. HC vs PDSF And HC vs PDSO On Dataset 2.
Channel Channel Overall Channel Channel Overall Channel Channel Overall Channel Channel Overall
Name accuracy Name accuracy Name accuracy Name accuracy

1 ’Fp1′ 88.89 33 ’AF3′ 89.67 1 ’Fp1′ 96.53 1 ’Fp1′ 97.52

2 ’Fz’ 90.42 34 ’AFz’ 95.79 2 ’AF3′ 92.73 2 ’AF3′ 94.22
3 ’F3′ 86.21 35 ’F1′ 92.72 3 ’F7′ 97.85 3 ’F7′ 98.18
4 ’F7′ 91.57 36 ’F5′ 85.44 4 ’F3′ 99.00 4 ’F3′ 98.84
5 ’FT9′ 84.29 37 ’FT7′ 88.51 5 ’FC1′ 94.71 5 ’FC1′ 97.85
6 ’FC5′ 86.21 38 ’FC3′ 77.01 6 ’FC5′ 97.85 6 ’FC5′ 96.53
7 ’FC1′ 88.89 39 ’C1′ 86.97 7 ’T7′ 91.74 7 ’T7′ 94.71
8 ’C3′ 84.67 40 ’C5′ 71.26 8 ’C3′ 93.72 8 ’C3′ 95.54
9 ’T7′ 78.93 41 ’TP7′ 83.52 9 ’CP1′ 94.22 9 ’CP1′ 97.52
10 ’TP9′ 82.38 42 ’CP3′ 83.14 10 ’CP5′ 98.18 10 ’CP5′ 96.86
11 ’CP5′ 75.86 43 ’P1′ 85.06 11 ’P7′ 93.39 11 ’P7′ 97.19
12 ’CP1′ 85.44 44 ’P5′ 86.59 12 ’P3′ 93.39 12 ’P3′ 98.84
13 ’P3′ 81.61 45 ’PO7′ 92.34 13 ’Pz’ 94.71 13 ’Pz’ 98.18
14 ’P7′ 85.44 46 ’PO3′ 81.99 14 ’PO3′ 90.09 14 ’PO3′ 97.52
15 ’O1′ 82.38 47 ’POz’ 93.49 15 ’O1′ 91.08 15 ’O1′ 95.58
16 ’Oz’ 87.74 48 ’PO4′ 85.06 16 ’Oz’ 92.24 16 ’Oz’ 96.86
17 ’O2′ 91.57 49 ’PO8′ 95.48 17 ’O2′ 92.24 17 ’O2′ 97.02
18 ’P4′ 92.72 50 ’P6′ 86.59 18 ’PO4′ 98.01 18 ’PO4′ 99.17
19 ’P8′ 91.95 51 ’P2′ 88.12 19 ’P4′ 98.18 19 ’P4′ 97.52
20 ’TP10′ 84.67 52 ’CPz’ 82.76 20 ’P8′ 98.18 20 ’P8′ 99.00
21 ’CP6′ 84.67 53 ’CP4′ 86.59 21 ’CP6′ 96.53 21 ’CP6′ 97.02
22 ’CP2′ 90.42 54 ’TP8′ 89.27 22 ’CP2′ 95.54 22 ’CP2′ 95.54
23 ’Cz’ 89.66 55 ’C6′ 93.10 23 ’C4′ 99.17 23 ’C4′ 97.85
24 ’C4′ 86.97 56 ’C2′ 86.59 24 ’T8′ 95.54 24 ’T8′ 97.19
25 ’T8′ 89.66 57 ’FC4′ 89.66 25 ’FC6′ 97.85 25 ’FC6′ 98.18
26 ’FT10′ 91.19 58 ’FT8′ 89.66 26 ’FC2′ 98.84 26 ’FC2′ 99.00
27 ’FC6′ 89.27 59 ’F6′ 89.27 27 ’F4′ 99.17 27 ’F4′ 99.17
28 ’FC2′ 85.82 60 ’AF8′ 85.44 28 ’F8′ 97.52 28 ’F8′ 99.84
29 ’F4′ 76.25 61 ’AF4′ 84.67 29 ’AF4′ 99.84 29 ’AF4′ 99.84
30 ’F8′ 83.14 62 ’F2′ 89.66 30 ’Fp2′ 97.85 30 ’Fp2′ 99.00
31 ’Fp2′ 85.06 63 ’FCz’ 86.97 31 ’Fz’ 98.18 31 ’Fz’ 99.17
32 ’AF7′ 87.36 32 ’Cz’ 97.85 32 ’Cz’ 99.17

Fig. 8. Visualization of channel-wise performance for dataset 1. Fig. 9. Visualization of channel-wise performance for dataset 2.

well as HC subjects. We conducted channel-wise accuracy analysis for or voltage maps, they visualize the electrical activity recorded by an EEG
PDSO vs HC and PDSF vs HC. The accuracy results for individual or other brain imaging techniques across different regions of the scalp.
channels in HC vs PDSF and HC vs PDSO on Dataset 2 are presented in Fig. 10 displays the topographic map obtained from dataset 1.
Table 4. Fig. 9 provides a visual representation of the varying perfor­ Similarly, Figs. 11 and 12 show the topographic maps of accuracy on
mances of each channel. Our analysis reveals that our developed model Dataset 2 for HC vs PDSF and HC vs PDSO, respectively. These figures
achieved the highest classification accuracy of 99.84% for channel AF4 reveal that in the detection of PD, frontal and central regions play a
in both HC vs PDSO and HC vs PDSF comparisons. AF4 denotes the crucial role in generating the representative characteristics of TFR. Thus,
“Anterior Frontal-right” location which is located on the right side of the utilizing the topographic maps obtained from accuracy, we have effec­
forehead, adjacent to AFz. It is important to note that both datasets tively demonstrated the role of various brain regions in PD detection.
demonstrate that channel locations in the central and frontal regions (e. Finally, we have plotted the topographic maps obtained on our devel­
g. AF4 and AFz) exhibit more representative features when compared to oped model to determine which portion of the brain has generated
other regions. Furthermore, the results demonstrate the effectiveness of representative characteristics for PD detection.
our developed technique in tracking changes during ON and OFF To achieve intense understandings into our developed model, we
medication phases of PD patients compared to HC subjects. assessed its performance using four key metrics. Table 5 presents the
In order to gain deeper insights into the brain regions most respon­ performance parameters obtained for
sible for PD detection and the generation of representative characteris­ Dataset 1 and dataset 2 using our proposed model. The analysis in­
tics, we generated topographic maps for both datasets. These maps play dicates that on Dataset 1, the model achieved a sensitivity (SEN) of
a significant role in understanding PD by providing valuable insights 96.09%, specificity (SPE) of 95.49%, positive predictive value (PPV) of
into the spatial distribution of brain activity. Referred to as scalp maps 95.35%, and F1 score of 95.42%. On Dataset 2, for the classification of

7
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Fig. 10. Topographic map of HC vs PD obtained for dataset 1.

Fig. 12. Topographic map of HC vs PDSO obtained for dataset 2.

Table 5
Performance parameters (percentage) of dataset 1 and dataset 2 using our
proposed AlexNet model.
Performance HC vs PDSO HC vs PDSF HC vs PD
Parameters (dataset 2) (dataset 2) (dataset 1)

Accuracy 99.83 99.83 95.79

Specificity 100.00 99.67 95.49
Sensitivity 99.67 100.00 96.09
Precision 99.67 100.00 95.35
F1-score 99.84 99.84 95.42

Table 6
Percentage Confusion Matrix for HC vs PD On
Dataset 1.

Fig. 11. Topographic map of HC vs PDSF obtained for dataset 2.

PDSF vs HC, our model achieved an SEN of 100%, SPE of 99.67%, PPV of
Table 7
100%, and F1 score of 99.84%. Meanwhile, for the classification of Percentage Confusion Matrix for HC vs PDSF
PDSO vs HC, our model obtained an SPE of 100%, SEN of 99.67%, PPV And HC vs PDSO On Dataset 2.
of 99.67%, and F1 score of 99.84%. These results demonstrate the
consistency and effectiveness of our developed model in accurately
detecting PD and HC EEG signals.
In order to determine the individual detection rates of HC and PD for
two datasets, we computed the confusion matrix. The results are pre­
sented in Table 6, showing the percentage accuracy obtained for HC and
PD classification on dataset 1.
Our model achieved an accuracy of 95.35% for PD detection and
96.21% for HC detection. Similarly, in the classification of PDSF, PDSO,
and HC on dataset 2, the confusion matrix results are displayed in
Table 7. The model demonstrated 100% accuracy in detecting HC and
PDSO, while achieving an accuracy of 99.67% for PDSF detection. These
findings highlight the effectiveness of our developed model in accurately performance of a binary classification model, such as those used for
classifying PD and HC subjects in both datasets. distinguishing between PD and healthy control subjects. The ROC and
Furthermore, we also assessed the performance of our proposed area under the curve (AUC) for our developed model on Dataset 1 is
model based on receiver operating characteristics (ROC) as it is a crucial shown in Fig. 13. As evident from the Figure, our developed model has
tool in the disease detection and diagnosis. It is widely used to assess the

8
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Fig. 13. ROC and AUC obtained for HC vs PD classification on Dataset 1. Fig. 15. ROC and AUC obtained for HC vs PDSO classification on Dataset 2.

WST was employed to capture distinctive temporal and spectral char­

acteristics, while the AlexNet CNN recognized complex spatial patterns
across various scales in the proposed framework. This integration
facilitated the precise identification of intricate EEG patterns associated
with PD, with the primary goal of developing a framework that effec­
tively differentiates between patients with PD and HC subjects using
EEG signal data. The framework also aims to distinguish individuals
with PD (off medication) and those with PD (on medication) from HC
subjects. The study conducted an investigation to pinpoint the most
influential EEG channels and brain regions providing significant infor­
mation for the effective identification of PD patients. The research in­
cludes a multi-pattern analysis, encompassing channel-wise
performance examination and brain region investigation through
topographic maps, achieving high performance in the effective identi­
fication of PD patients through comparative analysis. To assess the
framework’s effectiveness, tests were conducted on two publicly avail­
able datasets: the Iowa dataset (Dataset 1) and the UC San Diego dataset
(Dataset 2).
The research findings indicated that the AlexNet-based CNN model
achieved its highest accuracy of 95.79%, specifically for channel AFz in
dataset 1, surpassing all existing methods. The model exhibited a
sensitivity of 96.09%, specificity of 95.49%, positive predictive value
(PPV) of 95.35%, and an F1 score of 95.42% for the dataset 1. In dataset
2, the proposed model demonstrated the highest classification accuracy
Fig. 14. ROC and AUC obtained for HC vs PDSF classification on Dataset 2. of 99.84% for channel AF4 in both HC vs PDSO and HC vs PDSF. In
dataset 2, the model displayed a sensitivity of 100%, specificity of
obtained the AUC of 97.02%. Similarly, the AUC and ROC obtained on 99.67%, PPV of 100%, and an F1 score of 99.84% when classifying
dataset2 for HC vs PDSF and HC vs PDSO is shown in Figs. 14 and 15. PDSO vs HC. Across both datasets, the study revealed that channels
The figures reveal that our model has achieved the AUC of 100%, located in the central and frontal regions, such as AF4 and AFz, show­
denoting ideal performance on dataset 2. Thus, the ROC and AUC cased more representative features compared to other regions. Conse­
analysis shows that our model is highly consistent and accurate in binary quently, the results suggest that future research may not require the use
classification scenarios. of whole brain EEG channel data for identifying PD; focusing on central
and frontal regions should suffice for PD detection. Additionally, the
4. Discussions developed model’s performance was assessed through ROC, demon­
strating high consistency and accuracy in binary classification scenarios,
This study sought to introduce an innovative hybrid framework that with an AUC of 100% for UC San Diego and an AUC of 97.02% for the
combines a Time-Frequency Representation with a CNN based deep Iowa dataset.
learning model to automate the detection of PD from EEG signals. The To ensure consistency and reliability, a comparison was conducted

9
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

between the AlexNet-based CNN model and three other popular CNN highest accuracy compared to the VGG16, DarkNet19 and ResNet18
models, VGG16, DarkNet19 and ResNet18, within the proposed PD models for HC vs PD classification in dataset 1. The accuracy of the
detection framework. Like the AlexNet model, other three models: ResNet18 is slightly higher in dataset 2 compared to AlexNet, VGG16
VGG16, DarkNet19 and ResNet18 also utilized the best performing and DarkNet19 for both HC vs PDSO and HC vs PDSF cases. This trend
channels such as AFz for dataset 1 and AF4 for dataset 2. The experi­ remained consistent across sensitivity, specificity, precision, and F1-
mental procedures for VGG16, DarkNet19 and ResNet18 models were score, as illustrated in Fig. 16 (b)–(e), respectively. In conclusion, even
replicated five times, maintaining the same parameter settings (e.g. though the ResNet18 model shows slightly higher performance only for
batch size, epoch size) as employed in the proposed AlexNet model. dataset 2, but AlexNet model consistently exhibited superior perfor­
Fig. 16 (a)-(e) provides a comprehensive comparison of different per­ mance compared to the other three models in both datasets. Therefore,
formance parameters, including accuracy, sensitivity, specificity, pre­ we have decided that AlexNet is the best model for our proposed
cision, and F1 score, for the four CNN models: AlexNet, VGG16, framework.
DarkNet19 and ResNet18. Additionally, we conducted a comparison between our proposed
As depicted in Fig. 16(a), the AlexNet model demonstrated the method and existing approaches for the same PD EEG datasets that we

Fig. 16. Comparative analysis of performances among four CNN models within the proposed framework, (a) Accuracy comparison; (b) Sensitivity comparison; (c)
Specificity comparison; (d) Precision comparison; (e)F1-score comparison.

10
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

Table 8 utilization of fully connected layers, implementation of dropout regu­

Comparison of the proposed scheme with state-of-the-art method for the same larization, and extensive research support.
datasets.
Datasets Authors Methods Classes Overall 5. Conclusion and future plans
Accuracy

Dataset 1: Lee et al. Hjorth parameter and HC vs. PD 89.3% The key objective of this study was to determine the brain region that
Iowa [37] Gradient boosting provides vital information for efficient PD detection using EEG data and
dataset: decision tree + Gradient
to improve the performance of the proposed method compared to
boosting decision tree
Anjum Linear predictive HC vs. PD 85.7% existing approaches. To achieve this, a TFR-based AlexNet CNN frame­
et al. [14] coding + Hyperplanes work was developed for EEG-based PD diagnosis. The WST effectively
Sugden a channel-wise HC vs. PD 83.8% captured local and temporal properties of PD EEG data, while the
et al. [38] convolutional neural
AlexNet CNN model extracted complex features and detected spatial
network
Karaka & Gray-level co- HC vs. PD 85.71%
patterns at different scales, resulting in enhanced PD detection capa­
Latifo [2] occurrence matrix bilities. The proposed framework’s effectiveness was evaluated using
(GLCM) + SVM two real datasets: Iowa dataset (dataset 1) and San Diego dataset
Our SWT + Alexnet based HC vs. PD 95.79% (dataset 2). Through channel-wise analysis, the contribution of each
proposed CNN
EEG channel was assessed, leading to the identification of brain regions
Dataset 2: Qiu et al. Phase locking value and HC vs. 97.15%
UC San [1] Power spectral density PD_OFF 98.27% providing superior information for efficient PD diagnosis. Topographic
Diego + Proposed MCNN HC vs. maps, based on accuracy values of each channel, visually demonstrated
PD_ON the significance of AFz and AF4 channels, along with the frontal and
Khare Tunable Q wavelet HC vs. 96.13%
central brain regions, in facilitating efficient PD detection.
et al. [12] transform + Least PD_OFF 97.65%
square SVM HC vs.
The proposed model achieved impressive accuracy, sensitivity
PD_ON (SEN), specificity (SPE), positive predictive value (PPV), and F1 score
Aljalal CSP + LogEn + SVM/ Of-PD vs HC 99.41% for both datasets. On dataset 1, the model attained an accuracy of
et al. [39] KNN On-PD vs HC 95.76% 95.79%, SEN of 96.09%, SPE of 95.49%, PPV of 95.35%, and F1 score of
Loh et al. Gabor transform, and HC vs PD 99.44%
95.42% for PD vs. HC classification. On dataset 2, the model achieved an
[40] 2D CNN without drug 92.60%
HC vs PD accuracy of 99.84% for PDSF vs. HC classification, and also for PDSO vs.
with HC classification. The model also demonstrated excellent SEN, SPE,
medication PPV, and F1 score of 100%, 99.67%, 100%, and 99.84%, respectively,
Motin With the optimal Of-PD versus 87.09%
for PDSF vs. HC classification, and 99.67%, 100%, 99.67%, and 99.84%,
et al. [42] number + SVM HC
(polynomial)
respectively, for PDSO vs. HC classification. These results highlighted
Aljalal DWT + TShEn + KNN On-PD vs HC 94.21% the superiority of the TFR-based AlexNet CNN model over existing
et al. [41] Of-PD vs HC 99.89% methods in effectively classifying PD-related EEG data. Furthermore, we
Our SWT + AlexNet based HC vs PDSO 99.83% evaluated the performance of the AlexNet model in comparison to three
proposed CNN HC vs PDSF 99.83%
other CNN models, namely VGG16, DarkNet19 and ResNet18, within
the proposed framework. The experimental results indicate that the
used in this study. Table 8 presents the latest findings on EEG-based PD AlexNet-based CNN exhibited superior performance in both datasets
detection, specifically focusing on the overall accuracy of various when compared to the other three CNN models. Based on the optimistic
research works outcomes, the proposed model holds promising potential as a valuable
conducted on two datasets: the Iowa dataset (dataset 1) and the UC and enduring aid for experts and clinicians in diagnosing PD. Addi­
San Diego dataset (dataset 2). Within the context of the Iowa dataset, tionally, the combination model can be extended to other medical sig­
which consists of 14 PD patients and 14 control subjects, several studies nals, such as ECG, EOG, and EMG signals, suggesting potential
have been conducted on this dataset. The results depicted in Table 8 applications and practical significance in various medical domains.
highlight the significant advancements achieved by our proposed model
in terms of classification accuracy when compared to other relevant CRediT authorship contribution statement
studies, such as Lee et al. [37], Anjum et al. [14], Sugden et al. [38], and
Karaka & Latifo [2]. Our proposed model outperforms the other studies Siuly Siuly: Writing – review & editing, Writing – original draft,
by a large margin. Specifically, it achieves an accuracy that is 11.99% Methodology, Formal analysis, Data curation, Conceptualization. Smith
higher than Sugden et al. [38], 10.09% higher than Anjum et al. [14], K. Khare: Validation, Software, Formal analysis, Data curation. Enamul
10.08% higher than Karaka & Latifo [2], and 6.49% higher than Lee Kabir: Writing – review & editing, Validation, Investigation. Muham­
et al. [37]. mad Tariq Sadiq: Visualization, Validation, Software, Investigation.
The San Diego dataset, consisting of 15 PD patients and 16 controls, Hua Wang: Validation, Supervision, Project administration,
has been extensively studied by various researchers. The experiments Investigation.
conducted on this dataset include PD during ON medication (PDSO),
OFF medication (PDSF), and healthy controls (HC). In Table 8, it is
evident that our proposed method achieves the highest classification Declaration of competing interest
accuracy for PD during ON medication (PDSO). Additionally, for PD
during OFF medication (PDSF) classification, our proposed model The authors declare that they have no known competing financial
demonstrates nearly identical accuracy (99.83%) compared to the interests or personal relationships that could have appeared to influence
highest accuracy obtained by Aljalal et al. [41] (99.89%). These the work reported in this paper.
results collectively demonstrate that the TFR based AlexNet CNN
model we propose exhibits superior classification performance References
compared to parallel models when applied to the datasets utilized in this
study. The superiority of our proposed model in EEG-based PD detection [1] L. Qiu, J. Li, J. Pan, Parkinson’s disease detection based on multi-pattern analysis
can be attributed to their deeper architecture, larger receptive fields, and multi-scale convolutional neural networks, Front. Neurosci. 16 (2022) 957181,
https://doi.org/10.3389/fnins.2022.957181.

11
S. Siuly et al. Computers in Biology and Medicine 174 (2024) 108462

[2] M.F. Karakaş, F. Latifoğlu, Distinguishing Parkinson’s disease with GLCM features [23] X. Shi, T. Wang, L. Wang, P. Liu, N. Yan, Hybrid convolutional recurrent neural
from the Hankelization of EEG signals, Diagnostics 13 (2023) 1769, https://doi. networks outperform CNN and RNN in task-state EEG detection for Parkinson’s
org/10.3390/diagnostics13101769. disease, in: Asia-Pacific Signal and Information Processing Association Annual
[3] Smith K. Khare, Varun Bajaj, U. Rajendra Acharya, PDCNNet: an automatic Summit and Conference, 2019, https://doi.org/10.1109/
framework for the detection of Parkinson’s disease using EEG signals, IEEE Sensor. apsipaasc47483.2019.9023190.
J. 21 (15) (AUGUST 1, 2021). [24] S. Lee, R. Hussein, M.J. McKeown, A Deep convolutional-recurrent neural network
[4] World Health Organization. Parkinson Disease. Available online: Parkinson architecture for Parkinson’s disease EEG classification, in: IEEE Global Conference
disease, World Health Organization (WHO), https://www.who.int/news-room/fact on Signal and Information Processing, 2019, https://doi.org/10.1109/
-sheets/detail/parkinson-disease#:~:text=Global%20estimates%20in%202019% globalsip45357.2019.8969309.
20showed,of%20over%20100%25%20since%202000.. [25] J. Hong, Y. Luo, M. Mou, J. Fu, Y. Zhang, W. Xue, T. Xie, L. Tao, Y. Lou, F. Zhu,
[5] S. Selvaraj, S. Piramanayagam, Impact of gene mutation in the development of Convolutional neural network-based annotation of bacterial type IV secretion
Parkinson’s disease, Genes Dis 6 (2) (2019 Feb 27) 120–128, https://doi.org/ system effectors with enhanced accuracy and reduced false discovery, Briefings
10.1016/j.gendis.2019.01.004. PMID: 31193965; PMCID: PMC6545447. Bioinf. 21 (5) (2020) 1825–1836, https://doi.org/10.1093/bib/bbz120. PMID:
[6] A. Lee, R.M. Gilbert, Epidemiology of Parkinson disease, Neurol. Clin. 34 (4) (2016 31860715.
Nov) 955–965, https://doi.org/10.1016/j.ncl.2016.06.012. Epub 2016 Aug 18. [26] Zheng, et al., AnnoPRO: a strategy for protein function annotation based on multi-
PMID: 27720003. scale protein representation and a hybrid deep learning of dual-path encoding,
[7] A. Reeve, E. Simcox, D. Turnbull, Ageing and Parkinson’s disease: why is Genome Biol. 25 (1) (2024) 41, 2024). PMID: 38303023.
advancing age the biggest risk factor? Ageing Res. Rev. 14 (100) (2014 Mar) [27] Mou, et al., A transformer-based ensemble framework for the prediction of protein-
19–30, https://doi.org/10.1016/j.arr.2014.01.004. Epub 2014 Feb 3. PMID: protein interaction sites, Research 6 (2023) 240. PMID: 37771850.
24503004; PMCID: PMC3989046. [28] Wang, et al., A task-specific encoding algorithm for RNAs and RNA-associated
[8] M.F. Beal, Mitochondria, oxidative damage, and inflammation in Parkinson’s interactions based on convolutional autoencoder, Nucleic Acids Res. 51 (21)
disease, Ann. N. Y. Acad. Sci. 991 (2003) 120–131. (2023) e110, 2023). PMID: 37889083.
[9] Parkinson’s Foundation. Parkinson’s Statistics. Accessed on April 18, 2023, from [29] J. Hong, Y. Luo, Y. Zhang, J. Ying, W. Xue, T. Xie, L. Tao, F. Zhu, Protein functional
https://www.parkinson.org/Understanding-Parkinsons/Statistics. annotation of simultaneously improved stability, accuracy and false discovery rate
[10] K. Li, B. Ao, X. Wu, Q. Wen, E. Ul Haq, J. Yin, Parkinson’s disease detection and achieved by a sequence-based deep learning, Briefings Bioinf. 21 (4) (2020 Jul 15)
classification using EEG based on deep CNN-LSTM model, Biotechnol. Genet. Eng. 1437–1447, https://doi.org/10.1093/bib/bbz081. PMID: 31504150.
Rev. 11 (2023 Apr) 1–20, https://doi.org/10.1080/02648725.2023.2200333. [30] Iowa dataset, The Iowa dataset analyzed for this study can be found in the
Epub ahead of print. PMID: 37039259. following link: https://bit.ly/3pP1pts..
[11] P.D. Barua, S. Dogan, T. Tuncer, M. Baygin, U.R. Acharya, Novel automated PD [31] A.P. Rockhill, N. Jackson, J. George, A. Aron, N.C. Swann, Uc San Diego Resting
detection system using aspirin pattern with EEG signals, Comput. Biol. Med. 137 state eeg data from patients with Parkinson’s disease. California, CA, Tech. Rep.
(2021 Oct) 104841, https://doi.org/10.1016/j.compbiomed.2021.104841. Epub (2020).
2021 Sep 6. PMID: 34509880. [32] H.W. Loh, C.P. Ooi, E. Palmer, P.D. Barua, S. Dogan, T. Tuncer, M. Baygin, U.
[12] Smith K. Khare, Varun Bajaj, U. Rajendra Acharya, Detection of Parkinson’s disease R. Acharya, GaborPDNet: gabor transformation and deep neural network for
using automated tunable Q wavelet transform technique with EEG signals, Parkinson’s disease detection using EEG signals, Electronics 10 (2021) 1740.
Biocybern. Biomed. Eng. 41 (2) (2021) 679–689, https://doi.org/10.1016/j. [33] Z. Liu, G. Yao, Q. Zhang, J. Zhang, X. Zeng, Wavelet scattering transform for ECG
bbe.2021.04.008. ISSN 0208-5216. beat classification, Comput. Math. Methods Med. 2020 (2020), https://doi.org/
[13] A.P.S. De Oliveira, M.A. De Santana, M.A. Andrade, J.A.S. Gomes, M.T. Rodrigues, 10.1155/2020/3215681. Article ID 3215681, 11 pages, 2020.
W.P.D. Santos, Early diagnosis of Parkinson’s disease using EEG, machine learning [34] J. Andén, S. Mallat, Multiscale scattering for audio classification, in: International
and partial directed coherence, Res. Biomed. Eng. 36 (3) (2020) 311–331, https:// Society for Music Information Retrieval Conference, 2011, pp. 657–662. Miami,
doi.org/10.1007/s42600-020-00072-w. Florida, USA.
[14] F. Anjum, S. Dasgupta, R. Mudumbai, A.D. Singh, J.F. Cavanagh, N.S. Narayanan, [35] Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton, ImageNet classification with
Linear predictive coding distinguishes spectral EEG features of Parkinson’s disease, deep convolutional neural networks, Commun. ACM 60 (6) (2017) 84–90, https://
Park. Relat. Disord. 79 (2020) 79–85, https://doi.org/10.1016/j. doi.org/10.1145/3065386.
parkreldis.2020.08.001. [36] Ahmet Esad Top, Hilal Kaya, Alexnet classification of EEG signals by using transfer
[15] R. Balestrino, A. Schapira, Parkinson disease, Eur. J. Neurol. 27 (1) (2019) 27–42, learning on convolutional neural networks via spectrogram, in: International
https://doi.org/10.1111/ene.14108. Conference on Engineering Technologies (ICENTE18), 2018, pp. 156–161.
[16] R. Yuvaraj, U.R. Acharya, Y. Hagiwara, A novel Parkinson’s disease diagnosis Index [37] S.B. Lee, Y.J. Kim, S. Hwang, H. Son, S.K. Lee, K.I. Park, et al., Predicting
using higher-order spectra features in EEG signals, Neural Comput. Appl. 30 (4) Parkinson’s disease using gradient boosting decision tree models with
(2018) 1225–1235, https://doi.org/10.1007/s00521-016-2756-z. electroencephalography signals, Parkinsonism Relat. Disorders 95 (2022) 77–85,
[17] E. Naghsh, M.F. Sabahi, S. Beheshti, Spatial analysis of EEG signals for Parkinson’s 10. 1016/j.parkreldis.2022.01.011.
disease stage detection, Signal, Image Video Proc. 14 (2) (2020) 397–405, https:// [38] R. Sugden, P. Diamandis, Generalizable electroencephalographic classification of
doi.org/10.1007/s11760-019-01564-8. Parkinson’s Disease using deep learning, medRxiv (2022).
[18] O. Faust, Y. Hagiwara, U. Raghavendra, R. Yuvaraj, N. Arunkumar, [39] M. Aljalal, S.A. Aldosari, K. AlSharabi, A.M. Abdurraqeeb, F.A. Alturki, Parkinson’s
M. Murugappan, U.R. Acharya, A deep learning approach for Parkinson’s disease disease detection from resting-state EEG signals using common spatial pattern,
diagnosis from EEG signals, Neural Comput. Appl. 32 (15) (2020) 10927–10933, entropy, and machine learning techniques, Diagnostics 12 (5) (2022) 1033.
https://doi.org/10.1007/s00521-018-3689-5. [40] H.W. Loh, C.P. Ooi, E. Palmer, P.D. Barua, S. Dogan, T. Tuncer, et al., GaborPDNet:
[19] M. Murugappan, W.B. Alshuaib, A.K. Bourisly, S.K. Khare, S. Sruthi, V. Bajaj, gabor transformation and deep neural network for Parkinson’s disease detection
Tunable Q wavelet transform based emotion classification in Parkinson’s disease using EEG signals, Electronics 10 (2021) 1740, https://doi.org/10.3390/
using Electroencephalography, PLoS One 15 (11) (2020) e0242014, https://doi. electronics10141740.
org/10.1371/journal.pone.0242014. [41] M. Aljalal, S.A. Aldosari, M. Molinas, et al., Detection of Parkinson’s disease from
[20] N. Wagh, Y. Varatharajah, EEG-GCNN: augmenting Electroencephalogram-based EEG signals using discrete wavelet transform, different entropy measures, and
neurological disease diagnosis using a domain-guided graph convolutional neural machine learning techniques, Sci. Rep. 12 (2022) 22547, https://doi.org/10.1038/
network, Signal Process. (2020). http://arxiv.org/pdf/2011.12107.pdf. s41598-022-26644-7.
[21] S. Xu, Z. Wang, J. Sun, Z. Zhang, Z. Wu, T. Yang, G. Xue, C. Cheng, Using a deep [42] M.A. Motin, M. Mahmud, D.J. Brown, Detecting Parkinson’s disease from
recurrent neural network with EEG signal to detect Parkinson’s disease, Ann. electroencephalogram signals: an explainable machine learning approach, in: 2022
Transl. Med. 8 (14) (2020) 874, https://doi.org/10.21037/atm-20-5100. IEEE 16th International Conference on Application of Information and
[22] M. Koch, V.J. Geraedts, H. Wang, M.R. Tannemaat, T.G. Back, Automated machine Communication Technologies (AICT), Washington DC, DC, USA, 2022, pp. 1–6,
learning for EEG-Based classification of Parkinson’s disease patients, Int. Conf. Big https://doi.org/10.1109/AICT55583.2022.10013589.
Data (2019).

12