Neural Computing and Applications (2024) 36:9297–9311

https://doi.org/10.1007/s00521-024-09569-2(0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

A novel approach for Parkinson’s disease detection using Vold-Kalman

order filtering and machine learning algorithms
Fatma Latifoğlu1 • Sultan Penekli1 • Fırat Orhanbulucu2 • Muhammad E. H. Chowdhury3

Received: 22 November 2023 / Accepted: 24 January 2024 / Published online: 27 February 2024
 The Author(s) 2024

Abstract
Parkinson’s disease (PD) is the second most common neurological disorder caused by damage to dopaminergic neurons.
Therefore, it is important to develop systems for early and automatic diagnosis of PD. For this purpose, a study that will
contribute to the development of systems for the automatic diagnosis of PD is presented. The Electroencephalography
(EEG) signals were decomposed into sub-bands using adaptive decomposition methods, such as empirical mode
decomposition, variational mode decomposition, and Vold-Kalman order filtering (VKF). Various features were extracted
from the sub-band decomposed signals, and the significant ones were determined by Chi-squared test. These important
features were applied as input to support vector machine (SVM), fitch neural network (FNN), k-nearest neighbours (KNN),
and decision trees (DT), machine learning (ML) models and classification was performed. We analysed the performance of
ML models by obtaining accuracy, sensitivity, specificity, positive predictive value, negative predictive values, F1-score,
false-positive rate, kappa statistics, and area under the curve. The classification process was performed for two cases: PD
ON-HC and PD OFF-HC groups. The most successful method in this study was the VKF method, which was applied for
the first time in this field with the approach specified for both cases. In both instances, the SVM algorithm was employed as
the ML model, with classifier performance criterion values close to 100%. The results obtained in this study seem to be
successful compared to the results of recent research on the diagnosis of PD.

Keywords Parkinson’s disease classification  Electroencephalography signals  Vold-Kalman order filtering 

Machine learning algorithms

1 Introduction

1.1 Background

Parkinson’s disease (PD) is the second most common

& Fatma Latifoğlu
[email protected] neurodegenerative disease globally, particularly affecting
people over the age of 50 [1, 2]. Patients with PD usually
Sultan Penekli
[email protected] have motor symptoms, such as mild tremors, slowing of
movements, gait or posture disturbance. In addition to
Fırat Orhanbulucu
[email protected] these symptoms, non-motor symptoms such as depression,
loss of smell, constipation, or sleep problems are also
Muhammad E. H. Chowdhury
[email protected] observed [1, 3, 4]. Clinical evaluation of PD according to
these symptoms can be difficult [5]. Different neuroimag-
1
Department of Biomedical Engineering, Erciyes University, ing measurement techniques have been used by clinicians
Kayseri, Türkiye in the diagnosis of PD. Some of these techniques are based
2
Department of Biomedical Engineering, Inonu University, on physiological signals, such as EEG, and electromyog-
Malatya, Türkiye raphy, while others are based on magnetic resonance
3
Department of Electrical Engineering, Qatar University, imaging types or computed tomography. Measurement
Doha, Qatar

123
9298 Neural Computing and Applications (2024) 36:9297–9311

techniques using physiological signals are more advanta- achieved remarkable success in diagnosing PD and evalu-
geous in terms of time and cost [6]. Among these signals, ating treatment efficacy through the analysis of multiple
EEG is the most frequently used measurement method in aspects from EEG data with ML algorithms. They obtained
neurodegenerative diseases, such as PD [7]. 99% success in non-drug usage and 95–98% success in
EEG is a measurement method that records the electrical drug use. Motin et al. [14] proposed an EEG-based auto-
activity in the cerebral cortex with the help of multiple mated PD monitoring technique, using the SVM classifier,
electrodes through the scalp and is widely used to diagnose which resulted in an accuracy of 87.10%, sensitivity of
whether there is any abnormality in the brain [7, 8]. This 93.33%, and specificity of 81.25%. Using an Artificial
method helps expert clinicians to interpret a neurological Neural Network model and selected features from EEG
disorder or analyse a medical condition [8]. However, signals, Kamalakannan et al. [15] achieved a commendable
traditional diagnostic methods sometimes encounter con- accuracy rate of 93.3% in classifying PD. Karakaş and
ditions that are invisible to clinicians and therefore difficult Latifoğlu introduced a new method for analysing EEG
to classify [9]. For this reason, expert clinicians need faster signals from PD patients. They used the grey-level co-oc-
and more reliable systems that can be used for early currence matrix method and various ML algorithms,
diagnosis of a neurological disorder such as PD [4]. The obtaining the best performance with the SVM classifier,
development of these systems will allow specialists to which resulted in an accuracy rate of 92.4% [16]. Fur-
rapidly detect PD, enabling early initiation of treatment. thermore, Biswas et al. [17] proposed an ‘‘Ensembled
Machine learning methods are mostly applied to EEG Expert System’’ ensemble model for early diagnosis of PD,
signals in the development of automatic diagnosis systems achieving an accuracy of 93.2% in predicting PD. These
[7, 9]. Because of the complex structure of non-stationary recent studies demonstrate the potential of EEG and ML
EEG signals, they can be difficult to analyse. To facilitate techniques in aiding the accurate diagnosis of Parkinson’s
these analyses, PD-healthy control (HC) groups can be disease, bringing valuable contributions to the field.
easily classified by applying features obtained by various In addition to the studies on the diagnosis of PD listed
nonlinear feature extraction methods [10]. In addition, above, other investigations using EEG signals, ML tech-
detailed information can be obtained at different frequen- niques, and deep learning techniques are reported in the
cies by decomposing EEG signals into sub-bands with literature [1, 4, 6, 18–20]. These studies and especially the
various methods. In this way, success can be achieved in ones conducted with the data set used in this study are
ML-based studies for the automatic diagnosis of neuro- mentioned in detail in the discussion section. Studies using
logical disorders. As a result of these studies, a more sys- DL models require larger datasets and can be difficult to
tematic decision-making process can be provided to implement because in most cases, codes are not publicly
clinicians [9]. In order to contribute to the automatic available to replicate the study [13]. Various studies have
detection of PD, EEG signals were decomposed into sub- been carried out to help diagnose PD, but it is seen that
bands using signal processing techniques in this study. there is a need for new studies in addition to these studies
Various features are extracted from these sub-bands and to support clinicians [9, 13]. For this reason, in this study, a
applied as input to ML models and an EEG–ML model- new study that can help early diagnosis of PD using EEG
based study is proposed. There are recent studies based on signals and ML algorithms is proposed. This study incor-
EEG signals and ML models for the diagnosis of PD. These porates an innovative approach that combines VKF and
studies are mentioned in the next sub-section. ML techniques for the diagnosis of PD. The fusion of VKF
and ML, in a manner not previously attempted, offers a
1.2 Related works novel and distinct perspective in the diagnosis of Parkin-
son’s disease. The developed model utilises VKF and ML
It is seen that there are successful studies that have recently methodologies to provide high diagnostic accuracy and
contributed to the literature based on EEG signals that have reliability, surpassing existing diagnostic methods for
performed the classification of PD with ML algorithms. Parkinson’s disease. In addition to early diagnosis of PD,
Recent studies on Parkinson’s disease diagnosis utilising drug efficacy is also investigated and analysed in the pro-
EEG signals and machine learning algorithms have shown posed study. Our motivation for conducting the proposed
promising results. study, the purpose of the study, and its contributions to the
Anjum et al. [11] developed the LEAPD model, literature are explained below.
achieving an 85.40% accuracy in distinguishing PD from
the HC group using various EEG signal data. Murugappan 1.3 Our purpose, motivation, and contributions
et al. [12] utilized tunable Q wavelet transform and neural
network methods, achieving an impressive 96.16% accu- EEG signals are nonlinear. For this reason, the analysis of
racy in PD detection. In another study, Aljalal et al. [13] EEG signals in the diagnosis of PD can be difficult and

123
Neural Computing and Applications (2024) 36:9297–9311 9299

time-consuming. Using EEG signals and ML models, this medication levodopa three times per day (PD ON) and
study aims to contribute to the development of systems for those who did not (PD OFF). On several days, PD ON and
the early and automatic diagnosis of PD. Detailed infor- PD OFF recordings were gathered [4, 11]. Reference
mation can be obtained from EEG signals by decomposing [21, 22] provides comprehensive information about the
them into sub-bands and obtaining features. For this reason, dataset and the preprocessing techniques used on the
in this study, EEG signals were decomposed into sub-bands dataset.
with three different methods, and various features were
obtained. These features were analysed with four different 2.2 Proposed study
ML models to classify PD ON (drug using)-HC and PD
OFF (non-medicated)-HC groups. This study uses EEG data and ML models to categorise PD
We were motivated by the technical quality of the data (ON–OFF)-HC groups. In the initial stage of the study, the
set used in the study (32 channels and a sampling fre- noises in the EEG signals received from 32 channels were
quency capable of providing high-resolution data) and its eliminated to achieve this. Using EMD, VMD, and VKF
novelty. Another motivator was the fact that the VKF procedures, the denoised EEG signals were divided into
approach used in this work has not been studied in studies eight subband components. Since all three approaches have
involving EEG medication efficacy data. a strong classification performance, eight sublevels were
We consider it a significant benefit to examine the drug decided upon for the breakdown procedure. In this manner,
effect in this study because the data set contains EEG more data was extracted from intricate EEG signals. These
signals in the form of PD ON and PD OFF. There are sub-bands were used to extract characteristics for Shannon
studies on the application of the VKF approach in several entropy, Hjort parameters, Power Sum, Kolmogorov
domains [20] when the studies are analysed. But using the complexity, and Katz fractal dimension. The significant
methodology used in this study, research has been done for features were identified by applying the Chi-squared test to
the first time on the identification of neurological illnesses lessen the processing burden during the classification pro-
using EEG data. The EMD and VMD methods, which are cess and to enhance the functionality of the classifier
common for breaking down signals into sub-bands, are models. Following the identification of the key features, the
compared with the performance of the VKF approach in classification process was carried out for the PD OFF-HC
this study rather than being evaluated separately. We think and PD ON-HC groups using four ML models (FNN,
that this work can add to the body of knowledge. In the SVM, KNN, and DT). As a result of the classification
discussion section, the study’s benefits and contributions process, accuracy, sensitivity, specificity, positive predic-
are described in more detail. tive value, negative predictive value, F1-score, false-posi-
tive rate, kappa statistics, and area under the curve rates
used to analyse the performance of ML models were
2 Material and method obtained. The software programme MATLAB 2022a
(MathWorks, Inc., Natick, MA, USA) was used to carry
2.1 Dataset description out every step of this work. Figure 2 provides a generic
flow chart outlining the steps in the proposed investigation.
The EEG dataset (version 1.0.4) shared as open source on Detailed information about the research steps explained in
the Open Neuro website by Rochill et al. [21] was used to
test the approaches suggested in this work. The University
of California, San Diego captured EEG signals. With the
help of 16 HCs (mean age: 63.5 9.6 years, gender: 9
females, 7 males) and 15 PDs (mean age: 63.2 8.2 years,
gender: 8 females, 7 men) who had mild-to-severe
Parkinson’s symptoms, EEG signals were recorded from
the participants. Each participant used their right hand.
Participants were instructed to sit comfortably and fix their
attention on the cross on the screen during the recording. In
this method, 32-channel BioSemi ActiveTwo systems were
used to record resting EEG at a 512-Hz sampling rate.
Figure 1 depicts the placements of the 32-channel EEG
electrodes utilised in the research. The PD group’s EEG
signals were divided into two groups: those who got ther-
apy with an equivalent amount of the dopaminergic Fig. 1 Layout of 32-channel EEG electrodes [13]

123
9300 Neural Computing and Applications (2024) 36:9297–9311

this section is explained in the following sections of the 2.3.2 Variational mode decomposition
study.
The VMD method presented by Dragomiretskiy et al. is
2.3 EEG signals preprocessing used as a time–frequency decomposition method [26].
With the VMD method, the signal complexity is reduced
EEG signal data is multidimensional and contains noise. As by capturing sudden power changes and the signals are
a result, a preprocessing stage is required before feature adaptively decomposed against noise. In the VMD method,
extraction to remove noise from the data. Getting features like EMD, signals are decomposed into IMFs. The VMD
and preprocessing them before the classification stage is a method overcomes the recursive approach that does not
critical step in the creation of ML models. Blinking allow backward error correction and the inability to ade-
movements, electrical noise, and other sounds were elim- quately handle noise, both of which are viewed as short-
inated from the EEG data used in this work before signal comings in the EMD method [24]. The VMD method uses
stage processing and feature extraction by using a a simultaneous approach to extract the IMF from the sig-
0.1–40 Hz finite impulse response (FIR) band-pass filter. nal. VMD is less sensitive to noise than EMD and leaves no
Independent component analysis was also utilised to residual noise [27]. As in the EMD method, EEG signals
remove interference artefacts from the stream. The are decomposed into eight subband components in this
denoised EEG data were divided into sub-bands using study. A sample of eight IMFs obtained using VMD is
EMD, VMD, and VKF signal processing techniques. shown in Fig. 4.
Because the EEG signal encompasses a variety of oscilla-
tions, these sub-bands are referred to as intrinsic mode 2.3.3 Vold-Kalman order filtering
functions (IMF).
The VKF approach proposed by Hvard Vold and Jan
2.3.1 Empirical mode decomposition Leuridan can use the frequency vector to separate the
components of a non-stationary signal [28]. The VKF
To break down a signal into IMFs, Huang et al. [23] approach can effectively extract the desired sequence
developed the EMD approach. EEG signals are hypothe- component and has been employed in studies in a variety of
sised to exhibit many oscillations at the same time. The domains [29]. The VKF approach is capable of effectively
EMD approach is used in this situation to split the signals decomposing the frequency components of complicated
into stationary sub-bands. The difference between zero and multicomponent signals like EEG [29, 30]. This fea-
crossing and extremum numbers in the EMD approach ture allows you to extract single-component signals from
should be smaller than one. Furthermore, the average value multicomponent signals. Signal components in the time
of the upper and lower envelopes created by local maxima domain can be retrieved directly using the VKF approach.
and minima should be zero [24]. The EMD approach, In contrast to EMD and VMD signal decomposition and
which has been used for various physiological signals in filtering approaches, the VKF method collects signal
research, both decomposes and removes noise [23, 25]. The components directly from the time domain. In this manner,
EMD approach was used to split EEG signals into eight the phase variation produced by time–frequency conver-
sub-bands in this investigation. Figure 3 depicts a sample sion is avoided. In comparison to the EMD approach, the
of eight IMFs derived using EMD. VKF method may change the centre frequency based on
the instantaneous frequency and avoid the overlaps created
by the EMD method [31]. The frequency vector was
obtained in four Hz intervals from 4 to 32 Hz while

Fig. 2 Flowchart of the proposed work

123
Neural Computing and Applications (2024) 36:9297–9311 9301

Fig. 3 Eight IMF examples

were obtained as a result of the
application of the EMD method

dissecting the EEG data into subband components using complexity, and Katz Fractal Dimension). These features
the VKF method. Various frequency value ranges were were recovered from each subband of signals divided into
attempted in the process of determining these frequency eight sub-bands using the EMD, VMD, and VKF methods.
ranges. The best categorization performance was discov- A total of seven features reflecting time–frequency domain
ered in the frequency band of ‘‘4,8,…,28,32’’ Hz. As a and nonlinear system properties were recovered during
result, EEG signals were divided into eight subband com- feature extraction (Table 1). The following are brief
ponents in this investigation utilising Second Generation descriptions of the study’s features.
VKF (Fig. 5). Shannon entropy is used to quantify how accurately
signal information can be measured. It is a quantity that
2.4 Feature extraction and selection quantifies how much and in what proportion a signal
generates information. It is an uncertainty metric that is
Before developing ML models, the feature extraction and frequently used to evaluate the degree of chaos in an EEG
selection process is used to efficiently describe a dataset. signal [32, 33]. Hjorth’s parameters (HP) are activity,
These strategies can minimise the number of accessible mobility, and complexity, which he brought to the litera-
features in a dataset, pick the most significant features, and ture. Activity is defined as being aware of the signal
improve the performance of ML models. Time–frequency strength. The average frequency is estimated using
domain and nonlinear studies were done on decomposed mobility. The complexity of the power spectrum is defined
EEG signals to extract several properties (Shannon as its bandwidth, frequency variation, or standard deviation
Entropy, Hjort parameters, Power Sum, Kolmogorov [34]. HPs are well-suited for non-stationary signals like

123
9302 Neural Computing and Applications (2024) 36:9297–9311

Fig. 4 Example of eight sub-

bands obtained as a result of
applying the VMD method

EEG and are frequently employed in research [33, 34]. The method for determining the relationship between actual
power of the EEG signal was estimated by summing the data and the model. The independence of two variables is
power of each frequency after it was translated from the tested using this method. Furthermore, the Chi-square test
time domain to the frequency domain using the Fourier explores if the distribution of one variable differs from that
transform. Kolmogorov complexity is a technique for of another [33].
determining an object’s complexity, which is measured as
the size of the shortest representation of information about 2.5 Classification
an object [32, 35]. Another technique for calculating
complexity is fractal dimension (FD), which uses the non- Classifier models use the classification process to predict
stationary features of EEG signals [36]. Higuchi and Katz the accurate labelling of the data supplied as input. In this
FD are the most often utilised variants of FD for detecting study, a classification approach was used to separate PD
rapid changes in EEG patterns and diagnosing various patients (PD ON and PD OFF) from HC groups using
disorders. In this study, the Katz FD type was favoured. variables derived from EEG data analysis. Classical ML
The goal of this study was to estimate the upper limit of methods such as FNN, SVM, KNN, and DT were
the performance of the preferred features. One of the employed in the classification process. The ten fold cross-
automatic feature selection methods, the Chi-square-based validation (10-CV) and leave-one-subject-out-cross-vali-
feature selection approach, was used to limit the number of dation (L-CV) approaches were utilised during the classi-
available features before the classification procedure. The fication procedure. The following is a summary of the ML
Chi-square (X2) statistic is a traditional statistical test algorithms utilised in this study:

123
Neural Computing and Applications (2024) 36:9297–9311 9303

Fig. 5 Example of eight sub-

bands obtained as a result of
applying the VKF method

Table 1 List of EEG features

Feature type Extracted features
used in this study
Time–frequency domain features Hjort parameters (HPs) (activity; mobility; complexity)
Power sum (PS)
Non-linear features Shannon entropy (SE)
Kolmogorov complexity (KC)
Katz fractal dimension (KFD)

A feed-forward artificial neural network (ANN) model, Support Vector Machine (SVM) is a machine learning
commonly known as a multilayer neural network, is what technique that decreases risk by utilising various solution
Fitch Neural Network (FNN) is. It usually consists of three approaches for linear or nonlinear problems. It is used to
layers: the input layer, the output layer, and the hidden solve binary or multiple classification problems and is
layer. Each layer of the neural network is linked to the highly fast [16, 34].
layer before it [16, 37]. The model utilised in the study was The k-nearest neighbours (KNN) classifier is a super-
constructed with the MATLAB software programme’s vised machine learning technique that leverages the nearest
‘‘fitcnet’’ function. An input layer, three hidden layers, and distance between data points to perform classification [16].
an output layer comprise the model. In general, the nearest neighbour is calculated by using

123
9304 Neural Computing and Applications (2024) 36:9297–9311

distance calculations such as Euclid and Chebychev [34]. The PPV rate measures the proportion of true positive
Finding the minimal values of these distances among all instances among all instances predicted as positive by the
data sets yields classes [38]. classifiers. It is also used as precision ratio in studies.
The Decision Tree (DT) model compares numerical tp
properties that consistently combine a set of tests. DT PPV ¼ ð4Þ
tp þ fp
classifiers are recognised as one of the most powerful
algorithms for data classification problems in a variety of The NPV rate measures the proportion of true negative
domains [39]. instances among all instances predicted as negative by the
classifiers.
2.6 Performance metrics tn
NPV ¼ ð5Þ
tn þ fn
Performance evaluation based on classifier prediction is an
The F1-score combines the precision and recall values
important topic for classification procedures. Some factors
of the classifier model using their harmonic mean.
govern the performance of any ML algorithm. True posi-
tive (tp), true negative (tn), false positive (fp), and false ðPrecision  RecallÞ
F1  score ¼ 2  ð6Þ
negative (fn) values are employed in the derivation of these ðPrecision þ RecallÞ
criteria [40]. These values were used to calculate the The FPR is the proportion of negative cases that are
accuracy (ACC), sensitivity (SENS), specificity (SPE), incorrectly identified as positive cases in the classification
positive predictive value (PPV), negative predictive value process.
(NPV), F1-score (F1), false-positive rate (FPR), and area
under the curve (AUC) rates in this investigation. In order fp
FPR ¼ ¼ 1  specificity ð7Þ
to measure the reliability between classifiers, kappa tn þ fp
statistic (KPS) values were obtained using a probabilistic The kappa statistic is a measure of the reliability of the
approach. The performances of ML models were compared classification process, which helps to evaluate the perfor-
based on the results. mance between classifiers in machine learning studies. As
TP: The classifier model correctly predicted the data the kappa statistical value, which is measured based on
labelled as PD. probability, approaches ? 1, the model performance is
TN: The classifier model correctly predicted the data considered to be good.
labelled as HC. The AUC rate is a performance measure of how well
FP: The classifier model incorrectly predicted the HC- two classes can be distinguished according to the area
labelled data and included it in the PD class. under the receiver operating characteristics (ROC) curve.
FN: The classifier model incorrectly predicted data AUC value approaching 1 indicates that the performance of
labelled PD and included it in the HC class. the classifier is perfect.
The accuracy rate is defined as the percentage of correct
predictions in a dataset.

Accuracy ¼
tp þ tn
ð1Þ
3 Experimental results
tp þ tn þ fp þ fn
The goal of this study was to classify PD (ON–OFF) and
The sensitivity rate refers to the ability of the classifier
HC groups using EEG signals. Three distinct signal pro-
models to correctly identify positive examples (PD in this
cessing approaches were used to extract various features
study). False negatives are important in disease detection.
from EEG signals for this purpose. These features were fed
The sensitivity rate measures false negatives against true
into four different ML algorithms to find the best suc-
positives. It is also used as the Recall rate in studies.
cessful signal processing method and ML algorithm.
tp Preprocessing was used to denoise EEG signals in the
Sensitivity ¼ ð2Þ
tp þ fn first step of this study. Three distinct signal processing
The specificity is the percentage of negative samples algorithms (EMD, VMD, and VKF) were utilised to
(HC groups in this study) detected by the classifier models. deconstruct the denoised EEG signals into sub-bands to
retrieve characteristics from each component. The perfor-
tn
Specificity ¼ ð3Þ mances of the three approaches were so compared.
tn þ fp
The second stage of the research identified seven fea-
tures from each band of EEG signals, which were then
divided into eight sub-bands using EMD, VMD, and VKF

123
Neural Computing and Applications (2024) 36:9297–9311 9305

algorithms. The investigation yielded a total of 1792 fea- study’s findings, the SVM algorithm provided the most
tures (32 channels, 8 sub-bands, and 7 features). The Chi- accurate classification results for the PD ON-HC group (10/
Squared test was used before the classification procedure L-CV). The FNN, KNN, and DT algorithms (10/L-CV)
since the number of extracted features is high in the performed the best in the categorization of the PD OFF-HC
number of data and may cause a processing load. Important group.
features that will be used as input to ML algorithms and
that are effective in classifying outcomes are determined in
this manner. Classification results for 5–10–15–20 char- 4 Discussions
acteristics were obtained using the Chi-squared test.
Experiments with varying numbers of features revealed Some issues can be detected in the early stages of PD until
that 5 features provided approximate classification perfor- the disease reaches the medium level [22]. As a result, it is
mance when compared to other feature numbers. However, critical to develop automatic diagnostic methods that can
in this study, the classification results produced by ana- provide early detection of PD to avoid disease development
lysing 5 features are provided to maintain a balance and assist experts. Several studies on the automatic diag-
between the number of data utilised and the number of nosis of PD have recently been completed, although it has
features used to interpret the classification results more been suggested that additional investigations are required
accurately and to reduce the processing burden. to contribute to the development of systems [9, 13]. As a
The final stage of the investigation involved classifica- result, this study presents research that can contribute to the
tion based on CV-10 and L-CV instances. In the CV-10 development of systems for the automatic diagnosis of PD
process, k = 10 was selected and the data consisting of utilising EEG data from the PD (OFF–ON) and HC groups.
EEG recordings were randomly divided into parts. While EEG signals were divided into sub-bands in this work using
10% of this data was allocated for testing, 90% was allo- the EMD, VMD, and VKF signal processing methods, and
cated for training and this process was performed 10 times. characteristics were retrieved from these sub-bands. The
In the L-CV process, the data obtained from one record was Chi-squared test was used to find critical parameters
used as the test set and all other records in the data set were influencing classification performance based on these fea-
used as the training set. As a result, the performance cri- tures. The determined features were analysed using four
teria obtained at each stage were averaged and the final different ML models, and the PD ON-HC and PD OFF-HC
results were generated. In the classification procedure, four groups were classified. In this way, the performance of
different ML algorithms (FNN, SVM, KNN, and DT) that signal processing methods as well as the performance of
are commonly utilised in biological investigations were ML models were assessed alongside the effects of drugs on
chosen. The classification results were evaluated using the PD. Multifaceted research has so been carried out in this
performance metrics ACC, SENS, SPE, PPV, NPV, FPR, way.
F1, KPS, and AUC. The EMD, VMD, and VKF approaches ML and DL approaches are commonly used with EEG
were used to classify PD ON-HC and PD OFF-HC groups signals in studies on the creation of automatic detection
in three separate cases. In this approach, the performance systems for PD and other neurological diseases
of signal processing methods was analysed using four [6, 7, 9, 10, 18, 41–43]. However, investigations employing
distinct ML algorithms, and the efficacy of dopaminergic DL models necessitate big data collection. Furthermore,
drugs in Parkinson’s patients was evaluated. investigations using DL models may necessitate a high
Tables 2 and 3 show the outcomes of the classification computational load and may be time-consuming to
processes done for the PD ON-HC and PD OFF-HC groups implement [13]. As a result, investigations using ML
using the EMD approach. Tables 4 and 5 show the out- models can assist specialists in making methodical deci-
comes of the classification processes done for the PD ON- sions. Because EEG signals are non-stationary, they have a
HC and PD OFF-HC groups using the VMD approach. complicated structure. As a result, by splitting EEG signals
Tables 6 and 7 present the results of the classification into numerous sub-bands before undertaking EEG–ML
processes done for the PD ON-HC and PD OFF-HC groups model-based automatic disease detection studies, vital
using the VKF approach. In the tables, performance eval- information can be collected, and the analysis of the EEG
uation results are expressed as percentage (%). signal can be facilitated with various features [10, 42]
When the results shown in Tables 2, 3, 4, 5, 6 and 7 are In this work, EEG data were divided into sub-bands
evaluated, it is clear that the VKF approach outperforms using EMD, VMD, and VKF adaptive signal processing
the VMD and EMD methods in all ML algorithms used in algorithms. Then, from these sub-bands, seven significant
the study. In almost all ML algorithms, the VKF approach features [34, 35, 42] that are commonly used in EEG
achieves a success rate of 100% or close to 100% in terms investigations were retrieved. Because EEG signals are
of classification performance criterion values. Based on the difficult and time-consuming to analyse, it is critical to

123
9306 Neural Computing and Applications (2024) 36:9297–9311

Table 2 EMD method PD ON-

Performance metrics/classifier FNN SVM KNN DT
HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 73.23/74.84 75.16/74.84 74.83/75.16 72.26/71.30

Sensitivity (%) 65.33/70 56.67/56.67 66/72 78/79.33
Specificity (%) 80.63/79.38 92.50/91.88 83.12/78.13 66.88/63.76
Positive predictive value (%) 76.51/76.13 87.79/86.82 79.10/76 69.22/67.72
Negative predictive value (%) 71.50/73.87 69.56/69.46 72.41/75.16 77.52/78.10
Area under the curve (%) 73.83/75.46 71.04/69.70 75.44/76.46 70.71/72.42
F1-score (%) 70.48/72.94 68.88/68.58 71.96/73.95 73.53/73.10
False-positive rate (%) 19.37/20.62 7.50/8.12 16.88/21.87 33.12/36.24
Kappa statistic (%) 68/68 68/61 61/74 36/33
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

Table 3 EMD method PD OFF-

Performance metrics/classifier FNN SVM KNN DT
HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 60.32/63.22 69.68/70.65 69.68/77.10 69.35/71.94

Sensitivity (%) 48.67/61.33 58/56 53.33/66.67 65.33/66.67
Specificity (%) 71.25/65 80.63/84.38 85/86.88 73.13/76.88
Positive predictive value (%) 62.90/63.32 74.78/77.63 81.18/86.10 69.57/73.10
Negative predictive value (%) 59.62/63.88 67.36/67.25 66.10/73.58 69.19/71.10
Area under the curve (%) 61.85/65.38 71.38/71.42 75.75/77.69 56.27/58.94
F1-score (%) 54.88/62.31 65.33/65.10 64.38/75.15 67.38/69.74
False-positive rate (%) 28.75/35 19.37/15.62 15/13.12 26.87/23.12
Kappa statistic (%) 68/68 68/61 61/74 30/35
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

Table 4 VMD method PD ON-

Performance metrics/classifier FNN SVM KNN DT
HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 70.33/67.10 73.55/76.77 79.03/84.52 77.42/76.65

Sensitivity (%) 63.53/58 54.67/59.33 64/80.67 80.67/80
Specificity (%) 76.88/75.63 91.25/93.13 93.13/88.13 74.38/73.13
Positive predictive value (%) 72.67/69.10 85.56/89 89.70/86.34 74.81/73.60
Negative predictive value (%) 69.51/65.91 68.23/70.95 73.57/83.21 80.86/79.65
Area under the curve (%) 71.21/65.90 73.33/74.63 77.08/87.52 79.73/69.88
F1-score (%) 67.68/63.10 66.71/71.20 74.70/83.41 77.63/76.67
False-positive rate (%) 23.12/24.37 8.75/6.87 6.87/11.87 25.62/26.87
Kappa statistic (%) 74/68 55/48 81/94 74/68
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

deconstruct them into sub-bands and extract features in study are analysed, it is clear that VMD results are suc-
order to obtain precise information. In PD studies, EMD cessful in supporting the investigations.
and VMD approaches are recommended and compared To get comprehensive information from EEG data, these
[43, 44]. In research comparing EMD with VMD signals must be decomposed into several sub-bands [47].
methodologies, VMD is consistently found to be superior Adaptive decomposition methods like VKF, VMD, and
[26, 45, 46]. When the results in Tables 2, 3, 4 and 5 of this EMD are extremely powerful and effective in processing

123
Neural Computing and Applications (2024) 36:9297–9311 9307

Table 5 VMD method PD

Performance metrics/classifier FNN SVM KNN DT
OFF-HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 67.74/76.21 71.29/66.77 82.90/86.45 76.45/83.55

Sensitivity (%) 62.67/68.67 65.33/64.87 80.67/85.33 70.67/74
Specificity (%) 72.50/74.38 76.88/68.75 85/87.50 81.88/92.50
Positive predictive value (%) 67.95/72.59 72.80/65.95 83.60/86.46 78.94/91.32
Negative predictive value (%) 68.10/71.42 70.39/67.54 82.62/86.52 74.98/79.43
Area under the curve (%) 73.41/76.38 78.50/77.92 84.33/87.20 81.17/82.88
F1-score (%) 65.20/70.58 68.86/65.30 82.11/85.90 74.58/81.75
False-positive rate (%) 27.50/25.62 23.12/31.25 15/12.50 18.12/7.50
Kappa statistic (%) 55/55 55/55 55/55 61/68
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

Table 6 VKF method PD ON-

Performance metrics/classifier FNN SVM KNN DT
HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 99.35/100 100/100 99.68/100 93.23/93.55

Sensitivity (%) 98.67/100 100/100 99.33/100 92/93.33
Specificity (%) 100/100 100/100 100/100 94.38/93.75
Positive predictive value (%) 100/100 100/100 100/100 93.95/93.33
Negative predictive value (%) 98.82/100 100/100 99.41/100 92.71/93.75
Area under the curve (%) 100/100 100/100 100/100 93.19/93.54
F1-score (%) 99.33/100 100/100 99.66/100 92.96/93.33
False-positive rate (%) 0/0 0/0 0/0 5.62/6.25
Kappa statistic (%) 100/100 100/100 100/100 94/87
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

Table 7 VKF method PD OFF-

Performance metrics/classifier FNN SVM KNN DT
HC classification results
10/L-CV 10/L-CV 10/L-CV 10/L-CV

Accuracy (%) 100/100 99.68/100 100/100 100/100

Sensitivity (%) 100/100 99.33/100 100/100 100/100
Specificity (%) 100/100 100/100 100/100 100/100
Positive predictive value (%) 100/100 100/100 100/100 100/100
Negative predictive value (%) 100/100 99.41/100 100/100 100/100
Area under the curve (%) 100/100 100/100 100/100 100/100
F1-score (%) 100/100 99.66/100 100/100 100/100
False-positive rate (%) 0/0 0/0 0/0 0/0
Kappa statistic (%) 100/100 100/100 100/100 100/100
10/L-CV, tenfold and leave-one-out cross-validation process; FNN, fitch neural network; SVM, support
vector machine; KNN, k-nearest neighbours; DT, decision tree

signals like EEG [30]. By recording immediate information When the studies are examined, it is clear that the VKF
from complicated signals, such as EEG, the VKF approach approach has been used in a variety of sectors [30, 31, 48].
can successfully separate the frequency components of the There is a study that compares the advantages and disad-
signal. Because of this aspect, we believe that the VKF vantages of the VKF approach to the EMD and VMD
approach will be more successful than the EMD and VMD methods in detail [30]. When compared to EMD, the VKF
methods in investigations including EEG inputs. approach is superior and yields successful results [31].

123
9308 Neural Computing and Applications (2024) 36:9297–9311

Table 8 Comparison table between the results of the presented study and the results of the existing studies
Study Dataset Methods Best classifier Best evaluation results (%)

Aljalal et al. [13] Open Common spatial pattern ? log energy entropy SVM PD ON-HC
Neuro (ACC):95.76
KNN PD OFF-HC
(ACC/F1):99.41/99.40
Motin et al. [14] Open Feature extraction SVM PD ON-HC
Neuro (ACC/F1):87.10/86.87
Kamalakannan Open Feature extraction ANN PD ON-HC
et al. [15] Neuro (ACC/FPR/F1):93.30/12.26/93.48
Barua et al. [42] Open Aspirin pattern and maximum absolute pooling KNN PD ON-HC
Neuro decomposition (ACC/F1):95.48/95.48
PD OFF-HC
(ACC/F1):93.57/93.57
Khare et al. [47] Open Tunable Q wavelet transform- feature Least square PD ON-HC
Neuro extraction SVM (ACC/AUC):97.65/98.56
PD OFF-HC
(ACC/AUC):96.13/97
Aljajal et al. [50] Open Discrete wavelet transform-feature extraction KNN PD ON-HC
Neuro (ACC/F1):94.21/94.40
PD OFF-HC
(ACC/F1):99.89/99.89
Present study Open VKF-feature extraction SVM PD ON-HC (ACC/AUC/FPR/F1/KPS):100/
Neuro FNN/KNN/ 100/0/100/100
DT PD OFF-HC (ACC/AUC/FPR/F1/
KPS):100/100/0/100/100

There has been no research comparing the VKF method to allowing detailed analysis of neural activity [49]. DL and
the VKF method. When the findings in Tables 2, 3, 4, 5, 6 ML models have been used in certain research with this
and 7 are examined, it is clear that the VKF approach dataset. There has been research that has yielded positive
outperforms the EMD and VMD methods in all ML results for the automatic detection of PD utilising DL
models. To the best of our knowledge, there is just one models [1, 4, 6, 20, 43]. Table 8 displays the findings from
study in which the popular adaptive parsing algorithms the studies that used ML models to evaluate the dataset in
VKF, VMD, and EMD are evaluated together and exam- this study. Because ML models were employed in this
ples of their applicability in various domains are provided study, papers using DL models were not included in the
[30]. The VKF method has been used for the first time in comparison table.
the research of EEG signals and neurological diseases, such Table 8 lists recent studies [13–15, 47, 50] that have
as the strategy used in this work. We believe that the study contributed to the development of automatic PD diagnosis
is unique and the first of its kind. We believe that com- methods. These research are all based on EEG readings and
paring the performance of the applied methodologies, such machine learning algorithms. This study analyses and
as the one used in this study, can raise acceptance of the compares the performance of ML models (SVM, KNN)
validity of the findings. that have been effectively utilised in the investigations
The Open Neuro [21] dataset, which is considered listed in Table 8. In addition, the performance of FNN and
legitimate and technically important in recent studies [49], DT models was examined in this work. As a result of the
was used in this study. In addition to being new, it is analysis, the features collected by the VKF approach, when
critical that this dataset is available in a reliable interna- combined with the FNN, SVM, and KNN models, out-
tional database, contains pharmacological activity, and has performed or were comparable to the studies in Table 8.
not been used sufficiently before this study. The strong When the results of this study are evaluated in terms of the
features of the dataset are that it has a sampling frequency classifier measurement metrics given in Table 8, it is seen
of 512 Hz and provides information from 32 channels, that ACC, AUC, F1, and KPS rates are 100%. It is seen that

123
Neural Computing and Applications (2024) 36:9297–9311 9309

the FPR rate is 0%. In classification studies, FPR being 0 or 5. The features extracted from the VKF-EMD-VMD
close to 0 is a desirable rate for classifier performance. adaptive subband decomposition methods were
ACC, AUC, F1, and KPS rates close to 100 and 100 are assessed with ML models in this work, and the three
considered to be desirable for classifier performance approaches’ performances were compared. In recent
[51–53]. From this point of view, it is seen that the results research for PD diagnosis, EMD and VMD approaches
obtained in the study are within the desired rates. In have been chosen [43, 44]. In addition to these
addition, it is seen that the results of the research are methods, the VKF method was employed for the first
superior to the results of similar metric rates obtained in the time in a study for the diagnosis of PD. For the first
studies given in Table 8. time, the performance of EMD-VMD-VKF approaches
To the best of our knowledge, this work is the first time was compared using the approach used in this work.
the VKF approach has been used for EEG readings on 6. This study also looked at two separate classification
neurological diseases. In this work, almost all of the ML cases, PD ON-HC and PD OFF-HC. In this manner, the
algorithms that used the VKF technique characteristics as influence of a dopaminergic medication, which is
input in the classification of PD ON-HC and PD OFF-HC significant for PD, was examined in terms of classifier
groups performed nearly perfectly. It can be shown that the performance measures.
results achieved in this study are more successful than
Although the results obtained in this study are promising
those obtained in previous investigations [13, 14, 47, 50].
for the diagnosis of PD, the study has some limitations.
EEG data were split into sub-bands and various features
These limitations can be summarised as follows:
were collected and assessed with ML models to classify PD
(ON–OFF)-HC groups in the experiments presented in Ref 1. The dataset used in this work is technically advanced
[47, 50]. It was discovered in this investigation that the and novel [49]. However, the fact that there are so few
results achieved by using the VKF method were superior to open source datasets that are recognised as authentic,
the findings obtained in the studies cited in Ref [47, 50]. In such as this one, limits the scope of our research. For
this regard, we believe that our work will contribute to the the findings of this study to be recognised, we believe
literature and is significant. The study’s merits and short- that the VKF approach should be used with similar data
comings are discussed further below. sets.
2. The minimal number of participants in the study’s data
4.1 Advantages and limitations set may be deemed insufficient for the construction of
ML models. We intend to address this shortcoming in
According to the results obtained in this study, this research future studies by using the VKF approach and ML
can contribute to the systems for automatic diagnosis of models for larger data sets.
PD. The advantages of this study can be summarised as 3. The results produced through the use of the VKF
follows: approach are promising for the diagnosis of PD.
However, to improve the method’s applicability, it
1. In this work, a novel and underutilised technically
should be evaluated with different PD datasets or in
powerful data set that is recognised as valid in research
research for the identification of various neurological
was employed to diagnose PD.
diseases.
2. The VKF approach was applied to EEG readings for
4. This study can be considered as a preliminary study. In
the first time in a study to diagnose neurological
future studies, systems for early and automatic diag-
diseases. Therefore, this study is unique research in this
nosis of PD can be developed in collaboration with
domain.
expert clinicians.
3. The analysis revealed that the VKF method used in this
study performed well in the ML models used in the
study when compared to similar recent studies [47, 50].
4. The accuracy, sensitivity, specificity, F1-score, AUC, 5 Conclusion
PPV, FPR, kappa statistic, and NPV of ML models
were evaluated in this study. The categorisation was This research presents an ML-based study for the auto-
carried out in four separate models using two different matic diagnosis of PD by applying the VKF method to
CV (10/L) methodologies. Furthermore, the processing EEG signals. In the presented study, the VKF method is
load was lowered by selecting the important features applied to the EEG signals of PD (ON–OFF)-HC groups
using the feature selection method. As a result, we and the signals are decomposed into sub-bands. Various
believe that the findings of this study are solid and features (SE, HPs, PS, KC, KFD) were extracted from
dependable. these signals and the significant ones were determined by

123
9310 Neural Computing and Applications (2024) 36:9297–9311

applying the Chi-squared test. The determined features neural network for Parkinson’s disease detection using EEG
were applied as input to four different ML algorithms to signals. Electronics 10(14):1740
2. Tysnes OB, Storstein A (2017) Epidemiology of Parkinson’s
classify PD ON-HC and PD OFF-HC groups. The effect of disease. J Neural Transm 124(8):901–905. https://doi.org/10.
dopaminergic drugs was also analysed by performing 1007/s00702-017-1686-y
classification according to PD ON–OFF states. Four dif- 3. Politis M, Wu K, Molloy S, Bain PG, Chaudhuri KR, Piccini P
ferent ML models were analysed and their performances (2010) Parkinson’s disease symptoms: the patient’s perspective.
Mov Disord 25(11):1646–1651
were compared. In addition, in this study, the VKF method 4. Shaban M, Amara AW (2022) Resting-state electroencephalog-
was compared with EMD and VMD methods, which are raphy based deep-learning for the detection of Parkinson’s dis-
widely used in research. According to the findings of the ease. PLoS ONE 17(2):e0263159
study, the VKF approach outperforms all ML algorithms 5. Perlmutter JS (2009) Assessment of Parkinson disease manifes-
tations. Curr Protoc Neurosci 49(1):10–11
when compared to other methods. We believe that the VKF 6. Khare SK, Bajaj V, Acharya UR (2021) PDCNNet: An automatic
approach results are promising for the development of framework for the detection of Parkinson’s disease using EEG
systems for the automatic diagnosis of PD and can aid signals. IEEE Sens J 21(15):17017–17024
expert doctors in decision-making. The VKF approach has 7. Yao S, Zhu J, Li S, Zhang R, Zhao J, Yang X, Wang Y (2022)
Bibliometric analysis of quantitative electroencephalogram
never been used to identify a neurological condition using research in neuropsychiatric disorders from 2000 to 2021. Front
EEG signals before. We believe that the VKF approach can Psych 13:830819
produce positive results in research for the diagnosis of 8. Dubey AK, Saraswat M, Kapoor R, Khanna S (2022) Improved
various neurological illnesses. In future studies, it is plan- method for analyzing electrical data obtained from EEG for better
diagnosis of brain related disorders. Multimed Tools Appl
ned to carry out studies with more samples on the inves- 81(24):35223–35244
tigation of Parkinson’s disease and drug effects in 9. Mei J, Desrosiers C, Frasnelli J (2021) Machine learning for the
cooperation with clinicians. As a result, we aim to support diagnosis of Parkinson’s disease: a review of literature. Front
the findings of this study by using the VKF approach in Aging Neurosci 13:633752
10. Li K, Ao B, Wu X, Wen Q, Ul Haq E, Yin J (2023) Parkinson’s
different data sets in future studies to demonstrate the disease detection and classification using EEG based on deep
effectiveness of the VKF method. CNN-LSTM model. Biotechnol Genetic Eng Rev
11. Anjum MF, Dasgupta S, Mudumbai R, Singh A, Cavanagh JF,
Narayanan NS (2020) Linear predictive coding distinguishes
Funding Open access funding provided by the Scientific and Tech- spectral EEG features of Parkinson’s disease. Parkinsonism Relat
nological Research Council of Türkiye (TÜBİTAK). Disord 79:79–85
12. Murugappan M, Alshuaib W, Bourisly AK, Khare SK, Sruthi S,
Data availability Datasets used are online available: San Diego Bajaj V (2020) Tunable Q wavelet transform based emotion
dataset: https://openneuro.org/datasets/ds002778/versions/1.0.4. classification in Parkinson’s disease using electroencephalogra-
Additional data and information can be provided upon request. phy. PLoS ONE 15(11):e0242014
13. Aljalal M, Aldosari SA, AlSharabi K, Abdurraqeeb AM, Alturki
FA (2022) Parkinson’s disease detection from resting-state EEG
Declarations signals using common spatial pattern, entropy, and machine
learning techniques. Diagnostics 12(5):1033
14. Motin MA, Mahmud M, Brown DJ (2022) Detecting Parkinson’s
Conflict of interest The authors declare no conflicts of interest in this disease from electroencephalogram signals: an explainable
study. machine learning approach. In: 2022 IEEE 16th ınternational
conference on application of ınformation and communication
Open Access This article is licensed under a Creative Commons
technologies (AICT). IEEE, pp 1–6
Attribution 4.0 International License, which permits use, sharing,
15. Kamalakannan N, Balamurugan SPS, Shanmugam K (2021) A
adaptation, distribution and reproduction in any medium or format, as
novel approach for the early detection of Parkinson’s disease
long as you give appropriate credit to the original author(s) and the
using EEG signal. Technology (IJEET) 12(5):80–95
source, provide a link to the Creative Commons licence, and indicate
16. Karakaş MF, Latifoğlu F (2023) Distinguishing Parkinson’s
if changes were made. The images or other third party material in this
disease with GLCM features from the hankelization of EEG
article are included in the article’s Creative Commons licence, unless
signals. Diagnostics 13(10):1769
indicated otherwise in a credit line to the material. If material is not
17. Biswas SK, Nath Boruah A, Saha R, Raj RS, Chakraborty M,
included in the article’s Creative Commons licence and your intended
Bordoloi M (2023) Early detection of Parkinson disease using
use is not permitted by statutory regulation or exceeds the permitted
stacking ensemble method. Comput Methods Biomech Biomed
use, you will need to obtain permission directly from the copyright
Engin 26(5):527–539
holder. To view a copy of this licence, visit http://creativecommons.
18. Sengur A, Akbulut Y, Guo Y, Bajaj V (2017) Classification of
org/licenses/by/4.0/.
amyotrophic lateral sclerosis disease based on convolutional
neural network and reinforcement sample learning algorithm.
Health Inf Sci Syst 5(1):1–7. https://doi.org/10.1007/s13755-017-
References 0029-6
19. Oh SL, Hagiwara Y, Raghavendra U, Yuvaraj R, Arunkumar N,
1. Loh HW, Ooi CP, Palmer E, Barua PD, Dogan S, Tuncer T, Murugappan M, Acharya UR (2020) A deep learning approach
Acharya UR (2021) GaborPDNet: gabor transformation and deep for Parkinson’s disease diagnosis from EEG signals. Neural
Comput Appl 32:10927–10933

123
Neural Computing and Applications (2024) 36:9297–9311 9311

20. Qiu L, Li J, Pan J (2022) Parkinson’s disease detection based on 37. Krogh A (2008) What are artificial neural networks? Nat
multi-pattern analysis and multi-scale convolutional neural net- biotechnol 26(2):195–197
works. Front Neurosci 16:957181 38. Istiadi I, Rahman AY, Wisnu ADR (2023) Identification of tempe
21. Rockhill AP, Jackson N, George J, Aron A, Swann NC (2020) fermentation maturity using principal component analysis and
UC San Diego resting state EEG data from patients with k-nearest neighbor. Sinkron: Jurnal dan Penelitian Teknik İnfor-
Parkinson’s disease. Available https://openneuro.org/datasets/ matika 8(1):286–294
ds002778/versions/1.0.4 39. Charbuty B, Abdulazeez A (2021) Classification based on deci-
22. George JS, Strunk J, Mak-McCully R, Houser M, Poizner H, sion tree algorithm for machine learning. J Appl Sci Technol
Aron AR (2013) Dopaminergic therapy in Parkinson’s disease Trends 2(01):20–28
decreases cortical beta band coherence in the resting state and 40. Phyu TZ, Oo NN (2016) Performance comparison of feature
increases cortical beta band power during executive control. selection methods. In: MATEC web of conferences, vol 42. EDP
NeuroImage: Clin 3:261–270 Sciences, p 06002
23. Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q, Liu 41. Orhanbulucu F, Latifoğlu F, Baydemir R (2023) A new hybrid
HH (1998) The empirical mode decomposition and the Hilbert approach based on time frequency images and deep learning
spectrum for nonlinear and non-stationary time series analysis. methods for diagnosis of migraine disease and investigation of
İn: proceedings of the Royal Society of London. Series A: stimulus effect. Diagnostics 13(11):1887
mathematical, physical and engineering sciences, 454(1971), 42. Barua PD, Dogan S, Tuncer T, Baygin M, Acharya UR (2021)
903–995 Novel automated PD detection system using aspirin pattern with
24. Pandey P, Seeja KR (2022) Subject independent emotion EEG signals. Comput Biol Med 137:104841
recognition from EEG using VMD and deep learning. J King 43. Rümeysa E, İleri R, Latifoğlu F (2021) A new approach to
Saud Univ-Comput Inf Sci 34(5):1730–1738 detection of Parkinson’s disease using variational mode decom-
25. Baydemir R, Latifoğlu F, Orhanbulucu F (2022) classification position method and deep neural networks. In: 2021 medical
mental workload levels from EEG signals with 1D convolutional technologies congress (TIPTEKNO). IEEE, pp 1–4
neural network. Eur J Res Dev 2(4):13–23 44. Safi K, Aly WHF, AlAkkoumi M, Kanj H, Ghedira M, Hutin E
26. Dragomiretskiy K, Zosso D (2013) Variational mode decompo- (2022) EMD-based method for supervised classification of
sition. IEEE Trans Signal Process 62(3):531–544. https://doi.org/ Parkinson’s disease patients using balance control data. Bio-
10.1109/TSP.2013.2288675 engineering 9(7):283
27. Jiang L, Zhou X, Che L, Rong S, Wen H (2019) Feature 45. Latifoğlu F, Bulucu FO, Ileri R (2021) Detection of amyotrophic
extraction and reconstruction by using 2D-VMD based on carrier- lateral sclerosis disease by variational modedecomposition and
free UWB radar application in human motion recognition. Sen- convolution neural network methods from event-relatedpotential
sors 19(9):1962 signals. Turk J Electr Eng Comput Sci 29(8):2840–2854
28. Vold H, Leuridan J (1993) High resolution order tracking at 46. Taran S, Bajaj V (2018) Clustering variational mode decompo-
extreme slew rates, using Kalman tracking filters (No. 931288). sition for identification of focal EEG signals. IEEE Sens Lett
SAE Technical Paper 2(4):1–4
29. Yan Z, Wang L, Hu A Feature extraction by enhanced time- 47. Khare SK, Bajaj V, Acharya UR (2021) Detection of Parkinson’s
frequency representation based on Vold-Kalman filter. Liang and disease using automated tunable Q wavelet transform technique
Hu, Aijun, Feature extraction by enhanced time-frequency rep- with EEG signals. Biocybern Biomed Eng 41(2):679–689
resentation based on Vold-Kalman filter 48. Duan X, Feng Z (2023) Time-varying filtering for nonstationary
30. Liu T, Luo Z, Huang J, Yan S (2018) A comparative study of four signal analysis of rotating machinery: principle and applications.
kinds of adaptive decomposition algorithms and their applica- Mech Syst Signal Process 192:110204
tions. Sensors 18(7):2120 49. Hamidi A, Yousefi M (2023) Classification of EEG signals to
31. Li Y, Han Z, Wang Z (2020) Research on a signal separation detect Parkinsons Disease using a computationally method. Pre-
method based on Vold-Kalman filter of improved adaptive print at arXiv:2305.02234
instantaneous frequency estimation. IEEE Access 50. Aljalal M, Aldosari SA, Molinas M, AlSharabi K, Alturki FA
8:112170–112189 (2022) Detection of Parkinson’s disease from EEG signals using
32. Henriques T, Ribeiro M, Teixeira A, Castro L, Antunes L, Costa- discrete wavelet transform, different entropy measures, and
Santos C (2020) Nonlinear methods most applied to heart-rate machine learning techniques. Sci Rep 12(1):22547
time series: a review. Entropy 22(3):309 51. Khan MS, Nath TD, Hossain MM, Mukherjee A, Hasnath HB,
33. Li X, Song D, Zhang P, Zhang Y, Hou Y, Hu B (2018) Exploring Meem TM, Khan U (2023) Comparison of multiclass classifica-
EEG features in cross-subject emotion recognition. Front Neu- tion techniques using dry bean dataset. Int J Cogn Comput Eng
rosci 12:162 4:6–20
34. Işık Ü, Güven A, Batbat T (2023) Evaluation of emotions from 52. Tharwat A (2020) Classification assessment methods. Appl
brain signals on 3D VAD space via artificial intelligence tech- Comput Inform 17(1):168–192
niques. Diagnostics 13(13):2141 53. Saeedi S, Rezayi S, Keshavarz H, Niakan Kalhori SR (2023)
35. Altıntop ÇG, Latifoğlu F, Akın AK, Bayram A, Çiftçi M (2022) MRI-based brain tumor detection using convolutional deep
Classification of depth of coma using complexity measures and learning methods and chosen machine learning techniques. BMC
nonlinear features of electroencephalogram signals. Int J Neural Med Inform Decis Mak 23(1):16
Syst 32(05):2250018
36. Chandrasekharan S, Jacob JE, Cherian A, Iype T (2023) Publisher’s Note Springer Nature remains neutral with regard to
Exploring recurrence quantification analysis and fractal dimen- jurisdictional claims in published maps and institutional affiliations.
sion algorithms for diagnosis of encephalopathy. Cogn Neurodyn.
1–14

123