Novel automated PD detection system using aspirin pattern with EEG signals

https://doi.org/10.1016/j.compbiomed.2021.104841Get rights and content

Highlights

  • A new local feature extractor is proposed by using the chemical graph of the aspirin.

  • A high accurate EEG signal classification method is proposed for PD detection using the proposed aspirin pattern.

  • This work attained over 93% classification accuracies for all cases using LOSO validation.

Abstract

Background and objective

Parkinson's disease (PD) is one of the most common diseases worldwide which reduces quality of life of patients and their family members. The electroencephalogram (EEG) signals coupled with various advanced machine-learning algorithms have been widely used to detect PD automatically. In this paper, we propose a novel aspirin pattern to detect PD accurately using EEG signals.

Method

In this research, the feature generation ability of a chemical graph is investigated. Therefore, this work presents a new graph-based aspirin model for automated PD detection using EEG signals. The proposed method consists of (i) multilevel feature generation phase involving new aspirin pattern, statistical moments, and maximum absolute pooling (MAP), (ii) selection of most discriminative features using neighborhood component analysis (NCA), and (iii) classification using k nearest neighbor (kNN) for automated detection of PD and (iv) iterative majority voting.

Results

A public dataset has been used to develop the proposed model. Two cases are created, and these cases consisted of two classes. Leave one subject out (LOSO) validation have been used to calculate robust results. Our proposal achieved 93.57% and 95.48% classification accuracies for Case 1 and Case 2 respectively.

Conclusion

Our developed automated PD model is accurate and equipped to be tested with more diverse EEG datasets.

Introduction

Parkinson's disease (PD) is one of the neurodegenerative and age-related diseases common in many countries [1]. It is caused due to loss of dopamine secretion from the basal ganglia region in the brain [2]. Aging is the main reason for the decrease in dopamine secretion in the age-related neurons of the human brain. The decrease in the amount of dopamine produced by neurons negatively affects the brain condition. Thus, PD is more prevalent in older people [[3], [4], [5]]. Studies have shown that approximately 0.3% of the world population is affected by this disease [6]. Approximately 80% of PD patients are 60 years old and above [7]. In addition, according to another study, the risk of developing PD is 1.5 times higher in men [8]. Moreover, this disease peaks between 85 and 89 years of aged people [7,8]. Behavioral, cognitive, and emotional disorders can be observed in PD patients. The most common symptoms in PD patients are slowed movements, difficulty speaking and shaking hands and feet during rest. This situation negatively affects the daily lives of PD patients [9,10].

In the past and today, the diagnosis of Parkinson's disease is made based on clinical and anamnesis findings [11]. The most important characteristics of PD are the slowed movements and hand tremors [12]. In addition, mood disorders such as depression and rapid eye movements (REM) sleep disorders can help in making a preliminary diagnosis of PD [13]. In addition to these physiological findings, various pathological findings are also encountered in PD. One of these findings is alpha synuclein (α-synuclein) [14,15]. This protein plays an important role in neurodegenerative disorders such as PD [16,17]. Although various physiological and pathological findings are used in the diagnosis of PD, these findings do not provide a complete evidence in distinguishing healthy and diseased individuals [18].

Nowadays, experts use new techniques and automated systems for the diagnosis of PD in addition to the identified findings [[19], [20], [21]]. Automated systems developed using EEG signals are used to diagnose PD disease [9,10]. The EEG signals are nonlinear and non-stationary in nature. Therefore, various nonlinear features are obtained from EEG signals [22] to diagnose mental diseases such as PD, epilepsy, schizophrenia, Alzheimer's disease using EEG signals [[23], [24], [25], [26], [27]].

Various deep learning and hand-crafted methods have been used to obtain high performance [[28], [29], [30]]. In this study, a new hand-crafted method is proposed for automatic detection of PD using EEG signals obtained from the OpenNeuro [31] dataset. The state-of-the-art techniques developed for PD using EEG signals are presented in the next section.

Many methods have been proposed for the automated diagnosis of PD using EEG signals in the literature. Few recent studies are presented below.

Khoshnevis and Sankar [32] proposed a method to detect PD using a higher-order statistics technique. Their study obtained an accuracy of 87.00% with an ensemble RUSBoosted trees classifier using 40 subjects (20 PD patients and 20 normal subjects). Anjum et al. [33] presented an approach for PD detection with EEG signals. Their study obtained data from the University of New Mexico (27 PD and 27 healthy) and the University of Iowa (14 PD and 14 healthy). They have collected data using a 64-channel EEG device in 3 categories: eyes-closed, eyes-open, and eyes open-closed. Their method obtained an accuracy of 85.40% with 5-fold and 10-fold cross-validation strategies. Murugappan et al. [34] introduced a study using tunable Q wavelet transform and probabilistic neural network for PD detection using 20 PD and 20 healthy subjects obtained from University Kebangsaan Malaysia medical center. Their method obtained 93.88% and 96.16% detection accuracy for PD patients and healthy subjects, respectively. Xu et al. [29] suggested a pooling-based deep recurrent neural network method for PD detection using 20 PD and 20 healthy subjects. The EEG signals were collected from People's Hospital of Zhengzhou University using 14 EEG channels with a sampling rate of 128-Hz. They reported specificity, sensitivity, and precision ratios of 91.81, 84.84, and 88.31%, respectively. Their method is computationally intensive. Yuvaraj et al. [3] presented a higher-order spectra features-based approach for PD detection using 20 PD and 20 healthy subjects with 14 EEG channels. Their method reported an accuracy of 99.62% with 10-fold cross-validation. Oh et al. [35] proposed an automated method for PD detection using 20 PD and 20 healthy subjects. They proposed a thirteen-layer convolutional neural network and obtained an accuracy of 88.25% with a ten-fold cross-validation strategy. Gunduz [36] presented an effective dimensionality reduction approach for the detection of PD. They reported an accuracy rate of 95.70% using a support vector machine (SVM) classifier. Naghsh et al. [37] proposed an automatic method for the diagnosis of PD. They used a combination of magnetoencephalography and EEG signals obtained from 20 subjects (10 PD and 10 healthy) in their study. Their proposed method obtained an accuracy of 95.00%. Chaturvedi et al. [38] proposed a PD detection method with machine learning algorithm. The main purpose of their study is to distinguish PD patients from healthy individuals. In their study, 50 PD subjects and 41 healthy subjects were selected. They attained an accuracy rate of 78.0% with random forest. Shah et al. [39] presented a compact deep hybrid network based on deep learning for PD detection using 12 EEG channels. This network is named Dynamical System Generated Hybrid Network in the study. PD-on and PD-off medication data of 28 subjects were used in this study. Their proposed dynamical system-generated hybrid network reported accuracy of 99.22% in classifying the two groups. This method has high complexity. Chu et al. [40] suggested a spatiotemporal EEG microstate analysis method. This method was performed for drug-free patients. Their study concluded that PD could be expressed with a unique spatial microstate differently from healthy subjects. Vanegas et al. [18] applied PD detection approach with machine learning algorithms. In the study, they aimed to detect the most dominant EEG spectra. They concluded that PD patients develop an abnormally high background noise level and an overreaction gain profile for contrast response function. Gil-Martin et al. [41] used convolutional neural networks for PD detection. For this purpose, the drawing movements of the subjects were used. In their study, Parkinson Disease Spiral Drawings Using Digitized Graphics Tablet dataset [42] was used for experimental results. According to the results obtained, the accuracy rate was calculated as 96.50%.

The primary motivation to develop a novel model is to achieve high PD detection performance using multi-channel EEG signal datasets. Hence, a multilevel feature generator-based automatic EEG signal detection model is presented. Multilevel feature generators have used pooling functions minimum, maximum and average. These three (minimum, maximum and average) pooling decomposers have a routing problem. Hence, maximum absolute pooling (MAP) [43,44] is used to decompose the signal in this work. The MAP routes both minimum and maximum values depending on their condition. In the feature generation, two feature generators are used to extract both textural and statistical features. To generate textural features, a graph-based feature generation function is presented. In this function, four aspirin graphs are used to investigate feature generation ability of the molecular structure of aspirin. The proposed aspirin function generated 576 features from the EEG signal. Also, 30 statistical features are generated using linear and nonlinear statistical features. The presented features generator has four levels and (576 + 30) × 4 = 2424 features are extracted from the EEG signals. The NCA features selector selected 256 features (LBP generates 256 features) from an EEG signal. Therefore, we selected 256 features from the generated 2424 features. In the classification phase, kNN classifier automatically to detected PD using selected 256 features.

The novelties and main contributions of this work are as follows. Graph based learning methods are very popular in the literature because they reached high classification performance. A novel graph based feature generator, graph of the aspirin molecular structure is used to create a pattern. Furthermore, three binary feature generation functions are used in the proposed new aspirin pattern. Moreover, four cases are defined using the used PD dataset. The proposed Aspirin pattern based EEG classification model reached high classification performances the whole cases.

Section snippets

Materials

In this work, public EEG dataset is used to detect PD automatically. The details of this dataset is given below.

OpenNeuro [31] EEG dataset was collected at the University of San Diego in the resting state. In this dataset, 15 PD (8 Female, 7 Male) and 16 healthy subjects (9 Female, 7 Male) were used. The average ages of PD and healthy subjects were 63.2 ± 8.2, 63.5 ± 9.6, respectively. The average disease duration of PD and healthy individuals is the same in both and is 4.5 ± 3.5. In addition,

Overview of the proposed model

EEG signals have characteristic features and these features are the hidden salient signatures present in the EEG signals. The primary aim of this work is to find of the hidden patterns present in the EEG signals to predict PD accurately. To validate this idea, this work proposes a new hand-crafted feature extraction based automatic PD diagnosis model using EEG signals. The main steps of this model are feature generation, feature selection, and classification. The MAP [43,44] decomposition,

Results

The presented aspirin pattern-based EEG signal classification model was implemented on the MATLAB (2020b) programming environment. A simple configured laptop with 8 GB RAM, intel i7-7700 processor with 3.2 GHz and 512 GB disk was used. We have used two cases to obtain the results and the details of the cases are described below.

Case 1

This case consists of healthy and PD without drug classes. It has 3032 (1532 healthy, and 1500 PD without drug files) observations.

Case 2

This case consists of healthy and PD

Discussions

This work presents a new aspirin pattern-based EEG signal classification model. Aspirin pattern is proposed to investigate feature extraction ability of a chemical graph using an EEG signal as the input signal. We have used two cases and these cases are defined in Section 4. Moreover, LOSO validation have used to obtain the results. Our presented model attained 93.57% and 95.48% classification accuracies for Case 1 and Case 2, respectively. We have used kNN classifier for classification.

In our

Conclusion

A novel aspirin pattern feature generator-based automated PD detection system is proposed using EEG signals. This work employs a new aspirin pattern-based feature generation function and NCA feature selector coupled with a kNN classifier for automated PD detection. In order to show the superiority of the proposed model, LOSO validation has been used in this work. Our model attained 93.57% and 95.48% classification accuracies for Case 1 and Case 2 respectively. Moreover, the results of our model

Declaration of competing interest

There is no ‘Conflict of Interest’ in the publication of the manuscript “Novel automated PD detection system using aspirin pattern with EEG signals”.

References (60)

  • V. Jahmunah et al.

    Automated detection of schizophrenia using nonlinear signal processing methods

    Artif. Intell. Med.

    (2019)
  • C. Sridhar et al.

    Diagnosis of attention deficit hyperactivity disorder using imaging and signal processing techniques

    Comput. Biol. Med.

    (2017)
  • E. Aydemir et al.

    A Tunable-Q wavelet transform and quadruple symmetric pattern based EEG signal classification method

    Med. Hypotheses

    (2020)
  • M. Baygin et al.

    Automated ASD detection using hybrid deep lightweight features extracted from EEG signals

    Comput. Biol. Med.

    (2021)
  • M.F. Anjum et al.

    Linear predictive coding distinguishes spectral EEG features of Parkinson's disease

    Park. Relat. Disord.

    (2020)
  • H. Gunduz

    An efficient dimensionality reduction method using filter-based feature selection and variational autoencoders on Parkinson's disease classification

    Biomed. Signal Process Contr.

    (2021)
  • S.A.A. Shah et al.

    Dynamical system based compact deep hybrid network for classification of Parkinson disease related EEG signals

    Neural Network.

    (2020)
  • C. Chu et al.

    Spatiotemporal EEG microstate analysis in drug-free patients with Parkinson's disease

    Neuroimage: Clinical.

    (2020)
  • M. Baygin et al.

    Automated arrhythmia detection with homeomorphically irreducible tree technique using more than 10,000 individual subject ECG records

    Inf. Sci.

    (2021)
  • S. Dogan et al.

    Application of substitution box of present cipher for automated detection of snoring sounds

    Artif. Intell. Med.

    (2021)
  • S. Raghu et al.

    Classification of focal and non-focal EEG signals using neighborhood component analysis and machine learning algorithms

    Expert Syst. Appl.

    (2018)
  • T. Tuncer et al.

    Automated detection of Parkinson's disease using minimum average maximum tree and singular value decomposition method with vowels

    Biocybernetics and Biomedical Engineering

    (2020)
  • Y. Liao et al.

    Use of k-nearest neighbor classifier for intrusion detection

    Comput. Secur.

    (2002)
  • Y. Xu et al.

    Coarse to fine K nearest neighbor classifier

    Pattern Recogn. Lett.

    (2013)
  • S.K. Khare et al.

    Detection of Parkinson's disease using automated tunable Q wavelet transform technique with EEG signals

    Biocybernetics and Biomedical Engineering

    (2021)
  • A. Ortiz et al.

    Parkinson's disease detection using isosurfaces-based features and convolutional neural networks

    Front. Neuroinf.

    (2019)
  • M. Murugappan et al.

    Emotion classification in Parkinson's disease EEG using RQA and ELM

  • R. Yuvaraj et al.

    A novel Parkinson's Disease Diagnosis Index using higher-order spectra features in EEG signals

    Neural Comput. Appl.

    (2018)
  • Q.W. Oung et al.

    Empirical wavelet transform based features for classification of Parkinson's disease severity

    J. Med. Syst.

    (2018)
  • R. Yuvaraj et al.

    Hemispheric asymmetry non-linear analysis of EEG during emotional responses from idiopathic Parkinson's disease patients

    Cognitive neurodynamics

    (2016)
  • Cited by (24)

    • EEG classification based on Grassmann manifold and matrix recovery

      2024, Biomedical Signal Processing and Control
    • An accurate automated speaker counting architecture based on James Webb Pattern

      2023, Engineering Applications of Artificial Intelligence
    View all citing articles on Scopus
    View full text