Diagnosis of Parkinson’s disease using EEG and fMRI

Parkinson’s disease (PD) is a brain disorder that leads to shaking, stiffness, and difficulty with walking, balance, and coordination. Parkinson’s symptoms usually begin gradually and get worse over time. We have developed a framework to find solution for early diagnose of PD by investigating the topological properties of functional brain networks within fMRI and EEG Signals. After the construction partial correlation matrices of 160 regions from Dosenbach brain from fMRI image, six features were extracted. As well as extracted five features from EEG signals and these 11 inputs were given as input to adaboost classifier. This system has produced 93.45% accuracy and the outcome is significantly higher accuracy when compared to earlier works.


Introduction
Parkinson's disease (PD) is a progressive nervous system disorder that affects movement because of less dopamine secretions in human Brain. An estimated seven to 10 million people worldwide have Parkinson's disease [8]. The prevalence of the disease ranges from 41 people per 100,000 in the fourth decade of life [9] to more than 1,900 people per 100,000 among those who are 80 and older. There is no homogenous and large epidemiological data on PD [23] from India. The prevalence [27] rate of 14.1 per 100,000 amongst a population of 63,645 from rural Kashmir in the northern part of India. The prevalence rate [23] over the age of 60 years was 247/100,000. A low prevalence rate [30] of 27/100,000 was reported from Bangalore, in the southern part of India, and 16.1/100,000 from rural Bengal, in the eastern part of India. The prevalence rate of 328.3/100,000 among a population [30] of 14,010 Parsis living in colonies in Mumbai, Western India.
According to studies [11,7,13,15,28,25], because of late diagnose, elderly people are having difficulty in stiffness, walking, balancing, talking and coordination. Doctors rely on EEG and MRI to establish the precise nature of disease. Manual inspection of images for diagnose is both a tedious and time-consuming process. On top of that, doctors will get too tired to perform the task reliably after examining numerous images. Automatic interpretation of medical images can relieve some of the labour-intensive work of the doctors thus improving the accuracy of the diagnosis.
In this research, we have emphasised diagnosing PD through the properties of fMRI & EEG signal. So, quick and accurate diagnosis is crucial to the success of any prescribed treatment. Depending on the human experts alone for such a critical matter have cause intolerable errors. Hence, the idea of automatic diagnosis procedure has always been an appealing one. The main goal is to diagnose PD from EEG and fMRI images using MATLAB software.
The motivations of this system are: (i) saving time in diagnosis for PD patients and (ii) to reduce human errors because doctors in hospitals manually inspect a large number of fMRI /EEG images for diagnosing Parkinson. Manual inspection is tedious and time consuming. A tired radiologist has been found to miss diagnosis among healthy ones. Computer vision system can help to screen fMRI and EEG images for suspicious cases and alarm the doctors.
The overall objective is to diagnose Parkinson through medical images using image processing techniques, the specific objectives are.
(i) To improve the accuracy of detecting parkinson over the recent identification methods. (ii) To develop a method to diagnose very faster than the existing methodology.
The main aim of this work is to analyse Parkinson's disease detection by investigating the topological properties of functional brain networks within fMRI and EEG Signals of Healthy Control (normal) and PD patients. The main contribution in this work is a hybrid technique is used to diagnose PD in earlier stage.
The highlights of this work are.

Related works
Martin Gottlich et al. [16] used graph-theory based technique to measure whole-brain intrinsic connectivity. The network parameters used in this study would be helpful to track disease state and characterize subtypes of PD patients related to cognitive dysfunctions or other non-motor symptoms. The future research will be focused on how network changes according to the stages of disease or the type of medication. Hugo-Cesar Baggio et al. [3] reported that complex network analysis through resting-state fMRI is very helpful in investigating functional changes related to cognitive decline in PD. The limitation of this work is that the patients were evaluated only in on state, under the influence of dopaminergic medication. Yongbin Chen et al. [6] had used wholebrain functional connectivity as a classification feature to identify PD patients and healthy controls. The performance of this method had yielded an accuracy of 93.62%. The limitation in this study is the authors defined that the reliable results cannot be produced form the AAL regions as nodes. Secondly, the loss of global topological information is caused due to the concatenation of functional connections into a vector for subsequent feature selection. Zhang D et al. [36] reported that they had used network centrality, seed-based functional connectivity, and network efficiency analyses for full map of abnormal connectivity networks in PD with tremor that is distributed over cortical, subcortical, cerebellum, and brainstem sites and observed the functional changes were recorded. The main limitation in this study is the number of participants involved is limited as well as this study had not answered how the process of reshaping the wholebrain network topology is done by the abnormal component, such as large-scale small-world organization and intermediate-scale community structure.
Abhishek M.S et al. [1] reported that voice signals played a major role in predicting Parkinson's disease and concluded that genetic algorithm is the most common method used to extract features voice signal properties and the prediction accuracy was 95%. Athanasios Tsanas et al. [31], the authors had used 132 dysphonia measures from sustained vowels and this work could be used to discriminate PD subjects from healthy controls and was tested. The authors reported that by combining the information from both sustained vowels and running speech to improve the accuracy, the efficiency would be increased. Alex Frid et al. [14] developed an automatic system for quantification and classification of Parkinson's disease directly from natural speech using the Machine Learning technique.
Jens Barth et al. [4] developed a method by combining hand and gait motor function impairment to differentiate between PD patients and healthy controls and an accuracy of 97% was obtained using Adaboost classifier. Sama et al. [26] presented a method to detect dyskinesia and to characterize motor states for PD patients limited to 15 patients. Shyamal Patel et al. [19] had used accelerometer data to estimate the severity of symptoms and motor complications for the patients in Parkinson's disease.
A. Valli and Dr.G. Wiselin Jiji [32] discussed on the application of image processing techniques to diagnose Parkinson's disease diagnosis. Peter Drotar et al. [12] had used different handwriting modalities to classify PD. They reported that handwriting as a biomarker had produced only 89% accuracy. Pereira et al. [21] reported that handwritten trace as features for automatic classification of PD. The main problem in this study was high variability of the dataset, which was found difficult to be diagnosed. Salama A. Mostafa et al. [18] reported that they had used multiple feature evaluation and classification methods for improving the diagnoses of Parkinson's disease and the direction had been given for the future study by linking the dataset properties with the feature evaluation in the MFEA. Timothy J. Wroge et al. [34] reported that disease diagnosis and prediction is possible through automated machine learning architectures using only non-invasive voice biomarkers as features and an accuracy of 85% was noted. A.M. Ardi Handojoseno et al. [17] stated that combining the spatial, spectral and temporal features of surface EEG had helpful in the FOG in PD and to predict transition from walking to freezing with 87% sensitivity and 73% accuracy, the approach was used. Rajamanickam Yuvaraj et al. [35] had used HOS features extracted from EEG signals for diagnosis of PD patients and using the SVM classifier, an optimum mean classification accuracy of 99.62%, mean sensitivity and specificity of 100.00 and 99.25% was got. Passos et al. [20] had used deep neural network called ResNet-50 to learn the patterns and extracted features from images draw by patients and produced 96% of identification rate. Prajapati et al. [22] reported that using topological properties of functional brain networks within healthy controls (HCs) and PD patients are extracted from fMRI images to diagnose PD or HC and it was found that essential insights from network changes to the clinically relevant information for the PD progression had been provide by it. Mosarrat Rumman et al. [24] had used SPECT images with ANN had produced an accuracy of 94%, sensitivity of 100% and specificity of 88%. Calimeri et al. [5] reported that to provide an effective support for neurologists in studying the evolution of neurological disorders, they had used the combination of Answer Set Programming and ANN. They had created graphs called weighted graph onnectome. The authors reported that connectome will vary based on the severity of disease.
The review of literature so far makes us feel that the method discussed here have focused on detecting some techniques to diagnose PD. Thus, there is a need to find a new method which could be more effective in diagnosing by considering all properties of Brain. Moreover, The review shows that the severity of Parkinson disease is not properly analysed because it didn't combine both fMRI and EEG Signals. A method is needed to analyse the disease in terms of localization and patterns. Accuracy of detection is another one main issue in the review because any misdiagnosed will cause major effect on the human body. Thus, there is a need to diagnose disease with high accuracy by reducing the error rate. The main contribution for this paper is to extract different patterns from fMRI and EEG signals for accurate diagnosis.

Methodology
An Architecture is proposed for Parkinson's disease detection by investigating the topological properties of functional brain networks within fMRI and EEG Signals of Healthy Control (normal) and PD patients. Figure 1 shows the overview of Proposed Architecture. For fMRI the functional whole-brain connectome was constructed by thresholding partial correlation matrices of 160 regions from Dosenbach brain atlas. 160 × 160 functional correlation matrix was constructed using the Pearson correlation. From the graph theory approach, network metrics were analysed. For EEG spatial and Bispectrum features are extracted. Finally, Adaboost Classifier is used to classify whether it is normal or PD.

Dataset used
The input images were taken from the OpenNEURO which is open science neuro informatics database storing datasets from human brain imaging research studies. The dataset contains raw fMRI scans, raw EEG in Brain Vision format where the subjects include fMRI of 100 patients with Parkinson's and 100 with Healthy control, EEG signals were recorded in closed eye resting state for 100 patients with Parkinson's and 100 with Healthy Control. In Parkinson's 25 Male patients and 75 Female patients and in Healthy Control 10 Female and 90 male patients were taken. Figures 2 and 3 for the input Healthy Control and PD and Figs. 4 and 5 are Input image EEG for Healthy Control and PD respectively.

Network construction and functional connectome
The two fundamental elements of the network are edges and nodes, where nodes represent the brain regions and edges depict the functional connectivity between two brain regions or nodes. The Region of interest for this category are frontoparietal, cingulo-opercular, sensorimotor, occipital, and cerebellum were selected from Dosenbach atlas [29] to draw functional connectome and the output is shown in Fig. 6.
GRETNA tool is used to find out the functional connectivity matrices from brain images [33] and edges for the connectivity has been calculated using Pearson corelation coefficients [2]. The output is shown in Figs. 7 and 8. Each region is considered as a node to construct the brain network and the output is shown in Figs. 9 and 10.

Feature extraction: fMRI
We have extracted [10] the parameters betweenness mean, betweenness standard deviation, page rank mean, page rank standard deviation, centrality degree mean, centrality degree standard deviation, clustering coefficient, assortivity, closeness, centrality closeness values from network created from fMRI.
Where, v * is the node with maximum betweenness and c bet ð:Þis the normalized betweenness.
Page rank Where, D is a diagonal matrix, I is an N x N identity matrix, a is weight on the edges from vertex v.
Where, A is a matrix with vertices i ,j where i 6 ¼ j

Clustering Coefficient
Where, d v ð Þ is the degree of vertex v, N v ð Þ is set of all nodes that are a distance 1 from a vertex v and A is the matrix.
Where, e jk refers to the joint excess degree probability for nodes with excess degrees j and k.
q k is a normalized distribution of a randomly selected node, given by q k ¼ ðkþ1Þp k P jj p j and σ q is the standard deviation of the distribution q k Where, d (u, v) is the distance to all the other nodes in the network.

Feature extraction: EEG
The bispectrum is an advanced signal processing technique based on higher order statistics which considers both the amplitude and the degree of phase coupling of a signal. In contrast to traditional power spectrum, which quantifies the power of a time series over frequency, higher order spectral (HOS) analysis employs the Fourier transform of higher order correlation functions to explore the existence of quadratic (and cubic) non-linear coupling information [35]. The extracted bispectrum features are Bispectrum, Cumulant, Element frequency and Lag vectors. Spatial features include Wavelet Coherence mean, Wavelet Coherence SD, Wavelet Cross Spectrum mean, Wavelet Cross Spectrum SD [17].

Bispectrum
Bðf Where, F denotes the Fourier transform of the signal, and F* its conjugate.
Where, k nz represents the nth order of the obtained variable (z). denotes the nth-order cumulant of the i th component random variable.k n;x n Lag vector Where, sign is the signum function that discards phase difference of 0 mod π. The LV ranges between 0 and 1, with 0 indicating no coupling of instantaneous coupling due to volume conduction and 1 indicating true, lagged interaction.
Wavelet Cross spectrum wcs n jk t; s ð Þ ¼ w n j ðt; sÞw n k ðt; sÞ * ð10Þ Where, t is the time and s is frequency (scale), as a result the WCS is complex valued.

Wavelet Coherence
Where T is the time around which the coherence is calculated, i is the current index, and f is the frequency. The summations are carried around a variable segment size Δ, which is inversely proportional to frequency.
Six feature descriptors from functional Connectome in fMRI and five feature descriptors in EEG have been extracted. The eleven feature vectors are considered and given as input to classifier in this study.

Classification
Using fMRI and EEG signal, we have performed classification operation, to identify whether the input is healthy control or Parkinson affected. In An AdaBoost classifier [10], we have used Naive Bayes classifier as base estimator. In this work, we have used n_estimators as 45 and learning rate as 1. The misclassified training samples get more weights, and the test error keeps decreasing even after 700 iterations.

Discussion
The implementation results were tested using various EEG & fMRI images. The performance figure shown here for an optimized MATLAB implementation was obtained. The performance was measured on 2.40 GHz Pentium(4) CPU with 512 MB of RAM running Microsoft Windows XP version 2002. All programs applied in simulating the algorithms are designed by MATLAB 7.0. All input data were annotated by Radiologists which is considered as ground truth data.
The proposed method performs the feasibility of a functional MRI and EEG based computational biomarker, which can quantify the functional connectivity patterns between healthy controls and PD patients. To differentiate PD and HCs, we utilized a novel graph theoretical approach to determine global and nodal measures from rs-fMRI data and Spatial, Bispectrum Features from EEG signals.
The present study examined the topological characteristics of brain functional connectomes among 100 PD and 100 HC subjects. The whole-brain functional connectome was constructed from rs-fMRI using a graph theory approach, which characterizes 160 regions from Dosenbach atlas. The mean correlation matrices were determined for both HC and PD groups. Features are extracted from both EEG and fMRI and given input to Adaboost classifier. Performance measures are stated below; Accuracy Where TP (true positive) was the number of pixels of each region detected both manually and the hybrid techniques. FP (false positive) was the number of pixels of each region detected by the hybrid techniques but not manually. TN (true negative) was the number of pixels of regions rejected both manually and the hybrid techniques. FN (false negative) was the number of pixels of regions detected manually, but not by the hybrid techniques. The result was compared with ground truth radiologist annotated data.
In this work, we had made 10-fold cross-validation, a single subsample is taken as the validation data for testing the model, and the remaining k-1 subsamples are used as training data. The main use advantage of this method is that all observations are used for both training and validation, and each observation is used for validation exactly once. The work is compared with earlier works and found that our proposed work has received higher accuracy. Table 1 shows the performance measures for the proposed work. Table 2 shows the comparison of performance measures with earlier works. The experimental result shows the good performance of the proposed approach which is comparable to state-of-art techniques (Fig. 11).  r Perceptron Neural Network classifier [17] K-Nearest Neighbour classifier [12] SVM classifier [35] Proposed Work Sens Specificity Precision Accuracy Fig. 11 Analysis of Performance Measure

Conclusion
In this work, by examining the topological organization of functional brain networks and EEG Signals, we have performed classification operation for Normal and PD patients using Adaboost classifier. We have got good accuracy when compared with other works. The drawback in this work is we have to increase accuracy by incorporating hybrid techniques. It can help the neurologists in faster and more accurate diagnosis during their screening itself.
Data availability Data available on request from the authors.
Code availability Software application.

Declarations
Conflict of interest We have no conflicts of interest to disclose.
Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.