A Review on Parkinson’s Disease Diagnosis using Machine Learning Techniques

Download Full-Text PDF Cite this Publication

Text Only Version

A Review on Parkinson’s Disease Diagnosis using Machine Learning Techniques

Anila M

Department of CSE,

Koneru Lakshmaiah Education Foundation Vaddeswaram,India

Dr. G. Pradeepini

Department of CSE

Koneru Lakshmaiah Education Foundation Vaddeswaram,India

Abstract Parkinson disease is a neurodegenerative disorder that affects nervous system and the root cause of it is falling rates of dopamine levels in the forebrain. It is a chronic degenerative disease with progressive illness, which means it develops new symptoms over time[20]. This happens with progressive neuronal loss in the substantia nigra of brain. People with PD cannot do their works as a normal human. Though clinical assessments considered ample amount of data that include various features, sometimes it is hard to decide whether a person is suffering from PD or not based on the type of data, feature selection methods help to solve this issue. Various methods are developed, proposed, and analyzed to detect the Parkinson disease, given the required data. This paper is a survey of predicting Parkinson disease using machine learning algorithms, various new technologies applied, and their accuracies achieved.

Keywords PD (Parkinson Disease), dopamine, SVM (Support Vector Machine), KNN (K Nearest Neighbor), ANN (Artificial Neural Network).

  1. INTRODUCTION

    Parkinson disease mainly effects central nervous system and is observed to be affected on many people globally. Most of the people suffering with PD are observed to be physically and emotionally draining. They even feel depressed, trouble concentrating on things, painful spasms etc. PD has a large spectrum of clinical features ranging from motor to nonmotor symptoms. Some of the motor symptoms are hypophonic speech, rigidity, resting tremor. Non-motor symptoms are as hallucinations, depression, constipation, sleeping disorders, cognitive impairment, and impulse control disorders. Non motor symptoms show sickness than motor symptoms [1,3]. Most of the cases, physicians find it difficult to envisage whether a given patient is already affected by the disease or is expected to develop the Parkinson's disease[7]. To conquer this, development of some computing model must be done that evaluates and summarizes the data of a given patient and predicts with adequate accuracy where he/she will have development of PD. Most of the PD patients are observed with symptoms called voice impairment which is known as dysphonia. There are several measures related to dysphonia, out of which voice related problem can be used to assess the patients at various stages[14].

    This paper is a survey of prediction of PD using Machine learning and Deep learning techniques that generated good models and potency of those algorithms in terms of accuracies achieved, also about different methodologies applied.

  2. LITERATURE SURVEY

      1. Importance of Voice data:

        Speech or voice data is assumed to be 90% helpful to diagnose a person for identifying presence of disease. In general, Person with PD suffer from speech problems, which can be categorized into two: hypophonia and dysarthria. Hypophonia indicates very soft and weak voice from a person and dysarthria indicate slow speech or voice , that can hardly be understood at one time and this causes because of damage to central nervous system. So, most of the clinicians who treat PD patients observe dysarthria and try to rehabilitate with specific treatments to improvise vocal intensity.

      2. Survey carried out for the diagnosis of PD with different algorithms and approaches:

        Several strategies are recorded for early detection of PD based on the different ML techniques. But accuracy in detection and classifying within the time is very important or else, it causes development of more symptoms. There are different kinds of data, brain MRI images, Voice data, posture images, senor captured data, handwritten data, using which we can predict whether person is having PD or not. Out of all those , speech or voice data helps in identifying PD accurately.

        Eduardo Tolosa et al proposed a twofold fully automatic approach with 3D images has shown promising results in their experimentation [4].

        Max A. Little et al presented a new dysphonia measure, pitch period entropy (PPE) and used a kernel support vector machine and has achieved classification accuracy of 91%[10].

        RAINER SCHO¨ NWEILER et al identified a different approach which used voice analysis with ANN and got good results but observed that cost-effectiveness remains to be a challenge[5].

        Marius Ene et al suggested NN based approach with three types of internal methods and discriminated persons having PD with healthy persons[7].

        DAVID GIL A, MAGNUS JOHNSON B found that with a smaller number of neurons at hidden layer both training set and test sets performed poorly. With higher number of neurons, the training set performed well with high risk of

        over fitting. The ideal solution for this layer was found to be 13 neurons[8].

        Ipsita Bhattacharya et al identified the ROC curve variation and identified that values of TP and FP rates show changes while increase in the CV folds[13].

        Freddie Åström et al proposed unique approach of parallel neural networks and then outcome of each neural network is assessed by using a rule-based system for the decision. During the training process, data that is not yet learned of each neural network is collected and applied in the training set of the later neural network. This helped to increase prediction accuracy[14].

        Athanasios Tsanas et al developed a novel algorithm based on speech signals but its questionable as most of the features are not considered here, only 10 features are used[15].

        Hui-Ling Chen et al proposed FKNN centered system using a 10-fold cross validation method[17].

        Mohammad S Islam et al has compared various ML techniques based on their performance accuracies in determining whether person is having PD or not and mentioned that new classifier may be built to get better accuracies[18].

        Bo Penga et al suggested Computer Aided Analysis with image data and used BrainLab software for processing the images and calculate thickness of the cortex, volume of gray matter, and surface area of the cortex on each region of interest (ROI). Use of Multilevel ROI-based features improved the classification performance[19].

        Derya Avci and Akif Dogantekin proposed another approach using Genetic Algorithm-Wavelet Kernel-Extreme Learning and achieved good accuracy results[22].

        R Prashanth identified that multimodal features can be used to predict PD in earlier stage[23].

        Satyabrata Aich proposed a unique approach by using Genetic algorithm and PCA as feature selection methods and applied seven ML algorithms for classification, that saved time and productivity while doing pattern classification with two categories such as PD and not PD[25].

        Leandro A. Passos compared ResNet-50 , Optimum-Path Forest (OPF) classier with Support Vector Machines (SVM) and Bayes and achieved 96% of identication rate[37].

        Deepak Gupta followed a different approach cuttlefish algorithm and used for feature selection,

        different fitness functions approximations are used to improve cuttlefish algorithm and is termed as Optimized cuttlefish algorithm (OCFA).Decision tree and K-Nearest Neighbor classifiers are applied and achieved 94% of accuracy in detecting PD effected patients [36].

        Salama A. Mostafa proposed

        1. Multiple Feature Evaluation Approach (MFEA) of a multi-agent system (ii) Implementation of five clasification schemas which are Decision Tree, Random Forests, Neural Network, Naïve Bayes and Support Vector Machine on the Parkinsons diagnosis before and after applying their approach, and (iii)Author approach witnessed the following average rate of accuracies : Decision Tree achieved accuracy of 10.51%, Naïve Bayes shown 15.22%, Neural Network is found with 9.19%, Random Forests and SVM performed with 12.75% and 9.13% respectively.[34]

    S.No

    Author Name

    Year

    Methodology

    Input data

    Performances

    1

    Ali H. Al-Fatlawi et al

    2016

    Deep belief network, Restricted Boltzmann Machines ,Back propagation algorithm

    Voice data

    Acc: 94%

    2

    Marius Ene et al

    2008

    Probabilistic neural network (PNN)

    Speech samples

    Accuracies ranging between 79% and 81%

    3

    David Gil A, Magnus Johnson B

    2009

    ANNs and SVMs

    Speech

    90%

    4

    Chien-Wen Cho et al

    2009

    Principal component analysis with linear discriminant analysis.

    Voice samples

    95.49%

    5

    Max A. Little et al

    2009

    SVM

    Voice recordings

    classication performance of 91.4%

    6

    Resul Das et al

    2010

    Neural Networks, DMneural, Decision Tree and Regression

    Speech

    Score of 92.9% is achieved

    7

    C. Okan Sakar & Olcay Kursun

    2010

    SVM

    Speech data

    classification accuracy:92.75%

    8

    Zachary C.Lipton et al

    2016

    Long Short-Term Memory (LSTM-RNN) with forget gate, MLP

    Voice data

    Several accuracies are compared.

    9

    Ipsita Bhattacharya et al

    2010

    Used LibSVM for classifying along with random split of the dataset, and determine accuracy for the different kernel functions

    speech

    Improved average accuracy achieved.

    10

    Freddie Åström et al

    2011

    Used a different neural network to minimize the probability of outcome with error

    Voice data

    Total nine parallel neural networks are arranged and achieved development of 8.4%

    for the prediction of PD compared to single network

    11

    Athanasios Tsanas et al

    2012

    Speech signal processing algorithms, RF,SVM

    Voice signals

    99%

    12

    Indrajit Mandal et al

    2017

    Multinomial logistic regression, rotation forest together with SVM and PCA, ANN, boosting methods

    Speech

    100% accuracy achieved with sparse multinomial logistic regression and linear logistic regression, observed sensitivity:0.983 and specificity: 0.996

    13

    Hui-Ling Chen et al

    2013

    FKNN,SVM

    Speech

    96.07% obtained by the FKNN dependent system using a 10- fold CV method

    14

    Tarigoppula V.S Sriram et al

    2013

    SVM,KNN,NB,RF

    Voice data

    Random Forest shown better accuracy

    15

    Mohammad S Islam et al,2014

    2014

    SVM, Random Tree and Feedforward Back- propagation built Artificial Neural Network.

    Speech

    90% recognition accuracy

    16

    Oana Geman et al

    2015

    SVM,DNN

    Voice data

    90% accuracy achieved

    17

    Bo Penga et al

    2015

    t-test, SVM, and Minimum Redundancy and Maximum Relevance.

    Speech impairment data

    Proposed method used multilevel ROI-based features and is observed better classification accuracy..

    18

    Othman Ibrahim , Mehrbakhsh Nilashi, & Ali Ahani

    2016

    PCA is used for feature selection, EM, ANFIS and Support Vector Regression (SVR).

    Voice data

    SVM:AUC-0.9623 ANFIS:AUC-0.848

    19

    Hui-Ling Chen et al

    2016

    Extreme learning machine and kernel ELM

    Speech samples

    10- fold cross validation through 10 runs achieved 96.47% accuracy

    20

    Derya Avci and Akif Dogantekin et al

    2016

    Genetic Algorithm, wavelet kernel and Extreme Learning Machines(ELM).

    Voice data

    96.81%.

    21

    Thomas J. Hirschauer

    2015

    EPNN (Enhanced Probabilistic Neural Network

    Speech

    98.6 %

    22

    Lígia Sousa et al

    2019

    DNN, KNN,PCA (for optimizing feature set)

    Voice samples

    93.4% for the binary classication,84.7% for multiclass classication.

    23

    Leandro A. Passos

    2018

    ResNet-50 , Optimum-Path Forest (OPF) classier

    HandPD dataset, speech

    96% of identication rate using speech samples.

    24

    Deepak Gupta

    2018

    Optimized cuttlefish algorithm ,Decision tree, KNN

    Speech data and Handwritten data are used to

    evaluate the proposed model.

    94%

    25

    Shreya Bhat

    2018

    Along with advanced machine learning methods, Neuroimaging modalities also used

    Image data,

    speech, ,MRI, EEG

    (Various implementations are discussed)

    26

    Hariharan et al

    2014

    Gaussian mixture with PCA and LDA. SVM classifier

    Speech data

    100%

    27

    Zhang et al

    2017

    Stacked autoencoders, KNN

    Speech

    In the range of 94-98%

    28

    Oung et al

    2018

    Classifiers used are KNN, PNN, ELM classifiers.

    Motion and Speech

    KNN:93.26% PNN: 95.22%

    ELM: 95.93%

    29

    Hlavnicka et al

    2017

    Zero-crossing rate, variance of autocorrelation function.

    Speech

    Accuracy: 71.30%

    Sensitivity: 56.70%

    Specificity: 80%

    30

    Salama A. Mostafa

    2018

    Decision Tree, Random Forests, Naïve Bayes, Support Vector Machine and Neural Network.

    Voice data

    Avg rate of improved accuracies achieved are: Decision Tree: 10.51%,

    Random Forests: 12.75%

    Naïve Bayes:15.22%, Support Vector Machine: 9.13%, Neural Network: 9.19%

    31

    Rainer schonweiler et al

    2000

    Artificial neural networks, Regression tree

    Voice data

    Varius combinations of methods applied and achieves improved accuracies.

    Table 1: Summary of the survey of various methodologies and their performances

    It is important to note that , out of all ML techniques, ANN and SVM classifier are used most of the proposed algorithms to aid faster and accurate the prediction.

    As per the survey, we observed that most of the models used voice/speech data for efficient diagnosis of the disease and because it is preferred by most of the therapists to consider voice data as relevant feature.

  3. ARCHITECTURE OF ANN

    The following figure represents architecture of Artificial Neural network with an input layer, hidden layer(s) and an output layer. Number of hidden layers for each network varies from one another.

  4. DISCUSSIONS

    Machine Learning techniques has got prominent role as they are applied in variety of domains especially in the healthcare. Unlike traditional methods, the models generated by applying ML techniques show dynamic outputs as data is fed into it. One shall make note that significant and narrow research is needed to obtain knowledge in diagnosing the disease. Various machine learning algorithms and techniques are being proposed rapidly, out of which some are observed to be promising with the results and few demonstrated their usage in different fields. Advantage with the ML generated models is that when more data is used, the precision values gets increased and the much accuracy in predictions can be gained.

    Every circle in the above network represents a neuron at which the inputs and corresponding weights are processed layer by layer.

    Input Layer:

    An input layer accepts large volumes of data as input to build the neural network. The data can be in the form of text, image, audio, etc. In general, the input layer contains features of the dataset, each node of input layer in the above architecture represents one feature

    Hidden Layer:

    Every hidden layer receives the input feature along with their weights, where weights of every feature indicates their contribution towards the decision or prediction. Hidden layer processes the data at each node by performing complex computations and helps in feature extraction. Nodes at first hidden layer receives product of input feature with its weights value and is passed as input

    to next hidden layer and so on. Choosing number of hidden layers and number of nodes for every hidden layer varies with the problem as well as dataset.

    Output Layer:

    At output layer Processing of nodes are determined by the functions called as activation functions like tanh, sigmoid, ReLU. Depending on the kind of dataset and criteria , one can decide suitable activation function. Output layer receives the output generated by last hidden layer as input and generates an output in the desired form.

  5. CONCLUSION

    This paper is an effort to present broad review about Parkinson disease diagnosis system that have applied various machine learning techniques. The summary of results obtained by different researchers is made available in literature survey table , almost all the authors/researchers made great efforts to predict the Parkinson disease with novel approaches. It can be identified that maximum of all ML techniques used by various authors worked better but developing a very faster classifier using novel architecture of neural network combined with specific approach may work better. To achieve this, we try to implement artificial neural network with different number of hidden layers and number of nodes in future and compare all the accuracies.

  6. REFERENCES

  1. Claas Ahlrichs ae al, Parkinson's Disease Motor Symptoms In Machine Learning: A Review, Health Informatics- An International Journal (HIIJ) Vol.2,No.4,November 2013.

  2. ERIKA ROVINI et al, Comparative Motor Pre-clinical Assessment in Parkinsons Disease Using Supervised Machine Learning Approaches, Annals of Biomedical Engineering (2018) https://doi.org/10.1007/s10439-018-2104-9.

  3. Pratibha Surathi et al Research in Parkinson's disease in India: A review, Ann Indian Acad Neurol. 2016 Jan-Mar; 19(1): 9 20.doi: 10.4103/0972-2327.167713

  4. Eduardo Tolosa et al, Decision Support System for the Diagnosis of Parkinsons Disease, SCIA 2005, LNCS 3540, pp. 740749, 2005.

  5. Rainer Scho¨ Nweiler et al, Novel Approach to Acoustical Voice Analysis Using Artificial Neural Networks ,JARO 01: 270282 (2000) DOI: 10.1007/s101620010020.

  6. Gert Cauwenberghs et al, Incremental and Decremental Support Vector Machine Learning, Proceedings of NIPS 2000,Pages 384-394.

  7. Marius Ene et al, Neural network-based approach to discriminate healthy people from those with Parkinsons disease, Annals of the University of Craiova, Math. Comp. Sci. Ser. Volume 35, 2008, Pages 112116 ISSN: 1223-6934

  8. DAVID GIL A, MAGNUS JOHNSON B, Diagnosing Parkinson by using Artificial Neural Networks and Support Vector Machines, Global Journal of Computer Science and Technology, 2009.

  9. Chien-Wen Cho et al, A vision-based analysis system for gait recognition in patients with Parkinsons disease, Expert Systems with Applications 36 (2009) 70337039

  10. Max A. Little et al, Suitability of Dysphonia Measurements for Telemonitoring of Parkinsons Disease, IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 56, NO. 4, APRIL 2009.

  11. Resul Das et al, A comparison of multiple classification methods for diagnosis of Parkinson disease, IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 13, NO. 2, MARCH 2002.

  12. C. Okan Sakar & Olcay Kursun, Telediagnosis of Parkinsons Disease Using Measurements of Dysphonia, J Med Syst (2010) 34:591599 DOI 10.1007/s10916-009-9272-y.

  13. Ipsita Bhattacharya et al, SVM Classification to Distinguish Parkinson Disease Patients, Conference: Proceedings of the 1st Amrita ACM-W Celebration on Women in Computing in India, September 16-17, 2010, Tamilnadu, India.

  14. Freddie Åström et al, A parallel neural network approach to prediction of Parkinsons Disease, 2011 Elsevier, doi:10.1016/j.eswa.2011.04.028

  15. Athanasios Tsanas et al, Novel Speech Signal Processing Algorithms for High-Accuracy Classification of Parkinsons Disease, IEEE Transactions On Biomedical Engineering, Vol. 59, No. 5, May 2012.

  16. Indrajit Mandal et al, New machine-learning algorithms for prediction of Parkinson's disease, International Journal of Systems Science, DOI:10.1080/00207721.2012.724114.

  17. Hui-Ling Chen et al, An efficient diagnosis system for detection of Parkinsons disease using fuzzy k-nearest neighbor approach, Expert Systems with Applications 40 (2013) 263271.

  18. Mohammad S Islam et al, Performance Comparison of Heterogeneous Classifiers for Detection of Parkinsons Disease Using Voice Disorder (Dysphonia), 3rd International Conference On Informatics, Electronics & Vision 2014.

  19. Bo Penga et al, Computer Aided Analysis of Cognitive Disorder in Patients with Parkinsonism using Machine Learning Method with Multilevel ROI-based Features, J Med Syst (2015) 39:179 DOI 10.1007/s10916-015-0353-9.

  20. Mehrbakhsh Nilashi, Othman Ibrahim & Ali Ahani, Accuracy Improvement for Predicting Parkinsons Disease Progression, Scientific Reports,2016 | 6:34181 | DOI: 10.1038/srep34181.

  21. Hui-Ling Chen et al, An efficient hybrid kernel extreme learning machine approach for early diagnosis of Parkinsons disease, Neuro Computing (Elsevier) Volume 184, 5 April 2016, Pages 131-144.

  22. Derya Avci and Akif Dogantekin et al, An Expert Diagnosis System for Parkinson Disease Based on Genetic Algorithm-Wavelet Kernel- Extreme Learning Machine, Hindawi Publishing Corporation Parkinsons Disease Volume 2016, Article ID 5264743, 9 pages http://dx.doi.org/10.1155/2016/526474.

  23. R Prashanth, High-Accuracy Detection of Erly Parkinson's Disease through Multimodal Features and Machine Learning, ,Volume 90, June 2016, Pages 13-21

  24. Freddie Astrom, A parallel neural network approach to prediction of Parkinsons Disease, Expert Systems with Applications, Volume 38,

    Issue 10, 15 September 2011, Pages 12470-12474

  25. Satyabrata Aich, A Supervised Machine Learning Approach using Different Feature Selection Techniques on Voice Datasets for Prediction of Parkinsons Disease, ICACT Transactions on Advanced Communications Technology (TACT) Vol. 7, Issue 3, May 2018.

  26. O. Faust, Y. Hagiwara, J. H. Tan, S. L. Oh and U. R. Acharya, "Deep learning for healthcare applications based on physiological signals: a review," Computer Methods and Programs in Biomedicine, vol. 161, pp. 1-13, 2018.

  27. M. Hariharan, K. Polat and R. Sindhu, "A new hybrid intelligent system for accurate detection of Parkinson's disease," Computer Methods and Programs in Medicine, vol. 113, pp. 904-913, 2014

  28. Q. W. Oung, M. H, S. N. Basah, H. Lee and V. Vijean, "Empirical wavelet transform based features for classification of Parkinson's disease severity," Journal of Medical Systems, vol. 42, p. 29 , 2017

  29. Y. Zeinalia and B. Story, "Competitive probabilistic neural network," Integrated Computer-Aided Engineering, vol. 24, no. 2, pp. 105-118, 2017

  30. T. Hirschauer, H. Adeli and T. Buford, "Computer-aided diagnosis of Parkinson's disease using an enhanced probabilistic neural network," Journal of Medical Systems, vol. 39, no. 179, p. (12 pages), 2015.

  31. M. Ahmadlou and H. Adeli, "Enhanced probabilistic neural network with local decision circles: a robust classifier," Integrated Computer- Aided Engineering, vol. 17, no. 3, pp. 197-210, 2010.

  32. M. Abrahams, "Diagnostic markers in the early detection of Parkinson's disease," 2012.

  33. F. L. Pagan, "Improving outcomes through early diagnosis of Parkinson's disease," American Journal of Managed Care, vol. 18, no. 7, pp. 176-182, 2012.

  34. S. A. Mostafa, A. Mustapha, M. A. Mohammed et al., Examining multiple feature evaluation and classication methods for improving the diagnosis of Parkinsons disease, Cognitive Systems Research , 2018.

  35. National Parkinson Foundation (a), "What is Parkinsons?" [Online]. Available: http://www.parkinson.org/understanding-parkinsons/what- is-parkinsons. [Accessed 08 03 2017].

  36. Deepak Gupta, Optimized cuttlefish algorithm for diagnosis of Parkinsons disease, Cognitive Systems Research, Volume 52, December 2018, Pages 36- 48https://doi.org/10.1016/j.cogsys.2018.06.006

  37. Leandro A. Passos, Parkinson Disease Identication using Residual Networks and Optimum-Path Forest, SACI 2018, IEEE 12th International Symposium on Applied Computational Intelligence and Informatics, May 17-19, Timiúoara, Romania

Leave a Reply

Your email address will not be published. Required fields are marked *