INTRODUCTION:
Parkinson’s disease is a neuro-degenerative disorder which affects quality of life of an estimated 10 million people worldwide3. A tell-tale marker of this disease is a decrease in the dopamine levels in the brain which could be attributed to the degeneration of dopaminergic neurons. The onset of the disease may be suggested by tremor, rigidity, slowness of movement and postural instability14. Such symptoms may not present itself in the same format in all the cases but rather vary in combinations and severity but generally is chronic and degenerative. What interests us here is that among all diagnosed cases of PD, 90% of the cases show some sort of vocal impairment which consists of deterioration of normal production of vocal sounds, which is medically termed as Dysphonia2.
The impact of this disease stands at a whopping 1-2% of people worldwide in the age range of 60 years and above14.
The challenge currently faced by Medical Science is the early detection of PD24in the affected patients. If diagnosed early, the patients can improve their quality of life even if the disease progresses15. However this is difficult as PD symptoms have overlap with the symptoms of other diseases and hence PD might go undetected, or worse, diagnosed inaccurately. Another complication is that traditionally, the diagnosis of PD18 comprises of various procedures like taking an extensive neurological history of the patient and observing the patient’s motor skills under different circumstances. I would also like to point out the sheer absence of a definitive laboratory test for the detection of PD27, hence making diagnosis nearly impossible at the earlier stages of the disease, where the motor symptoms aren’t severe enough for detection. Therefore, monitoring progress of the patient and effective diagnosis requires persistent visits by the patient.
Any way to diagnose PD25 without continual clinic visits would be more than appreciated by patients and would prove to be beneficial for them. As PD patients show unique and characteristic vocal features, their voice recordings would prove to be an invaluable procedure for diagnosis as it is also non-invasive. Hence tests using speech prove to be a good tool to audit PD. This opens up an entirely new door of opportunities for Machine Learning algorithms to work on in-house voice recording datasets of PD patients and then go on to efficiently predict or diagnose PD20. Voice features from the audio clips can be extracted by passing the recordings through signal processing algorithms and a these features can be used to build a classification and regression model to predict a rating on the Unified Parkinson’s Disease Rating Scale (UPDRS)15. UPDRS, which indicates presence and progression of PD, has been as the most commonly used scale for measuring and assessing PD.
This work is organised as follows the next section entails the literature survey conducted on the existing research work on the domain. The third section talks about the methodologies adopted and gives an outline of the dataset we’ve worked with. The fourth Section talks about the implementation part and elaborates on all the Machine Learning algorithms used in the making of this project. The subsequent section discusses the results obtained and the paper then winds up with the conclusion.
EXISTING LITERATURE:
The following section entails the literature survey that was conducted for the purpose of better understanding the problem at hand and to explore possible solutions.
Resul Das compares four different types of classification algorithms in the machine learning domain for enabling the diagnosis of Parkinson’s disease. In the study a SAS based software has been used to model various classifiers which can recognise the presence of PD. The classifiers employed are – DMNeural, Neural Network, Regression and Decision Tree. Various evaluation methods were run to evaluate the efficiency of the classifiers and the accuracy of NN yielded 92.9% correct classification rate.
The work by Cheng et.al aims at presenting an efficient diagnosis system but using Fuzzy K- Nearest Neighbour algorithm for doing so2 .Initially Principal Component Analysis is applied to construct the optimal feature set and the FKNN model is built on it. The model is then compared with Support Vector machine based approaches. The conclusion was that the FKNN based systems outperform the SVM based systems.
The paper by Hariharan et.al is different from the rest as in it proposed a hybrid intelligent system3 for detection and subsequent diagnosis of PD. The pre-processing for the proposed system is done using model based clustering (Gaussian Mixture Model), feature reduction is performed using Principal Component Analysis, Linear Discriminant Analysis, Sequential Forward Selection and Sequential Backward Selection. Classification was done using three supervised classifiers such as least square support vector machine, Probabilistic Neural Network and General Regression Neural Network. The dataset used was from the UCI ML5 data base.
The work by Astrom and Koker makes use of a parallel feed-forward neural network structure4 for the prediction of PD. This was because a set of nine parallel neural networks simply yielded an 8.4% improvement when in comparison with a single neural network. The output of each Neural Network is evaluated by using a rule based system for the final output. The unlearned data for each NN is collected and carry forwarded to the training set of the next NN.
The work by Challa K and others works to predict PD using non-motor symptoms such as Rapid Eye Movement, Sleep Disorder and Olfactory Loss5. Another paper by Agarwal K et al aims to predict Parkinson’s by using extreme machine learning techniques called as Extreme Learning Method6. The simple architecture of the proposed method and the inbuilt learning of Data makes the prediction scheme a very reliable one. One of the major drawbacks of all the above research work is tediousness of conducting tests for PD Patients. The Paper by S.Arora et al7 tackles exactly that domain by enabling smartphones to collect data on PD Patients and using that data for prediction.
The work by Asgari and Shafran works not only on speech data, but refines speech data from PD Patients. The work aims to find a versatile sample of voice production, use processing algorithms to extract useful information8 and then work on that. The work by Sriram et al creates a tool called “ParkDiag”9 and it reviews characteristics in voices as an identification for PD diagnosed patients. The paper by Jangwon Kim et al10, describes a fully automated method to predict the disease severity on the UPDRS scale. The work by Lewis.S et al investigates the heterogeneity in PD patients11 across age groups and symptoms shown.
The work by Little et al lays foundation to the validity and importance of dysphonia as a symptom to detect or predict PD12.This helps us pick dysphonia as the guiding metric in our proposed method. The paper by Pablo Martínez-Martínet al.13 aims to make a cut-off between different severity levels of PD Patients by making use of an unconventional psychiatric metric. The paper by Nilashi, M., Ibrahim, O., and Ahani, A presents an accuracy improvement for PD Prediction by using a unique combination of noise removal, clustering and prediction methods14. The paper byTsanas, A., Little, M. A., McSharry, P. E., and Ramig, L. O presents an accurate and effective remote telemonitoring of symptoms of PD patients by non-invasive speech tests.15
The proposed work builds upon all this existing literature and aims to predict Parkinson’s disease with acceptable accuracy.
METHODOLOGY:
Architecture:
The dataset is taken from the UCI ML repository. Exploratory data analysis is done long with visualizations of attributes to understand how the data characteristics of the data. Dimensionality of the data is then reduced using Principle component analysis. The data is then split into a training dataset and test dataset. The ML algorithms and the neural network algorithm is fit on to the training dataset. The trained model is then used to predict the values from the test dataset and the accuracy is calculated. Figure 1 represents the proposed model for classification,
Fig.1: Architecture of the system
Dataset:
The project aims to build a machine learning model which will predict the severity of PD using Parkinson’s Telemonitoring dataset from UCI ML repository.
The dataset used in this study is collection of features from voice recordings taken from 42 candidates who have been diagnosed with early-stage Parkinson's disease recruited to a six-month trial of a telemonitoring device for remote symptom progression monitoring. These recordings were captured by the device at the homes of the patients. Each row in the dataset corresponds to one of 5,875 voice recording from the aforementioned individuals.
The objective of this study is to predict the motor and total UPDRS scores ('motor_UPDRS' and 'total_UPDRS') from the 16 vocal attributes which will give an idea about the progression of PD.
The attributes in the dataset:
(i) Subject number:
Unique Id of each individual
(ii) Subject Age:
Age of the patient
(iii) Subject gender:
‘0’ indicates male, ‘1’ indicates female
(iv) Test time:
Time since the individual has been recruited into the trial.
(v) UPDRS:
This is a clinician’s scale for recording symptoms related to Parkinson’s disease. The UPDRS metric is a scale with 44 sections, and each of those sections addresses a symptom at one different part of the body. A summation of these 44 section will yield a value called UPDRS score which ranges between 0 (perfectly healthy) to 176(total disability)
(vi) Motor UPDRS:
Motor UPDRS score given by the clinician, linearly interpolated - this forms sections 18-44 from the UPDRS sections
(vii) Total_UPDRS:
total UPDRS score given by clinician, linearly interpolated - this includes all 44 sections.
(viii) 16 other biomedical voice measures such as :
Jitter Percentage, Jitter (Absolute), Jitter (RAP), Jitter (PPQ5), Jitter (DDP), Shimmer, Shimmer(dB), Shimmer:APQ3, Shimmer:APQ5, Shimmer:APQ11, Shimmer: DDA, NHR, HNR, RPDE, DFA, PPE
Prediction:
The objective of this study is to accurately predict the motor and total UPDRS scores ('motor_UPDRS' and 'total_UPDRS') from the 16 voice measures using various machine learning methods and compare the results. For the same purpose we will be making use of SVM, Decision Trees, Linear Regression and Neural Networks. We’ve used Root Mean Squared Error (RMSE). RMSE is defined as the square root of the average of the square of the total error. It’s one of the most common uses measure of accuracy for prediction problems.
RMSE= (1)
Where,
n=number or observations
Xi = Each observation
X=Mean of observations
MACHINE LEARNING MODELS:
Support vector regression (SVR):
Support Vector Machine(SVM) or Support Vector Regression (SVR) is a supervised machine learning algorithm. It can be used for Regression or Classification. In this algorithm , each data item is plotted on a graph as a point in an n-dimensional space where the value of each feature is converted as a value of a particular corresponding coordinate. Then, a hyper-plane that differentiates the two classes is found, hence classifying the points. In SVM regression, the data item yis plotted on a graph in an n-dimensional space using a fixed mapping, and then a linear model is constructed in this feature space. The linear model f(y,w )is given by
f(y,w)=
(2)
where
hi(y), i=1,..,n denotes a set of nonlinear transformations, and c is the “bias” term. The bias term is often dropped when the data has zero mean.
Decision tree Regression:
Decision Tree is a machine learning technique which is used for classification problems and regression problems. This algorithm tries to solve problems by using a tree representation where each leaf node corresponds to a class label and each internal node corresponds to a feature. The best feature is placed in the root node and then the training set is split into subsets in such a way that each subset contains the same value as the feature. This process is repeated until all leaf nodes are found.
To predict a class label, we first start from at the root, and we follow along the branches of the tree by comparing the value at the root and the value at the attribute and then jump to the next node. The value of the attribute is compared with the values at other internal nodes until a leaf node is reached, which has the class label.
Linear Regression:
Linear regression is a machine learning model used to model a dependent variable based on one or more independent variable. If there, is only one independent variable, then it is simple linear regression and if there is more than one independent variable, then it is called multiple liner regression. Usually ordinary Least squares approach is used to model linear regression, but other methods like least absolute deviation, ridge and lasso regression are also used.
In a dataset of n tuples and m attributes, the relationship between the dependent variable Q against the independent variables Pi,i=1,....,m,is given by,
Qi = α0 + α1P1 + αmPm i = 1, , n (3)
Resilient Back propogation:
Resilient Back propogation(RPROP) is a supervised feed forward artificial neural network created by Martin Riedmiller and Heinrich Braun in 1992.This is a First order optimization algorithm which considers only the sign of the partial derivative of all patterns, and acts independently on each "weight". If the sign of the partial derivative of the total error function has changed compared to the previous iteration, the update value for that weight is multiplied by a factor η−, where η− < 1. If there is no change in the sign, the update value is multiplied by a factor of η+, where η+ > 1. After calculating all the update values, each weight is changed by its own update value in the direction opposite to that of its partial derivative, so that total error function is minimized. This algorithm has one of the fastest weight updating mechanism.
RESULTS AND DISCUSSION:
For the error calculation of the predicted scores, We have used Root Mean Squared Error (RMSE). RMSE is defined as the square root of the average of the square of the total error. It’s one of the most common uses measure of accuracy for prediction problems.
The Machine Learning models used for the predictions were elicited in the previous section. This section will go on to discuss the results achieved. The RMSE values of total UPDRS and motor UPDRS were calculated for the predictions by use of linear regression, SVR, decision tree and resilient back propogation algorithm. While comparing total UPDRS and motor UPDRS, all the above mentioned algorithms tend to predict motor UPDRS much more accurately as the error is lower in the latter. When a comparison was conduction between the 4 chosen Machine Learning Algorithms Support Vector Regression demonstrated the best results while the Artificial Neural Network algorithm- Resilient Back Propogation proved to be second best. The error (RMSE) values mentioned in Table 1, is tolerable to a limited extent as the UPDRS scale ranges between 0-176.
Table 1: Results- UPDRS Scores
|
ML Algorithm |
Total UPDRS (RMSE) |
Motor UPDRS (RMSE) |
|
Linear Regression |
9.93 |
7.61 |
|
Support Vector Regression |
7.49 |
6.06 |
|
Decision Tree Regression |
9.96 |
7.5 |
|
Resilient backpropagation |
8.95 |
6.5 |
CONCLUSION:
This paper aimed to accurately predict the motor UPDRS and total UPDRS scores from the 16 voice features from the Parkinson’s Telemonitoring Dataset. Various Machine Learning models such as SVM, Decision Trees, Linear Regression and Resilient BP were trained on the dataset and their accuracy was measured. The ML algorithms were also compared and contrasted in light of the particular data. We were able to achieve desirable accuracy and predict the UPDRS scores in the expected way. The limitations of the current work would be that no matter how automated the process of Parkinson’s prediction becomes, there still will be a need for human intervention, intelligence and experience to make the diagnosis an accurate one. For future works, the dataset could be modelled on other more fitting Machine Learning models to improve accuracy of prediction
REFERENCES:
1. Das, R. (2010). A comparison of multiple classification methods for diagnosis of Parkinson disease. Expert Systems with Applications, 37(2), 1568-1572.
2. Chen, H. L., Huang, C. C., Yu, X. G., Xu, X., Sun, X., Wang, G., and Wang, S. J. (2013). An efficient diagnosis system for detection of Parkinson’s disease using fuzzy k-nearest neighbor approach. Expert systems with applications, 40(1), 263-271.
3. Hariharan, M., Polat, K., and Sindhu, R. (2014). A new hybrid intelligent system for accurate detection of Parkinson's disease. Computer methods and programs in biomedicine, 113(3), 904-913.
4. Ĺström, F., and Koker, R. (2011). A parallel neural network approach to prediction of Parkinson’s Disease. Expert systems with applications, 38(10), 12470-12474.
5. Challa, K. N. R., Pagolu, V. S., Panda, G., and Majhi, B. (2016, October). An improved approach for prediction of Parkinson's disease using machine learning techniques. In Signal Processing, Communication, Power and Embedded System (SCOPES), 2016 International Conference on (pp. 1446-1451). IEEE.
6. Agarwal, A., Chandrayan, S., and Sahu, S. S. (2016, March). Prediction of Parkinson's disease using speech signal with Extreme Learning Machine. In Electrical, Electronics, and Optimization Techniques (ICEEOT), International Conference on (pp. 3776-3779). IEEE.
7. Arora, S., Venkataraman, V., Zhan, A., Donohue, S., Biglan, K. M., Dorsey, E. R., and Little, M. A. (2015). Detecting and monitoring the symptoms of Parkinson's disease using smart phones: A pilot study. Parkinsonism & related disorders, 21(6), 650-653.
8. Asgari, M., and Shafran, I. (2010, August). Predicting severity of Parkinson's disease from speech. In Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE (pp. 5201-5204). IEEE.
9. Sriram, T. V., Rao, M. V., Narayana, G. S., and Kaladhar, D. S. V. G. K. (2016). ParkDiag: A Tool to Predict Parkinson Disease using Data Mining Techniques from Voice Data. International Journal of Engineering Trends and Technology (IJETT), 31(3), 136-140.
10. Kim, J., Nasir, M., Gupta, R., Segbroeck, M. V., Bone, D., Black, M. P., and Narayanan, S. S. (2015). Automatic estimation of Parkinson's disease severity from diverse speech tasks. In Sixteenth Annual Conference of the International Speech Communication Association.
11. Lewis, S. J. G., Foltynie, T., Blackwell, A. D., Robbins, T. W., Owen, A. M., & Barker, R. A. (2005). Heterogeneity of Parkinson’s disease in the early clinical stages using a data driven approach. Journal of Neurology, Neurosurgery and Psychiatry, 76(3), 343-348.
12. Little, M. A., McSharry, P. E., Hunter, E. J., Spielman, J., and Ramig, L. O. (2009). Suitability of dysphonia measurements for telemonitoring of Parkinson's disease. IEEE transactions on biomedical engineering, 56(4), 1015-1022.
13. Martínez-Martín, P., Rodríguez-Blázquez, C., Alvarez, M., Arakaki, T., Arillo, V. C., Chaná, P., and Serrano-Duenas, M. (2015). Parkinson's disease severity levels and MDS-Unified Parkinson's Disease Rating Scale. Parkinsonism and related disorders, 21(1), 50-54.
14. Nilashi, M., Ibrahim, O., and Ahani, A. (2016). Accuracy improvement for predicting Parkinson’s disease progression. Scientific reports, 6, 34181.
15. Tsanas, A., Little, M. A., McSharry, P. E., and Ramig, L. O. (2010). Accurate telemonitoring of Parkinson's disease progression by noninvasive speech tests. IEEE transactions on Biomedical Engineering, 57(4), 884-893.
16. Dua, D. and Karra Taniskidou, E. (2017). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science.
17. J Bagyalakshmi, S Vijayaraj, Sindhu, TK Ravi. Method Development and Validation of Erythrosine (E127) Using RP-HPLC Coupled With PDA Detector. Asian J. Research Chem. 1(2): Oct.-Dec. 2008; Page 95-96.
18. Atul H Kategaonkar, Dhaval M Patel, Bhushan P Itkar, Vishnu P Choudhari, Bhanudas S Kuchekar, Ana G Nikalje. Spectrophotometric and Stability Indicating RP-HPLC-PDA method for Simultaneous Determination of Finasteride and Tamsulosin in Combined Tablet Dosage form. Asian J. Research Chem. 2(4):Oct.-Dec. 2009 page 414-419.
19. TentuNageswara Rao, Muralidhar Reddy Avuthu, B. Venkata Reddy, SNVS Murthy. Forced degradation study of TerbinafineHCl by HPLC with PDA detector. Asian J. Research Chem. 2016; 9(11): 561-565.
20. Shishir Shrivastava, Ajay Kumar Akasapu, Lokesh Sharma. Knowledge Discovery in Spatial Databases. Int. J. Tech. 1(1): Jan.-June. 2011; Page 23-28
21. Parth R. Shah, Pratik J. Gupta, Rushant S. Sabnis. Recent Advancementsin Smart Phone Charging. Int. J. Tech. 2017; 7(1): 69-78.
22. K. Shukla, S.S. Shukla, V. Jain, R. Pandey, S. Jain, Swarnlata Saraf, S. Saraf. Fingerprinting of Traditional Medicines through RAPD Technology: - A Newer Approach. Research J. Pharm. and Tech. 1(2): April-June. 2008; Page 63-68.
23. Susanta Kumar Rout, Saptarshi Dutta, Mukul Sengupta, Srimanta Das. Evaluation and Comparison Study of Acetylcholine and Histamine on Isolated Chick, Rat, Guinea pig and Rabbit Ileum with Special Reference to PD2 Value in Pharmacology Practical. Research J. Pharm. and Tech. 4(5): May 2011; Page 715-718.
24. Neveda Baskeran. Iron Deficiency and Brain Disorders. Research J. Pharm. and Tech. 7(3): Mar., 2014; Page 352-353.
25. P. Maheshwari, I. Somasundaram. Health Related Quality of Life Measurement in Asthma and Chronic Obstructive Pulmonary Disease. Research J. Pharm. and Tech. 2016; 9(5): 518-520.
26. Harikumar Rajaguru, Sunil Kumar Prabhakar. Wavelet Neural Networks, Elman Back propagation and Multilayer Perceptrons for Epilepsy Classification from EEG Signals. Research J. Pharm. and Tech 2018; 11(4):1301-1306.
27. Himadri Shekhaar Baul, Muniyan Rajiniraja. Favorable binding of Quercetin to α-Synuclein as potential target in Parkinson disease: An Insilico approach. Research J. Pharm. and Tech. 2018; 11(1): 203-206.
Received on 12.07.2018 Modified on 21.08.2018
Accepted on 16.09.2018 © RJPT All right reserved
Research J. Pharm. and Tech 2019; 12(2):644-648.
DOI: 10.5958/0974-360X.2019.00114.8