Research of Development, Application, and Validity of Various Machine Learning Algorithms to Predict The Disease Progression in Parkinson’s Disease

Parkinson’s disease (PD) is a progressive, neuro-degenerative disease mainly affecting the elderly and is considered the second leading neurodegenerative disease after Alzheimer’s disease. It is mainly characterized by motor and non-motor features affecting the patient’s movement, gait, balance, and swallowing. Current research methods lack a comprehensive understanding of PD disease progression as several clinical and non-clinical factors are involved in disease manifestation and progression which often lead to heterogenicity. PD has much more complex disease manifestations to predict than just the motor/non-motor symptoms. One of the major concerns to predict PD in the early stages is that PD symptoms overlap with the symptoms of other diseases such as Multiple Sclerosis and Alzheimer’s disease.

The Unified Parkinson’s Disease Rating Scale (UPDRS) is widely-used clinical symptoms rating scale for PD and serves as a baseline assessment tool to detect the presence and to rate the severity. Although current medical approaches can lessen the symptoms, there is no such standard therapy to cure PD. Therefore, early diagnosis is of critical importance to predict the disease progression to help patients survive and improve their quality of lives. Moreover, monitoring PD progression is often cumbersome as it involves a patient-directed care management plan where disease symptoms and UPDRS scores were determined in the clinic for every 3, 6 or 12 months. Since most of the patient affected were elderly and the methods are time-consuming, it is not logistically and financially feasible for both the physicians and the patients. In addition, most of the current methods used for tracking PD rely on expert clinical raters, from which PD symptoms assessment may be difficult due to inter-individual variability.

Therefore, for all these reasons, periodical remote monitoring of PD progression and UPDRS scores has emerged as an alternative solution to track the PD patients with low-cost non-invasive methods. Acknowledging the significance of remote patient monitoring in PD, healthcare providers have been developing and implementing several new approaches at the organizational levels to enhance the remote tracking of the disease progression in PD patients. Apart from the motor and non-motor symptoms, speech impairment is prevalent in 70-90% of PD subjects. Therefore, voice recordings would be useful to identify PD and to track the disease progression.

Further, PD subjects experience characteristic voice and speech patterns, which could serve as potential indicators for early detection. However, it is difficult for a physician to characterize these speech patterns and voice recordings manually for the same patient or several patients over the months. Machine learning (ML) methods programmed in wearable devices were proven to be a promising technology to automatically rate the disease severity scores using the UPDRS scale. These methods apply statistical or mathematical algorithms on input data to draw arbitrary patterns, common structures or data points in the dataset to make predictions for new input data (outcome). In this context, a systematic review of the available literature is warranted to understand the development, application, and validity of various machine learning algorithms to do perform a comparative analysis that helps clinicians to predict the disease progression in PD and designers in designing the wearable devices for remote monitoring of PD patients.

Although there is no specific test exists to diagnose PD, traditional diagnosis involves a neurologist examining the medical history of brain and evaluating the subject’s motor skills in various methods knowing that the traditional methods are prone to inter-individual variability. Another problem is that the early symptoms of PD often overlap with the symptoms of other diseases such as Alzheimer’s disease, multiple sclerosis, Huntington’s disease, and dementia with Lewy bodies leading to diagnostic errors. Due to lack of standard laboratory tests or methods to diagnose PD, early diagnosis of PD has become difficult where most of the motor symptoms are not severe in early stages of PD which require regular monitoring of motor symptoms in clinical settings.

PD often occurs among elderly adults, and a constant monitoring of disease progression is warranted through regular clinic visits. Remote patient monitoring has been gaining increased attention to track the disease progression using various noninvasive methods such as monitoring the speech patterns in subjects suffering from PD. Since speech impairment is prevalent in 70-90% of PD subjects, voice recordings would be useful to identify PD and to track the disease progression. Further, PD subjects experience characteristic voice and speech patterns, which can serve as potential indicators for early detection, low cost noninvasive and time-consuming diagnostic tools for PD. The reason behind the increased attention for diagnosis of PD using speech patterns is mainly due to rapid development in telediagnosis and telemonitoring in the medical field. Also, these methods are less expensive, and the devices are often easy to self-monitor by the subjects which reduce the patient visits to clinics allowing the patients to self-track their disease progression. Although medications and surgical interventions could control the PD disease progression by alleviating the motor symptoms, there is no method to cure PD. Current research methods lack a comprehensive understanding of PD disease progression as several clinical and non-clinical factors are involved in the disease manifestation and progression often leading to heterogenicity. Therefore, early diagnosis is of critical importance to predict the disease progression to help the patients improve and survive a quality of life.

The Unified Parkinson’s Disease Rating Scale (UPDRS) is the universal and widely-used clinical rating scale for assessing the clinical spectrum of PD and serves as a baseline assessment for PD. Understanding the relationship between UPDRS scores and patient’s speech signal features have been extensively studied for predicting PD in early stages. However, clinicians cannot use UPDRS scores to evaluate and rate the voice recordings of patients manually in a large dataset. Since voice recordings of patients usually occupy a large space, it is almost impossible for clinicians to manually evaluate them because it is a time-consuming process. Machine learning systems were proven to be a promising technology to automatically rate the disease severity scores using the UPDRS scale. Therefore, computer-aided systems using machine learning algorithms are developed to detect and monitor the disease progression objectively. Most of the studies used advanced machine learning algorithms to extract the relevant or most significant features (feature extraction) from the database that contributes to the PD (UPDRS scores). Voiced and unvoiced segments were extracted from the model built using Gaussian Mixture Model Universal Background Model (GMM-UBM) using Support Vector Regression algorithm. The model predicted the PD with Pearson’s correlation of 0. 60 for MDS-UPDRS scores. Remote tracking of PD disease progression was performed using regression methods such as Support Vector Machines (SVM), Least Square Support Vector Machines (LS-SVM), Multilayer Perceptron Neural Network (MLPNN) and General Regression Neural Network (GRNN) to predict the observed UPDRS scores. Results indicated that LS-SVM outperforms all other regression methods tested for the dataset. Minimum redundancy maximum relevance feature selection algorithm tested on the speech signals of PD resulted in 90. 3% accuracy and 90. 2% precision and Mathews correlation value of 0. 73 using random forest model. This finding showed that simple random forest was better than other methods such as bagging, boosting, SVM, and decision tree methods. To support the incremental updates of data, Incremental Support Vector Regression (ISVR) approach was implemented to predict the UPDRS scores. The study confirms that the self-organizing map (SOM), non-linear iterative partial least squares (NIPALS), and ISVR techniques are effective in predicting the Total-UPDRS and Motor-UPDRS (Nilashi et al. 2018). Although clustering was not the main objective of this study, advanced methods such as Principal Component Analysis (PCA) and Expectation Maximization (EM) were performed to cluster the multi-colinear PD speech data.

Novel regression techniques such as Adaptive Neuro-Fuzzy Inference System (ANFIS) and SVR were performed to predict the PD progression which remarkably improved the accuracy of PD prediction. However, this model was limited by a number of training samples in each cluster built by PCA and EM algorithms. UPDRS assessment for telemonitoring data was tested using linear and non-linear regression techniques such as least squares (LS), Iteratively Reweighted LS (IRLS), and Least Absolute Shrinkage and Selection Operator (LASSO). Computational approaches such as neural networks could develop more accurate models for disease prediction. Artificial Neural Networks (ANN) and ANFIS were used for predicting the PD using speech data. The advantages of using neural networks are that it is easy to train the models for mapping over the input data. However, it was noticed that ANFIS models comparatively taking more time for training the model. The best model implemented in this paper produced MAE = 5:33, MSE = 44:69 and R = 0:61. Data mining techniques using decision trees are performed on PD telemonitoring dataset from the University of California Irvine (UCI) to build a prediction model based on decision tree classification. The model produced an accuracy of 85. 08% with a classification error of 14. 92%. Bagged decision trees (random forests) outperformed linear regression, SVR, and Markov models over two different PD speech datasets (from Synapse. org and UCI machine learning library). The accuracy of prediction using random forests was improved to 2% on a scale of 0 to 176 with a root mean square error (RMSE) of 2. Novel and robust machine learning algorithms such as sparse multinomial logistic regression, rotation forest ensemble with support vector machines and principal components analysis, artificial neural networks, boosting methods were performed on the PD speech dataset. The multilinear logistic regression model produced an accuracy of 100% with sensitivity and specificity of 0. 983 and 0. 996 respectively.

Automatic classification of tremor activity was performed using decision trees, SVM, discriminant analysis, random forest, and K-nearest neighbor. The method produced a root mean square error (RMSE) of 0. 410 for the tested algorithms and the performance of algorithms has an accuracy of 85. 5%. The authors validated the results using a leave-one-out method rather than traditional cross-validation which has better precision and accuracy. Although the study concludes that the method was successful in automatic sorting of tremors using UPDRS scores and ML algorithms, there were few limitations discussed in the paper such as UPDRS bias towards low scores due to which lower precisions are recorded for a higher UPDRS score.

In another study, a new hybrid algorithm, genetic algorithm, and SVM were used to classify the PD patients based on their voice recordings. Genetic algorithm optimized the 14 key voice variables and SVM was used for classification of them. Although the study achieved an accuracy of 94. 5%, it resembled the results of previous work and was not tested for performance assessment of the new algorithm.