Alot of research has been done to predict Parkinson’s disease in a person,however less work has been done to predict its severity in patients. Theseworks have used various machine learning techniques like random forests,support vector machines, decision trees, dynamic neural network, and Hidden Markov Models. For effectual diagnosis of the Parkinson’s Disease (PD), several types ofclassification techniques were employed by Das2. The performance score of theclassifiers was computed based upon various evaluation methods. As per theresults of application scores, they found that Neural Networks (NNs) classifierobtains the best result, giving 92.9% accuracy.
Genain et al.3 have used Bagged decision trees (random forests) to predict continuous measures of Parkinson’sDisease Severity from voice recordings of patients. They have used theUniversity of California Irvine Parkinsons Telemonitoring Dataset and datasetfrom Synapse.org and achieved model accuracy to an RMSE of 2.
Akash P. Manwatkar et al. 4 have created a classifier that extracts featuresfrom voice recordings of Parkinson’s disease patients and detects the severitylevel of the disease. The dataset used was provided by Synapse.org, containingvoice recording of patients along with their demographic information as well astheir PDRS and Hoehn & Yahr scale score.
Several voice filtering algorithmssuch as cepstral mean normalization, band path filter, and spectral subtractionwere applied to pre-process the data and then combinations of machine learningalgorithms such as support vector machine and decision tree classified eachrecording to the severity. The obtained result had an accuracy of 62.2% inpredicting Hoehn Yahr score and RMS error of 9.57 in predicting PDRS score.
Inthe research work, ‘Multiclass classification of Parkinson’s disease usingdifferent classifiers and LLBFS(Local Learning Based Feature Selection) featureselection algorithm’, done by Ben Malek et al.5, a 40-features dataset hasbeen used, where after analyzing all the features, 9 were selected out of 40 toclassify PWP subjects into four classes, based on their unified Parkinson’s diseaseRating Scale (UPDRS). The various techniquesused for this work were Subspace discriminant, KNN and SVM which provided anaccuracy of 96.5%, 67.2% and 90.1% respectively.