Parkinson’s Disease Detection Via Machine Learning Using Data Splitting and Validation Methods

Parkinson’s disease (PD), a neurological disorder, negatively affects the lives of patients and their caregivers. PD, which is very difficult to diagnose early by examining the clinical characteristics of the person, can be diagnosed using voice recordings. However, the inconsistent performance results of the models obtained from the evaluation of voice recordings through machine learning techniques limit the usability of these models to aid in diagnosing physicians. This study used a database of 195 voice data obtained from 31 individuals, 23 of whom have PD. The classification of the voices as healthy or patient was based on the 22 features in the database. The split ratios 90/10, 80/20, 70/30, 50/50 and 30/70 were used to select the training and test phase data, respectively. In addition, each split ratio was evaluated using 10-fold cross-validation, 5-fold cross-validation, holdout validation and resubstitution validation methods in the training phase, which is the initial process that will directly affect the other classification procedures. In addition, the classification process was performed using quadratic discriminant analysis, support vector machine, ensemble bagged tree, k-nearest neighbours and neural network classifiers. All procedures were repeated 10 times to ensure consistency of results and randomisation of split ratios. As a result, the k-nearest neighbours classifier with 80/20 splitting ratio and 10-fold cross-validation was determined to be the most successful among the compared methods with 95.64±3.21% accuracy. Therefore, it can be seen that much more successful results can be obtained by analysing only the effects of the existing parameters of the classifiers.