A comparison of classification methods for diagnosis of Parkinson's

Abstract

Parkinson's is a neurological health problem and one of the most common diseases affecting more than four million people worldwide. Recent studies have shown that deterioration of vocal cords, especially from Parkinson's, provides important information in the diagnosis and follow-up of the disease. In this study, a database of biomedical voice recordings from 32 people of different ages and genders was used to diagnose Parkinson's disease. With this database, the performance comparison of the machine learning algorithms k-Nearest Neighborhood (k-NN) and Naïve Bayes (NB) classifiers were performed. Seven different distance measurement methods (Chebychev, Correlation, Cosine, Euclidean, Hamming, Mahalanobis, and Spearman) for the k-NN and five different distribution methods (Uniform kernel, Epanechnikov kernel, Gaussian kernel, Triangular kernel and Normal distribution) for the NB classifier were performed in the performance process and separate tests were performed. The data obtained from these tests were compared with statistical measurements. In experimental studies, we used 10-fold cross validation technique for Parkinson dataset. Better results were obtained from k-NN classification algorithm than Naive Bayes classification algorithm. While k-NN mean accuracy score was 82.34%, this ratio was obtained as 74.15% for NB. Mahalanobis distance measurement method was found to give better results.