The detection of Parkinson disease using machine learning models

Since one of the most important symptoms of Parkinson's disease (PD) is dysphonia that is difficulty in speaking due to PD and a physical disorder of the mouth, tongue, throat, vocal cords, the diagnosis of disease is difficult to diagnose early with the widely used standard scales such as Unified Parkinson's Disease Rating Scale (UPDRS). Therefore, in this study, feature vectors were extracted from patients and healthy volunteers by using statistical parameters in the time, frequency and time-frequency domains of vocal cords, and these vectors were applied as the inputs into machine learning (ML) based classifier models for PD diagnosis. For this aim, the frequency domain coefficients and frequency sub-bands coefficients were obtained by using discrete Fourier transform (DFT) and discrete wavelet transform (DWT) from vocal cord voice signals, respectively. The feature vectors were extracted from these obtained coefficients by using statistical parameters such as mean, geometric mean, harmonic mean and standard deviation. These feature vectors were applied as the inputs into classifier models based on commonly used ML methods for PD diagnosis. In the classification experiments implemented in time, frequency and time-frequency domains for PD diagnosis, the classifier models reached to the satisfactory levels in terms of total correct classification (TCC) ratios. In addition, when the classification experiments performed by using the parameters of entropy, Chi- inter-points slope that were evaluated in this thesis for the first time were repeated for PD diagnosis, the classifier models provided much higher TCC ratios especially when the feature vectors obtained from the frequency domain of the signals were used. As a result, this approach allowed to diagnose PD from feature vectors obtained by using statistical parameters in the frequency sub-bands of patients' voice signals in high success ratios with the help of ML methods.