Development of a clinical decision support system for early diagnosis in Parkinson's patients without cognitive impairment

Today, Parkinson's disease (PD) has become an increasingly important health problem all over the world due to its frequency and the problems it creates. The disease is difficult to diagnose, and effective treatment and medication for PD have not been fully achieved to date. In this thesis study, it is aimed to develop an inexpensive, effective prediction model that can detect PD at an early stage in patients without cognitive impairment, using resting-state electroencephalography (EEG) data through an artificial intelligence approach. For this purpose, in this thesis study, ready-made EEG data sets belonging to three different universities were classified with different machine learning methods by using new feature approaches and features used for the first time were proposed, and the results were presented comparatively with the methods in the literature. First of all, using the features extracted from the frequency domain power spectral density of the data, the PH and control group data were successfully classified. In this process, new spectral features have also been proposed with successful results. Secondly, classification was made by converting EEG data into audio data and extracting musical features from it, a method that was tried for the first time in PD. The study results were evaluated separately in terms of the number of channels used, feature type and machine learning methods used and presented in comparison with the literature. According to the most successful results obtained, 96.67% was achieved with 10-fold cross-validation using the CatBoost algorithm in the classification made with a single channel and seven spectral features in the Iova data set. However, in the classification made with EEG subbands, 18 central channels and four spectral features at 4-8 Hz, and 10 central channels and 11 spectral features at 8-12 Hz separately, again with the CatBoost algorithm, a 100% result was achieved in all metrics. According to the chroma based musical features obtained from the audio data, 96.67% accuracy was again achieved with the CatBoost algorithm in 31 channels. It appears that the results obtained from this study are better than the results obtained so far with the same data sets in the literature. Therefore, the results of the this thesis are thought to be a promising and preferable option for PD analysis and early clinical diagnosis.