Differentiating Early and Late Stage Parkinson’s Disease Patients from Healthy Controls

Parkinson’s disease (PD) is a neurodegenerative disease which is difficult to diagnose at an early stage. Brain imaging techniques like [18F]-fluorodeoxyglucose positron emission tomography (FDG-PET) may
aid to identify disease-related changes in cerebral glucose metabolism. The scaled subprofile model with principal component analysis (SSM/PCA) is applied to FDG-PET data to extract features and corresponding patterns of glucose metabolism which can be used to distinguish PD subjects from healthy controls. From a previous study, the decision tree (DT) classifier’s performance to separate the PD group from healthy controls was below chance level. This could be attributed to the small number of subjects in the dataset, combined with the early disease progression. In this study, we make use of an
additional PD dataset, consisting of subject brain images obtained at a later disease stage. The features m extracted by the SSM/PCA method are used for distinguishing PD subjects from healthy controls using three classification methods, that is, decision trees, Generalized Matrix Learning Vector Quantization (GMLVQ), and Support Vector Machine (SVM) with linear kernel. The classifiers are validated to determine their capability of classification given new subject data. We compare the classifiers performances on the distinct early-stage and late-stage datasets, as well on the combined datasets. We
also use the early and late-stage datasets interchangeably for training and testing the classifiers. We find that the DT classification performance on the late-stage dataset is considerably better than in the previous study, where we used early-stage data. For early-stage patients, the application of the GMLVQ
and SVM classifiers gives a significant improvement as compared to the DT classifier.