Disturbances in lipid composition and lipoproteins metabolism can play a crucial role in the pathogenesis of Parkinson's disease (PD) and other neurodegenerative diseases. The lipidomic strategy proposed here involves lipoprotein profiling using NMR spectroscopy and multivariate data pre-processing and analysis tools on 94 plasma samples (belonging to 38 early-stage PD patients, 10 PD-related dementia patients, 23 persons with Alzheimer's dementia, and 23 healthy control subjects) to firstly differentiate PD patients (irrespective of the stage of the disease) from persons with Alzheimer's disease (AD) as well as from controls, and then to discriminate among PD patients according to disease severity. The whole data set was subdivided into 86 training and 8 external test samples for validation purposes. A two-step classification scheme, based on linear discriminant analysis with variable selection accomplished by a stepwise orthogonalisation procedure, was proposed to optimise classification performance. Careful pre-processing of NMR signals was crucial to ensure data set quality. A total of 30 chemical shift buckets enabled differentiation between PD patients (regardless of disease severity), AD and control subjects, providing classification, cross-validation and external prediction rates of 100% in all cases. Only 15 variables were required to further discriminate between early-stage PD and PD-related dementia, again with 100% correct classifications, and internal/external predictions. The simplicity and effectiveness of the classification methodology proposed support the use of NMR spectroscopy, in combination with chemometrics, as a viable alternative diagnostic tool to conventional PD clinical diagnosis.